Journal of Computers - Academy Publisher
Journal of Computers - Academy Publisher
Journal of Computers - Academy Publisher
You also want an ePaper? Increase the reach of your titles
YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.
<strong>Journal</strong> <strong>of</strong> <strong>Computers</strong><br />
ISSN 1796-203X<br />
Volume 6, Number 9, September 2011<br />
Contents<br />
Special Issue: Changes in Computer Application for Economic Analysis <strong>of</strong> Law and Business<br />
Management<br />
Guest Editor: Malin Song, Dingding Pan, Jie Wu, Li Yang, Hongping Zhou, and Christopher<br />
Clemence<br />
Guest Editorial<br />
Malin Song, Dingding Pan, Jie Wu, Li Yang, Hongping Zhou, and Christopher Clemence<br />
SPECIAL ISSUE PAPERS<br />
Tax Evasion, Taxation Inspection and Net Tax Revenue: from an Optimal Tax Administration<br />
Perspective<br />
Bing Liu<br />
The Developmental Analysis <strong>of</strong> China’s Information Technology Services<br />
Wei Gao, Feng Wang, and Li Wang<br />
A Web Survey Program Based on Computer Technology and Its Application to Evaluation Model<br />
about Youth Self-organizations in China<br />
Ma-lin Song, Tong Yang, and Ya-qing Song<br />
The Research on the Influencing Factors <strong>of</strong> Financing Strategy <strong>of</strong> Woman Entrepreneurs in China<br />
Xiong Xiong, Rong Fu, Wei Zhang, Yongjie Zhang, and Lin Xiong<br />
A Spatial Econometric Analysis <strong>of</strong> China’s Manufacturing Agglomeration based on Geoda and<br />
Matlab<br />
Huayin Yu and Weiping Gu<br />
Application <strong>of</strong> Computer Technology in Efficiency Analysis <strong>of</strong> China Life Insurance Company<br />
Hongling Wu and XiaoFei Zeng<br />
A Bayesian Belief Net Model for Evaluating Organizational Safety Risks<br />
Li Song, Li Yang, Jing Han, and Jinkai Li<br />
Research and Application <strong>of</strong> J2EE and AJAX Technologies in Industry Report<br />
Min Hu, Ding-ding Pan, and Pei-en Zhou<br />
The Analysis <strong>of</strong> China New Energy Vehicle Industry Alliance Status based on UCINET S<strong>of</strong>tware<br />
Xiongfei Guo and Yingqi Liu<br />
Efficiency Evaluation Information System Based on Data Envelopment Analysis<br />
Jing Han and Malin Song<br />
An Optimal Inventory Control Model for a Supply Chain with Shortage Constraints<br />
Yinkuan Gu and Hongxia Zhang<br />
1797<br />
1799<br />
1805<br />
1812<br />
1819<br />
1825<br />
1832<br />
1842<br />
1847<br />
1852<br />
1857<br />
1862
Variable Selection for Credit Risk Model Using Data Mining Technique<br />
Kuangnan Fang and Hong Huang<br />
Corporate-, Product-, and User-Image Dimensions and Purchase Intentions —The Mediating Role <strong>of</strong><br />
Cognitive and Affective Attitudes<br />
Xian Guo Li, Xia Wang, and Yu Juan Cai<br />
A Microcomputer-Based Predictive Digital Current Programmed Control System for Three-phase<br />
PWM Rectifier<br />
Zhongjiu Zheng, Gu<strong>of</strong>eng Li, and Ninghui Wang<br />
Supply Chain Coordination under Return Policy with Asymmetric Information about Cost <strong>of</strong> Reverse<br />
Logistics Operations<br />
Ting Long Zhang<br />
Economic Development and Financial Support for Coal Resource Cities — A Panel Data Analysis<br />
Zuhuai Yuan, Li Yang, Jing Han, and Keliang Wang<br />
REGULAR PAPERS<br />
Solving the Sparsity Problem in Recommender Systems Using Association Retrieval<br />
YiBo Chen, ChanLe Wu, Ming Xie and Xiaojun Guo<br />
Integrated Structure and Control Design for Servo System Based on Genetic Algorithm and Matlab<br />
Dingzhen Li and Ruimin Jin<br />
A Model to Select System Core and Its Application<br />
Chongming Li and Yue Ding<br />
De-noise Comprehensive Research On Airplane Cockpit Signals Recorded by CVR<br />
Dao-Lai Cheng, Chui-JieYi, and Hong-Yu Yao<br />
Fuzzy Support Vector Machines Control for 6-DOF Parallel Robot<br />
Dequan Zhu, Tao Mei, and Lei Sun<br />
Parameters Optimization <strong>of</strong> Least Squares Support Vector Machines and Its Application<br />
Chunli Xi, Cheng Shao, and Dandan Zhao<br />
The Expected Value Model <strong>of</strong> Multiobjective Programming and its Solution Method Based on<br />
Bifuzzy Environment<br />
Mingfa Zheng , Bingjie Li, and Guangxing Kou<br />
A Method for Building Partially Connected Neural Network<br />
Gang Li, Xingsan Qian, Chunming Ye, and Lin Zhao<br />
A Cooperative Co-evolution PSO for Flow Shop Scheduling Problem with Uncertainty<br />
Bin Jiao, Qunxian Chen, and Shaobin Yan<br />
A Double Margin Based Fuzzy Support Vector Machine Algorithm<br />
Kai Li and Xiaoxia Lu<br />
A Modified Technique for Analysis <strong>of</strong> Synchronous Counters Constructed with Flip-flops<br />
Dangui Yan, Ruijun Tong, Chengchang Zhang, and Changyong Li<br />
A New Method <strong>of</strong> Detecting Multi-component LFM Signals Based on Blind Signal Processing<br />
Qiang Guo, Yajun Li, and Changhong Wang<br />
1868<br />
1875<br />
1880<br />
1886<br />
1891<br />
1896<br />
1903<br />
1913<br />
1920<br />
1926<br />
1935<br />
1942<br />
1949<br />
1955<br />
1962<br />
1971<br />
1976
Research on Self-built Digital Resource Backup Systems<br />
Li-zhen Shen<br />
Configuration Scheme for Small Scale Multi-FPGA Systems<br />
Chengchang Zhang, Lisheng Yang, Dangui Yan, and Changyong Li<br />
Order Bi-spectrum For Bearing Fault Monitoring and Diagnosis Under Run-up Condition<br />
Hui Li<br />
1983<br />
1988<br />
1994
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1797<br />
Special Issue on Changes in Computer Application<br />
for Economic Analysis <strong>of</strong> Law and Business Management<br />
Guest Editorial<br />
Nowadays, the computer is ubiquitous in the business world; many new computer applications have been developed<br />
for economic analysis <strong>of</strong> law and business management evaluation and forecasting in the past several years. With the<br />
rise <strong>of</strong> the Internet, computing is ever more integral to many disciplines than ever. As the average business becomes<br />
more and more computerized, so to does the science <strong>of</strong> studying business and economic analysis <strong>of</strong> law. These fields<br />
benefited greatly from the rise in quantity and power <strong>of</strong> the computer as well.<br />
“Tax Evasion, Taxation Inspection and Net Tax Revenue: from an Optimal Tax Administration Perspective”<br />
discusses the tax evasion and builds a general equilibrium model. In this paper, the interaction between tax declaration<br />
and taxation inspection is analyzed, and some policies and proposals about taxation inspection are proposed.<br />
“The Developmental Analysis <strong>of</strong> China’s Information Technology Services” makes an analysis on the<br />
development features, significance, present situation and existed problems <strong>of</strong> information technology services in China,<br />
and gives some relative suggestions on how to develop the information services <strong>of</strong> China better.<br />
“A Web Survey Program Based on Computer Technology and Its Application to Evaluation Model about<br />
Youth Self-organizations in China” studies the network and youth self-organizations based on web-platform, forecasts<br />
the developmental trend <strong>of</strong> adolescents by analyzing their current situation in China, and builds the evolution model for<br />
youth self-organizations.<br />
“The Research on the Influencing Factors <strong>of</strong> Financing Strategy <strong>of</strong> Woman Entrepreneurs in China” examines<br />
gender differences among Chinese entrepreneurs seeking financing pattern, studies the factors those affect women<br />
entrepreneurs’ financing strategies.<br />
“A Spatial Econometric Analysis <strong>of</strong> China’s Manufacturing Agglomeration based on Geoda and Matlab” uses<br />
spatial econometric methods to analyze the influencing factors <strong>of</strong> China’s provincial manufacturing Agglomeration by<br />
Geoda s<strong>of</strong>tware and Matlab network tools.<br />
“Application <strong>of</strong> Computer Technology in Efficiency Analysis <strong>of</strong> China Life Insurance Company” aims at<br />
studying the application <strong>of</strong> computer technology in efficiency analysis <strong>of</strong> China life insurance company.<br />
“A Bayesian Belief Net Model to evaluating Organizational Safety Risks” presents a methodology for<br />
organizational risk analysis for safety management.<br />
“Research and Application <strong>of</strong> J2EE and AJAX Technologies in Industry Report” analyzes the weakness <strong>of</strong> the<br />
traditional industry report, and proposes an industry report system based on J2EE and AJAX technologies.<br />
“The Analysis <strong>of</strong> China New Energy Vehicle Industry Alliance Status based on UCINET S<strong>of</strong>tware” uses the<br />
s<strong>of</strong>tware UCINET to draw up the picture <strong>of</strong> China’s new energy vehicle industry alliance network, and studies the<br />
cooperation relationships within the alliances through analyzing their elements.<br />
“Efficiency Evaluation Information System Based on Data Envelopment Analysis” studies the data envelopment<br />
analysis, and demonstrates the bridge between DEA and MIS for building efficiency evaluation information system.<br />
“An Optimal Inventory Control Model for a Supply Chain with Shortage Constraints” researches on the<br />
inventory decision model <strong>of</strong> the minimum total annual cost <strong>of</strong> the supply chain.<br />
“Variable Selection for Credit Risk Model Using Data Mining Technique” estimates long term default probability<br />
for developing appropriate credit risk model with the estimated default probability using Transition Matrix and mapping<br />
methods.<br />
“Corporate-, Product-, and User-Image Dimensions and Purchase Intentions” investigates the effects <strong>of</strong><br />
corporate-, product-, and user image dimensions on purchase intensions with cognitive and affective attitudes as<br />
mediator, and conducts a questionnaire survey.<br />
“A Microcomputer-Based Predictive Digital Current Programmed Control System for Three-phase PWM<br />
Rectifier” sets up a microcomputer control system for three-phase PWM rectifier using the floating-point digital signal<br />
processor TMS320LF2407 from Texas Instruments.<br />
“Supply Chain Coordination under Return Policy with Asymmetric Information about Cost <strong>of</strong> Reverse<br />
Logistics Operations” predicts the return policy and supply chain coordination in a channel <strong>of</strong> one supplier and one<br />
retailer.<br />
“Economic Development and Financial Support for Coal Resource Cities — A Panel Data Analysis” considers<br />
the high industry concentration <strong>of</strong> financial resources, which leads to a decline in financial resource allocation<br />
efficiency.<br />
We hope that the readers <strong>of</strong> this Special Issue could find and would enjoy something, such as the academic ideas,<br />
methods and enlightening form the papers in this Special Issue.<br />
© 2011 ACADEMY PUBLISHER<br />
doi:10.4304/jcp.6.9.1797-1798
1798 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
Guest Editors:<br />
Dr. Malin Song is an associate pr<strong>of</strong>essor in School <strong>of</strong> Statistics and Applied Mathematics, Anhui University <strong>of</strong> Finance and<br />
Economics. He is a Research Fellow in Economic Development Research Center, Anhui University <strong>of</strong> Finance and Economics. His<br />
major field <strong>of</strong> study includes Logistics, Environmental Economics and System Modeling and Analysis. Email:<br />
malinsong@gmail.com.<br />
Mr. Ding-ding Pan received his bachelor's degree in Computer Science and Technology from Anhui University <strong>of</strong> Architecture,<br />
Hefei, China (2008) and master's degree in Computer Application Technology from Hefei University <strong>of</strong> Technology, Hefei, China.<br />
His research interests include computer application, s<strong>of</strong>tware engineering. Email: panding1986@sina.com.<br />
Dr. Jie Wu, School <strong>of</strong> Management, University <strong>of</strong> Science and Technology <strong>of</strong> China, Hefei, Anhui Province, P. R. China, 230026.<br />
Phone: +8613966717485, Email: wujie012@ustc.edu<br />
Dr. Li Yang is a pr<strong>of</strong>essor in School <strong>of</strong> Economics and Management, Anhui University <strong>of</strong> Science & Technology, Huainan, Anhui,<br />
China. He is currently a doctor candidate in the School <strong>of</strong> Management at University <strong>of</strong> Science & Technology <strong>of</strong> China, Hefei,<br />
Anhui, China. His major field <strong>of</strong> study includes credit risk, strategic alliance and coal-mining eco-industrial park. E-mail:<br />
yangli081003@163.com.<br />
Dr. Hongping Zhou is Master Instructor in School <strong>of</strong> Engineering Science, University <strong>of</strong> Science and Technology <strong>of</strong> China. Her<br />
major fields <strong>of</strong> study are Computer Applications and Digital Circuit Design. She earned her Bachelors, Masters and Doctoral<br />
degrees in the School <strong>of</strong> Information Science and Technology, University <strong>of</strong> Science and Technology <strong>of</strong> China. Email:<br />
zhouhp@sina.cn.<br />
Dr. Christopher Clemence earned his Juris Doctorate at the University <strong>of</strong> Missouri-Kansas City School <strong>of</strong> Law. He practices<br />
real estate, energy, environmental and transaction law, and currently works for a Fortune 500 company in the United States. Email:<br />
conagher78@gmail.com.<br />
© 2011 ACADEMY PUBLISHER
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1799<br />
Tax Evasion, Taxation Inspection and Net Tax<br />
Revenue: from an Optimal Tax Administration<br />
Perspective<br />
Bing Liu<br />
School <strong>of</strong> Economics and Management/Anhui Normal University, Wuhu, China<br />
Abstract— Tax evasion has always been an important topic<br />
to tax theory researchers and the department <strong>of</strong> government.<br />
However, existing research results are confined to the<br />
unilateral action <strong>of</strong> taxpayers, neglect the interaction<br />
between the tax declaration and taxation inspection. This<br />
paper, from an optimal tax administration perspective,<br />
builds a general equilibrium model, in which, taxation<br />
inspection cost, net tax revenue and taxpayers personal<br />
expected utility maximization, are included, to analyze the<br />
interaction between the tax declaration and taxation<br />
inspection. Then it proposes some policies and proposals<br />
about taxation inspection.<br />
Index Terms—tax evasion, taxation inspection, net tax<br />
revenue<br />
I. INTRODUCTION<br />
Tax evasion is the economic activities that taxpayers<br />
through illegal channels to reduce their tax payable.<br />
Large-scale tax evasion will not only affect a<br />
government's fiscal revenue, lead to the failure <strong>of</strong> a<br />
government's macroeconomic indicators, distortion <strong>of</strong><br />
resource allocation and income distribution out <strong>of</strong><br />
control. Tax evasion is widespread in the world,<br />
according to the statistics, in developed countries, 22<br />
high-income countries which per capita GNP more than<br />
8626 U.S. dollars (such as Germany, Japan, Switzerland,<br />
United Kingdom, the United States and other countries),<br />
and 10 upper-middle-income countries which per capita<br />
GNP between 2786-8625 U.S. dollars (such as Argentina,<br />
Brazil, Chile and other countries), the tax loss is about<br />
35%. 9 lower-middle-income countries which per capita<br />
GNP between 696-2785 dollar U.S. dollars (such as<br />
Colombia, the Czech Republic, Indonesia and other<br />
countries) and 5 low-income countries which per capita<br />
GNP below 695 U.S. dollars (including Egypt, India,<br />
Zambia and other countries) loss 30% and 60%. Such as<br />
the U.S. the tax loss between 30%- 45%, the Netherlands,<br />
the tax loss between 22% -35%. Japan's tax loss vary<br />
according to the size <strong>of</strong> the taxpayer, Large, medium and<br />
small taxpayers, the loss around 20%, 40%, 60%.<br />
In China, tax evasion is very serious and has become a<br />
well-known fact, according to the scholars calculation: in<br />
1999, China's tax revenue loss was about 77.6 billion<br />
RMB(China yuan), if calculated with the amount that<br />
should be collected, the loss exceeded 100 billion RMB,<br />
© 2011 ACADEMY PUBLISHER<br />
doi:10.4304/jcp.6.9.1799-1804<br />
if measurement by the combination <strong>of</strong> factors method, the<br />
loss will reach 320 billion to 430 billion, was the entire<br />
tax revenue’s 30% -40%. Experts conservatively<br />
estimated that, in 2004, tax loss was at least 450 billion,<br />
was the entire tax revenue’s 15%. in 2004, the National<br />
Audit Office audited 788 enterprises from 19 provinces<br />
and cities, found that in 2002 the tax should be collected<br />
was 117.35 billion, but actually the amount was 103.96<br />
billion, pay less tax 13.38 billion, accounting for the tax<br />
that should be collected 11.41%; January to September,<br />
2003, the tax should be collected 103.78 billion, 918.84<br />
billion actually collected, less 118.94 billion, accounting<br />
for 11.46% <strong>of</strong> tax should be collected, in this way, tax<br />
wastage nationwide in 2007 up to 500 billion.<br />
Tax evasion has been the important issues which<br />
theory researchers and government departments<br />
concerned about all the time. as China's economy<br />
development and the tax system improvement, more and<br />
more scholars research the issue. In this paper, based on<br />
Chinese and foreign scholars previous studies, from the<br />
point <strong>of</strong> tax audit cost and net revenue, study a general<br />
equilibrium model about tax evasion, tax inspection and<br />
tax net income, give some suggestions about the optimal<br />
behavior <strong>of</strong> tax collection for tax administration<br />
reference.<br />
II. THEORY REVIEW<br />
A. A-S model<br />
The first use <strong>of</strong> modern methods <strong>of</strong> economics to study<br />
the problem <strong>of</strong> tax evasion is U.S. economist Kagan.P, he<br />
used the cash ratio method to estimate the scale <strong>of</strong> U.S.<br />
tax loss in "The total money supply and the<br />
corresponding currency demand" (1958). M G Allingham<br />
and A Sandmo constructed a theoretical model, it’s<br />
theory based on Gary Becker’s study on economics <strong>of</strong><br />
crime and A Sandmo’s research on risk and uncertainty.<br />
Allingham and Sandem’s tax evasion model established<br />
the theoretical framework <strong>of</strong> tax evasion, is a classic<br />
model <strong>of</strong> tax evasion, <strong>of</strong>ten referred to as A-S model.<br />
A-S model is a model <strong>of</strong> expected utility<br />
maximization, the basic assumptions are: (1) taxpayer’s<br />
cardinal utility maximization as objective function, and<br />
cardinal utility is a single function <strong>of</strong> income; (2) the<br />
taxpayer's marginal utility is positive and decreasing; (3)<br />
the taxpayer's actions are consistent with Von Neumann
1800 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
and Morgenstern action rules under uncertainty; (4)<br />
proportional tax system; (5) tax authorities’ net income<br />
maximization with the budget constrained; (6) tax<br />
inspectors discovered a constant probability; (7) penalty<br />
based on the difference between taxable income and<br />
reported income, rather than the tax evasion, and the<br />
penalty ratio is higher than tax rate. In addition, tax<br />
audition will not add cost to taxpayers, and the taxpayer's<br />
real income after checking can be drawn. With these<br />
strict assumptions, the taxpayer's objective function can<br />
be expressed as:<br />
E( U ) = ( 1−<br />
p)<br />
U ( w − tX ) + pU[<br />
W −Tx<br />
− r(<br />
W − X )]<br />
(1)<br />
Here, U is taxpayer’s disposable income utility, E (U)<br />
is expected utility; W is taxpayer’s real income, an<br />
exogenous variable, in the cases <strong>of</strong> incomplete<br />
information, the tax authority can not accurately grasp it;<br />
X for the taxpayer’s taxable income declared to the tax<br />
authority, a endogenous variable, 0 ≤ X ≤ W; p for the<br />
probability <strong>of</strong> taxpayer seized by the tax authority; t for<br />
the tax rate set by the tax authority, r is the penalty ratio<br />
when be investigated, 0
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1801<br />
in the income areas, ∆ ≥ 0<br />
+<br />
Y , Otherwise, in the loss<br />
areas and ∆ ≤ 0<br />
−<br />
+<br />
Y , ∆Y 、 −<br />
∆Y can be understood as<br />
the wealth value changes relative to the reference point.<br />
By (5), (6) we have:<br />
+ + 0 +<br />
∆Y<br />
= Y −Y<br />
= Y − ( 1−<br />
t)<br />
W = t(<br />
W − D)<br />
≥ 0<br />
(7)<br />
− + 0 +<br />
∆Y<br />
= Y −Y<br />
= Y − ( 1−<br />
t)<br />
W = −(<br />
s + λt)(<br />
W − D)<br />
≤ 0<br />
(8)<br />
To (7), (8) into (4) can be obtained<br />
+ β<br />
β<br />
⎪⎧<br />
( ∆ Y ) = [ t ( W − D ) ] , x ≥ 0<br />
V ( x ) = ⎨<br />
β ∈ [ 0,<br />
1 , ]θ<br />
− β<br />
β<br />
⎪⎩ − θ ( − ∆ Y ) = −θ<br />
[ ( λ t + s)(<br />
W − D ) ] , x ≤ 0 >1 (9)<br />
Here, ∆Y is the amount <strong>of</strong> gain or loss relative to the<br />
reference point, θ is the aversion coefficient, seized<br />
probability p(D)(gain), not being seized probability 1p(D)(loss).<br />
According to the Prospect theory, People tend<br />
to give objective probability a lower or higher subjective<br />
probability. Therefore, with the prospect theory,<br />
weighting function for the loss state (being seized) is<br />
W p(D)<br />
−<br />
, weighting function for the gain state(not being<br />
+<br />
seized) is<br />
W [ 1−<br />
p(<br />
D)]<br />
. According to (9) and weight<br />
function, the value function <strong>of</strong> taxpayer as follow:<br />
[ ][ ] [ ] β<br />
β<br />
+<br />
−<br />
V( D,<br />
t,<br />
s,<br />
λ , θ)<br />
= W 1−p(<br />
D)<br />
t(<br />
W−D)<br />
−W<br />
p(<br />
D)<br />
θ(<br />
λt<br />
+ s)(<br />
W−D)<br />
(10)<br />
In prospect theory, the taxpayer's goal is to maximize<br />
the value function V <strong>of</strong> (10). With boundary conditions,<br />
the relationship between variables can be gotten. The<br />
relationship between tax rates and less declaring income,<br />
for example, tax rate increase, ∂V/∂D will be strictly<br />
negative, that is, underreporting <strong>of</strong> income will allow<br />
taxpayers to increase the value <strong>of</strong> the function. Therefore,<br />
the escape <strong>of</strong> will increase with the tax rate increase,<br />
there is a positive correlation between them, which is the<br />
important difference with the traditional A-S model. It<br />
explains, in reality, why tax evasion is widespread in the<br />
high-income groups.<br />
C. Tax evasion cost-benefit model<br />
Let the real income <strong>of</strong> the taxpayer to be W, taxpayers<br />
are likely to avoid tax by declaration <strong>of</strong> low taxable<br />
income, set X for its declared income, then the taxpayers’<br />
hided income R for W-X, seized probability <strong>of</strong> tax<br />
evasion for P, 0 C, the results <strong>of</strong><br />
inequality is:<br />
t > mP + a + s<br />
(14)<br />
That is, when taxpayers expect that tax evasion penalty<br />
(mp) suffered when being seized plus the operating costs<br />
(a) <strong>of</strong> tax evasion and psychological costs (s) are less than<br />
the taxes paid in accordance with the statutory tax rate,<br />
then the taxpayer will choose to evade. Formula (12),<br />
(13) are derived with R respectively, then, relative to the<br />
concealed income, the marginal benefit (MB) and the<br />
marginal cost (MC) <strong>of</strong> tax evasion are get:<br />
MB = t<br />
(15)<br />
MC = a + mP + s + RPm'(<br />
R)<br />
+ Rs'(<br />
R)<br />
(16)<br />
To maximize taxpayers’ expectation, in accordance<br />
with the principles <strong>of</strong> economics, the marginal benefit<br />
must equals to the marginal cost, it is MB = MC. If R is<br />
the horizontal axis, the marginal benefit curve MB can be<br />
expressed as a t height horizontal line, the marginal cost<br />
curve MC is a tilting curve up the right. By formula (16)<br />
and the known conditions, when R = 0, MC has the<br />
minimum a + mP + s, so the starting point <strong>of</strong> MC curve<br />
(R = 0, MC = a + mP + s) is lower than the MB curve, the<br />
two curves intersect at E, the R* corresponding to the<br />
intersection E is the best concealing amount where the tax<br />
escaper’s expected income maximization, and the best<br />
amount <strong>of</strong> tax evasion is R*t.<br />
However, look at the existing research results, the<br />
starting point confined to the consideration <strong>of</strong> unilateral<br />
acts <strong>of</strong> the taxpayer, ignored the interactive relationship<br />
between the behavior <strong>of</strong> declaration and tax audition. In<br />
fact, the conditions <strong>of</strong> risk selection taxpayers facing,<br />
such as penalty amount and the seized probability, are<br />
closely related to the behavior <strong>of</strong> the tax authorities’<br />
audition. If only consider the risks to taxpayer regardless<br />
the conditions <strong>of</strong> risk, the findings will be unconvincing.<br />
Based on this, to maximize the government’s net tax<br />
revenue, this article, including the variable <strong>of</strong> tax audit<br />
expenses, construct a general equilibrium model which<br />
cover the utility’s net tax revenue maximization and the<br />
taxpayer’s expectation maximization, to analyze the<br />
interaction between the tax inspection and the tax<br />
evasion.
1802 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
III. MODEL CONSTRUCT<br />
Based on the A-S model’s parameters setting, assume,<br />
in a tax year, the taxpayer’s reported less taxable income<br />
audited by the tax authorities for b(W-X), 0 ≤ b ≤ 1.<br />
From the perspective <strong>of</strong> tax authorities, taxpayer’s tax<br />
evasion is tb (W-X), accordingly, the taxpayer to pay the<br />
fine by rtb (W-X). Expenditures for the tax department’s<br />
inspection for C, the relationship between b and C can be<br />
expressed as b = b (C), it certainly has:<br />
b c<br />
db<br />
= > 0<br />
dC ,<br />
That is, with the spending <strong>of</strong> the tax department<br />
inspection increased, the higher the taxpayer’s taxable<br />
income which being inspected out.<br />
2<br />
d b<br />
bCC = 2<br />
dC ,<br />
With the increase in audit expenses, bC into decline by<br />
the increase, bCC turned negative from positive,<br />
indicating that the maximum value b exists.<br />
To the interaction between tax audit and the taxpayer’s<br />
behavior, here should be noted that, on the choice <strong>of</strong> the<br />
tax department’s policy instruments, in order to curb tax<br />
evasion, seeking to maximize the tax revenue, tax<br />
inspection (inspection expenses), the tax rate adjustment<br />
and penalty rates and other policy tools are available, but<br />
in the specific application, these tools are different. Tax<br />
and penalty rates are legal areas, generally not free to<br />
change, relatively speaking, tax inspection efforts and<br />
configuration <strong>of</strong> inspection project, can have a moderate<br />
change based on the subjective views <strong>of</strong> the tax<br />
department, so check expenditures are choice variables,<br />
the tax rates and punishment rate systematic exogenous<br />
variables.<br />
To be consistent with the A-S model, here set the<br />
actual income <strong>of</strong> the taxpayer for W, an exogenous<br />
variable, under the conditions <strong>of</strong> incomplete information,<br />
the tax authorities can not accurately grasp; X for taxable<br />
income the taxpayer declare to the tax authorities, an<br />
endogenous variable, 0 ≤ X ≤ W; p for the seized<br />
probability taxpayers seized by the tax authorities; t for<br />
the tax rate set by the tax authorities, r is the ratio <strong>of</strong> fine<br />
when being investigated, 0
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1803<br />
''<br />
U ( W2)<br />
R2<br />
= − > 0<br />
'<br />
U ( W2)<br />
, 2 1 R R ><br />
, only the follow equation<br />
being established, the tax rate increases will lead the<br />
taxpayer to reduce the reported income.<br />
rCR2bc<br />
k <<br />
rR2b(<br />
W − X ) − X ( R1<br />
− R2)<br />
(25)<br />
Discussion above show that, when the flexibility <strong>of</strong> tax<br />
audition expenses to marginal tax base is not high, the tax<br />
department can not further improve the performance <strong>of</strong><br />
tax inspection, raising tax rates will induce taxpayers to<br />
increase tax evasion, thus reducing tax compliance.<br />
B. The income <strong>of</strong> taxpayer W changes<br />
The impaction <strong>of</strong> taxpayer’s income changes on the<br />
changes in taxpayer’s reporting income depends on<br />
taxpayer’s utility maximization behavior, the X <strong>of</strong><br />
formula (21) is derived <strong>of</strong> W:<br />
dX ( 1 − p ) U ′ ( W 1 )<br />
=<br />
[ ( 1 − rtb ) R 2 − R 1 ]<br />
dW<br />
tZ<br />
(26)<br />
R1<br />
> ( 1−<br />
rtb)<br />
When R2<br />
set up, the sign <strong>of</strong> formula (26) is<br />
positive, that is, the higher taxpayers’ income, the higher<br />
the income <strong>of</strong> its report.<br />
Formula (26) means that, the higher taxpayers’<br />
income, the higher its probability <strong>of</strong> being audited. When<br />
there is tax evasion, the probability <strong>of</strong> its being seized and<br />
the cost <strong>of</strong> being punished are higher, too. Therefore, the<br />
higher the taxpayers’ income, the more likely an honest<br />
declaration.<br />
As for the proportion <strong>of</strong> declared income accounted for<br />
reported real income, according to the definition, there is:<br />
⎛ X<br />
∂ ⎜<br />
⎝ W<br />
∂ W<br />
⎞<br />
⎟<br />
⎠<br />
=<br />
1<br />
W<br />
2<br />
⎛<br />
⎜ W<br />
⎝<br />
∂ X<br />
∂ W<br />
−<br />
X<br />
⎞<br />
⎟<br />
⎠<br />
(27)<br />
Formula (27) means that, if the conclusion that the<br />
taxpayer’s real income W and reported income X change<br />
in the same direction is right, whether the proportion <strong>of</strong><br />
the declared income accounted for reported real income<br />
increases with the rising <strong>of</strong> the real income depends on<br />
the flexibility <strong>of</strong> declared income, it is, if the flexibility <strong>of</strong><br />
the reported income greater than 1, then the proportion <strong>of</strong><br />
declared income accounted for reported real income will<br />
increase with the rising <strong>of</strong> the real income, or if less than<br />
1, it will fall.<br />
C. Penalty rate and seized probability change<br />
The punitive rate and the seized probability change, its<br />
impaction to the taxpayers’ compliance can get by<br />
deriving equation (21) with r and p respectively:<br />
∂X<br />
1 '<br />
''<br />
b<br />
= { − pU ( W2<br />
) − ptW(<br />
r −1)<br />
U ( W2)[<br />
∂r<br />
tZ<br />
b<br />
2<br />
C<br />
cc<br />
− b(<br />
W − X )]} > 0<br />
(28)<br />
2<br />
∂X<br />
1 '<br />
'<br />
''<br />
rtbC<br />
= − [ U ( W1)<br />
+ ( r −1)<br />
U ( W2)<br />
+ ( r −1)<br />
U ( W2)<br />
] > 0<br />
∂p<br />
tZ<br />
bcc<br />
(29)<br />
Formula (28) and (29) shows that, the penalty rate and<br />
the rate <strong>of</strong> seizures increased, both mean adverse to tax<br />
evasion, thereby, the taxpayer will increase the reported<br />
income people, and increase tax compliance at last.<br />
© 2011 ACADEMY PUBLISHER<br />
V. CONCLUSIONS<br />
To the taxpayer, the tax evasion should pay the relative<br />
cost, that is, will face the possible punitive price. The tax<br />
department is no exception, to increase taxes, prevent tax<br />
evasion, also must spend a lot <strong>of</strong> manpower and<br />
resources, tax inspection and tax evasion are against each<br />
other and influence each other. The net revenue <strong>of</strong> the<br />
government and the utility <strong>of</strong> the taxpayer maximize<br />
respectively, in the cases there exist an optimal<br />
equilibrium solution, whereby the paper establish a<br />
general equilibrium model in which the net revenue <strong>of</strong> the<br />
government and the utility <strong>of</strong> the taxpayer maximize<br />
respectively. By analyzing the model, can draw the<br />
following conclusions:<br />
(1) The equilibrium solution to the government’s<br />
optimal inspection expenditures, is defined as spending<br />
one dollar <strong>of</strong> audit costs, must be equal to the overdue tax<br />
and penalty when the tax evasion to be checked out. The<br />
equilibrium solution to the maximization <strong>of</strong> the<br />
taxpayer’s utility is, the paying less tax’s expected<br />
benefits the taxpayer obtained by reducing a unit <strong>of</strong><br />
reported taxable income, must be equal to the expected<br />
marginal cost <strong>of</strong> paying an overdue tax and being<br />
punished when the tax evasion to be checked out.<br />
The behavior <strong>of</strong> the government and the taxpayer is the<br />
opposite: the taxpayer increase (decrease) the tax evasion,<br />
would enable the government to increase (decrease) tax<br />
audit expenditures; on other hand, with the government’s<br />
tax audit expenditures increase (decrease), would enable<br />
the taxpayer to increase (decrease) their tax compliance<br />
correspondingly.<br />
(2) In general, the tax rate increase, the tax evasion<br />
will be expanded. However, the tax rate increase will<br />
increase the marginal income <strong>of</strong> the tax audition<br />
expenditures, incentive the tax audition spending go up,<br />
and promote to enhance inspection efforts. Only when the<br />
marginal benefit <strong>of</strong> tax inspection expenditures is 0, then<br />
the degree <strong>of</strong> tax audition is optimal. Therefore, Whether<br />
or not the taxpayer evade tax depends on the flexibility <strong>of</strong><br />
tax audition expenses to marginal tax base and the degree<br />
<strong>of</strong> risk <strong>of</strong> their being seized. When the flexibility <strong>of</strong> tax<br />
audition expenses to marginal tax base is not high, the tax<br />
department can not further improve the performance <strong>of</strong><br />
tax inspection, raising tax rates will induce taxpayers to<br />
increase tax evasion, thus reducing tax compliance.<br />
(3) The higher the taxpayers’ income, the higher its<br />
probability <strong>of</strong> being audited. When there is tax evasion,<br />
the probability <strong>of</strong> its being seized and the cost <strong>of</strong> being<br />
punished are higher, too. Therefore, the higher the<br />
taxpayers’ income, the more likely to declare honestly.<br />
As for the change direction <strong>of</strong> the proportion <strong>of</strong><br />
declared income accounted for reported real income, if<br />
the conclusion that the taxpayer’s real income and<br />
reported income change in the same direction is right,<br />
whether the proportion <strong>of</strong> the declared income accounted<br />
for reported real income increases with the rising <strong>of</strong> the<br />
real income depends on the flexibility <strong>of</strong> declared<br />
income, it is, if the flexibility <strong>of</strong> the reported income<br />
greater than 1, then the proportion <strong>of</strong> declared income
1804 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
accounted for reported real income will increase with the<br />
rising <strong>of</strong> the real income, or if less than 1, it will fall.<br />
(4) Penalty rate and the detected rate increase means<br />
that the tax evasion will assume greater costs once being<br />
detected, which raise more awareness <strong>of</strong> the risk<br />
management <strong>of</strong> the taxpayer, the taxpayer will increase<br />
the reported income, thereby increasing tax compliance.<br />
Practical significance <strong>of</strong> this study is, view China’s tax<br />
reality, the tax loss has the characteristics <strong>of</strong> wide range,<br />
large number, diversity. With the diversification <strong>of</strong><br />
economic entities and the diversification <strong>of</strong> mode <strong>of</strong><br />
operation, the means <strong>of</strong> illegal and crime tax-related has<br />
become increasingly complex. In short, with the everchanging<br />
tactics <strong>of</strong> corporate tax evasion, tax means more<br />
and more hidden, tax audit work has become increasingly<br />
difficult. In addition, as for the power <strong>of</strong> the tax audit and<br />
the inspection level, the seized probability <strong>of</strong> current tax<br />
authorities on tax evasion cases is very low, usually not<br />
more than 50% [7]. For the above, if to improve the<br />
seizure rate is bound to increase a large number <strong>of</strong> tax<br />
<strong>of</strong>ficials and the huge audit costs, the result will not<br />
necessarily bring about the increase in net revenue. As for<br />
the tax and punishment rates increase, not only to revise<br />
the relevant laws, but also will increase corporation’s tax<br />
burden, causing social dissatisfaction, in fact, it is<br />
feasible, too. Therefore, how to adjust the structure <strong>of</strong> the<br />
© 2011 ACADEMY PUBLISHER<br />
tax audit and expenditure, under the conditions <strong>of</strong> the<br />
existing human and material resources <strong>of</strong> the tax<br />
authorities, to achieve the maximum <strong>of</strong> the net tax<br />
revenue would be the optimal orientation <strong>of</strong> the tax<br />
administrative act.<br />
REFERENCES<br />
[1] Allingham, M.G. & A. Sandmo. 1972, Income tax evasion:<br />
a theoretical analysis. <strong>Journal</strong> <strong>of</strong> Public Economics, 1,<br />
pp.323-338.<br />
[2] Yitzhaki, S.. 1974, A note on income tax evasion: a<br />
theoretical analysis. <strong>Journal</strong> <strong>of</strong> Public Economics, 3,<br />
pp.201-202.<br />
[3] Lin Wei, 2001,Tax collection column. Tax Research,1,<br />
pp.18-23.<br />
[4] Ali, M. M., H. W. Cecil., & J. A. Knoblett., 2001, The<br />
effect <strong>of</strong> tax rates and enforcement policies on tax payers<br />
compliance: a study <strong>of</strong> self-employed taxpayers.<br />
Economics <strong>Journal</strong>, 29(2), pp.186-202.<br />
[5] Ruan Jiafu. 2005, On the tax status <strong>of</strong> non-compliance,<br />
causes and countermeasures, Modern Finance, 1,pp.78-89.<br />
[6] Lo Kuang, Xiao Yan Fen, 2007, Consider the tax<br />
compliance cost <strong>of</strong> tax evasion model, Tax Research, 1,<br />
pp.78-96.<br />
[7] Zhu Feng, 2007, Research tax evasion in China. Market<br />
Weekly, 4, pp.32-38.
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1805<br />
The Developmental Analysis <strong>of</strong> China’s<br />
Information Technology Services<br />
Wei Gao<br />
College <strong>of</strong> Statistics and Applied Mathematics /Anhui University <strong>of</strong> Finance and Economics, Bengbu, China<br />
Email: gaowei.@163.com<br />
Feng Wang and Li Wang<br />
College <strong>of</strong> Statistics and Applied Mathematics/Anhui University <strong>of</strong> Finance and Economics, Bengbu, China<br />
Abstract—Information technology has become an<br />
indispensable part to modern society. Information<br />
technology services as an independent industry has a<br />
pr<strong>of</strong>ound effect on the progress <strong>of</strong> whole society and play on<br />
an importance role the development <strong>of</strong> the overall socioeconomic.<br />
This article makes an analysis on the<br />
developmental features, significance, present situation and<br />
existed problems <strong>of</strong> China’s information technology services<br />
and gives some relative suggestions as to how to develop<br />
China’s information services better.<br />
Index Terms—Information Technology, Information<br />
Technology Services, Present Situation Strategy<br />
I. INTRODUCTION<br />
Your goal is to simulate the usual appearance <strong>of</strong> papers<br />
in a <strong>Journal</strong> <strong>of</strong> the <strong>Academy</strong> <strong>Publisher</strong>. We are requesting<br />
that you follow these guidelines as closely as possible.<br />
Information technology services industry uses<br />
networks, computers and other modern scientific<br />
technology to produce, collect, process, store, transmit<br />
and use information and it is a specialized industry<br />
collection for providing service to society by information<br />
products. Various economic fields are linked close to<br />
information technology. The application <strong>of</strong> global<br />
information technology is developing deeply and widely.<br />
The development <strong>of</strong> information technology makes a very<br />
large impact on the technology-related industries and it<br />
has a promotion effect on improving productivity <strong>of</strong> the<br />
related industries. But, information technology is an<br />
independent industry separated from others, whose<br />
contribution has a bright future. As a sunrise industry<br />
along with its rapid development, there is a certain degree<br />
<strong>of</strong> shortcomings and deficiencies. Therefore, this article<br />
attempts to analyze the existed problems in the<br />
development <strong>of</strong> information technology services. So as to<br />
make it play better in the related-industries and to further<br />
promote economic development.<br />
II. ANALYSIS OF CHINA'S INFORMATION TECHNOLOGY<br />
SERVICES FEATURES<br />
(a) Information technology services industry is a<br />
means <strong>of</strong> modern services industry. Users use modern<br />
© 2011 ACADEMY PUBLISHER<br />
doi:10.4304/jcp.6.9.1805-1811<br />
means to obtain information, not only using such tools as<br />
book-style index and abstracts but also using CD-ROM<br />
or on-line search to obtain needed information.<br />
(b) Information technology services industry is<br />
engaged in industry <strong>of</strong> information technology.<br />
Information technology services, not to industrial<br />
enterprises, which is a department <strong>of</strong> the third industry.<br />
(c) Information technology services industry is the<br />
services aimed at specific objects, i.e. finding specific<br />
information for specific users, and finding specific user<br />
for specific information.<br />
(d) Information technology services industry is a high<br />
value-added industry, compared with the general services<br />
department, information services with high value-added.<br />
(e) Information technology services industry is a high<br />
penetration industry. Information technology and<br />
information services can penetrate and spread to all areas<br />
<strong>of</strong> society and industry sectors and has an active leading<br />
role to the department <strong>of</strong> other industry.<br />
(f) Information technology services industry is one<br />
which provides special commodities. Information<br />
technology services industry is characteristic <strong>of</strong> wide<br />
varieties, wide range, high value <strong>of</strong> transfer, easy to copy,<br />
timeliness and sharing. Furthermore, information<br />
technology commodity achieving its own value must rely<br />
on specific technology and service.<br />
(g) Information technology services industry is<br />
intelligence-intensive industry. IT services has feature <strong>of</strong><br />
intelligence-intensive than other services department.<br />
Therefore, information technology services industry is the<br />
product <strong>of</strong> the knowledge economy era.<br />
(h) Information technology services, information<br />
circulation <strong>of</strong> commodities market has a special industry.<br />
The special nature <strong>of</strong> information goods, information<br />
goods exchange form and scope <strong>of</strong> the diversified<br />
characteristics.<br />
III. SIGNIFICANCE OF CHINA DEVELOPING OF IT<br />
SERVICES<br />
The development <strong>of</strong> IT services as an independent<br />
industry has far-reaching impact on development <strong>of</strong><br />
national economy and social progress. It belongs to a new<br />
style <strong>of</strong> industry with high value-added, high technology,<br />
low consumption <strong>of</strong> energy resources, low environmental
1806 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
pollution, high industry association leading role and good<br />
use <strong>of</strong> human resources. Its development has a promoting<br />
role to the other industry.<br />
A. The Promoting Role <strong>of</strong> the Development <strong>of</strong> IT Services<br />
to the Whole Society and the Development <strong>of</strong> National<br />
Economy<br />
(a) The Role <strong>of</strong> IT Services Development to National<br />
Education<br />
In the information age, information is the decisive<br />
resource. However, in our country, people’s<br />
understanding that information is a kind <strong>of</strong> resources is<br />
very vague. The waste <strong>of</strong> information resources is great<br />
and the efficiency <strong>of</strong> intensive utilization <strong>of</strong> information<br />
resources is very low. The development <strong>of</strong> modern IT<br />
services plays an actively educational role in improving<br />
and expanding the utilization efficiency <strong>of</strong> information<br />
resources, educating national people to cherish and use<br />
information resources and developing awareness <strong>of</strong><br />
people’s consumption. Trying hard to develop databases<br />
and consulting services as modern IT services contributes<br />
to progressively faster the concept “compensation for the<br />
use <strong>of</strong> information” and reverses vague understanding <strong>of</strong><br />
information resources.<br />
(b) The Development <strong>of</strong> IT Services Helping to<br />
Promote the Fusion <strong>of</strong> Industrialization and Information<br />
Now, one <strong>of</strong> the major tasks in our country is to<br />
achieve the integration <strong>of</strong> industrialization and<br />
information. IT services in upgrading and transformation<br />
<strong>of</strong> traditional industries forms a new road to<br />
industrialization by pouring new content into traditional<br />
industries and deepening the use information. In turn,<br />
new industrialization focusing on the information<br />
application also provides a broad market space and<br />
material and technical basis for the development <strong>of</strong><br />
information services. Thus, IT services can drive the<br />
integration <strong>of</strong> industrialization and industry through the<br />
promotion information technology. In the cycle <strong>of</strong><br />
interaction <strong>of</strong> Information technology and<br />
industrialization, the social and economic leaps and<br />
bounds will be achieved by promoting the formation <strong>of</strong><br />
modern agriculture, services, and new industries.<br />
(c) The Development <strong>of</strong> IT Services Industry Itself<br />
Being Able to Bring Scale Expansion and the Change <strong>of</strong><br />
the Mode <strong>of</strong> Economic Growth<br />
IT services belongs to an important part <strong>of</strong> information<br />
industry. IT services has become a highlight <strong>of</strong> the new<br />
century. The rapid development <strong>of</strong> IT services promotes<br />
the rapid development <strong>of</strong> information industry. Its<br />
contribution to the improvement <strong>of</strong> the information<br />
industry improves the proportion <strong>of</strong> the information<br />
industry in GDP. Information production, circulation and<br />
consumption scale keeps expanding and further<br />
stimulates people to create more demand for information.<br />
With the development <strong>of</strong> information technology, new<br />
growth points emerge in economic field such as IT<br />
industry, information services and Internet industries, and<br />
meanwhile, information can promote the growth <strong>of</strong> the<br />
total national economy through the optimization <strong>of</strong><br />
production systems. Information technology also helps<br />
break the time limit and geographic restrictions <strong>of</strong> the<br />
© 2011 ACADEMY PUBLISHER<br />
market and speeds up the market <strong>of</strong> information<br />
processing and circulation.<br />
(d) The Development <strong>of</strong> IT Services Industry Helping<br />
the Increase <strong>of</strong> Economic Efficiency and Promoting the<br />
Whole Social Economic Development<br />
IT services industry is viewed as the industry <strong>of</strong><br />
processing information whose information search and the<br />
improvement <strong>of</strong> transmission and switching efficiency<br />
has an important promoting role to the reduction <strong>of</strong><br />
information asymmetry and lowering transaction costs. In<br />
addition, IT services industry has a significant multiplier<br />
and diffusion effect. On the one hand, IT services is<br />
beneficial to raising the utilization efficiency and<br />
information content <strong>of</strong> productive factors in other<br />
industries through the penetration <strong>of</strong> all economic fields,<br />
directly or indirectly producing a significant impact. On<br />
the other hand, with its own development and innovation,<br />
some new type will create <strong>of</strong> such as virtual tourism and<br />
distance education and so on, directly creating<br />
employment and economic output.<br />
B. The Development <strong>of</strong> IT Services Industry Having a<br />
Promotion Effect on the Other Industries Related<br />
(a) The Development <strong>of</strong> IT Services Industry<br />
Providing a Huge Market Demand for the Development<br />
<strong>of</strong> and Third and Other Industries<br />
In recent years, the rapidly growing demand <strong>of</strong> IT<br />
services not only provides the market demand <strong>of</strong> rapid<br />
growth for the and third and other Industries but also<br />
putting forward an urgent request for the development <strong>of</strong><br />
IT services industry itself. IT services belongs to<br />
knowledge and technology-intensive industries whose<br />
rapid development is linked close to the manufacture <strong>of</strong> a<br />
large number <strong>of</strong> senior technical pr<strong>of</strong>essional talents. It<br />
provides broad market prospects for the development <strong>of</strong><br />
the related industries and educational career. Therefore<br />
the rapid development <strong>of</strong> IT services has a leading role to<br />
the gross economic output and scale <strong>of</strong> the third and<br />
related industries.<br />
(b) The Role <strong>of</strong> IT Services Industry Development<br />
Optimizing the Industrial Structure and S<strong>of</strong>tening<br />
National Economy<br />
IT services industry can act as “catalyst role” in the<br />
development <strong>of</strong> national economy. The function <strong>of</strong> IT<br />
services optimizing the industrial structure is not only<br />
reflected in the development <strong>of</strong> IT services itself directly<br />
promoting optimization <strong>of</strong> industrial structure but also<br />
reflected in IT services as a typical general purpose<br />
technology indirectly promoting optimization <strong>of</strong><br />
industrial structure. It leads to a series <strong>of</strong> production and<br />
change <strong>of</strong> the related industries, opens some new services<br />
industries and spawns a number <strong>of</strong> new “edge industries”.<br />
At the same time, information service industry has shown<br />
a s<strong>of</strong>tening effect on solving the employment. The<br />
pr<strong>of</strong>essional tendency <strong>of</strong> modern IT services industry<br />
reduces some pressure, creates new jobs and makes many<br />
new careers stand out.<br />
(c) The Effect <strong>of</strong> IT Services Development on the<br />
Service <strong>of</strong> the Related Industries<br />
The function <strong>of</strong> information services is embodied in<br />
the reform <strong>of</strong> traditional industrial information and the
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1807<br />
information technology support <strong>of</strong> the related industries.<br />
As for the traditional industrial departments, the<br />
development <strong>of</strong> IT services industry can raise developing<br />
efficiency <strong>of</strong> the new products, the technological content<br />
and value-added products. As for the construction<br />
industry, IT services like modern engineering consulting<br />
services should be developed vigorously, to s<strong>of</strong>ten the<br />
industrial industries and to achieve business value-added.<br />
In addition, to develop vigorously IT services industry<br />
can not only raise the added value level <strong>of</strong> the third<br />
industry but also update the traditional means <strong>of</strong> service<br />
industry.<br />
IV. TENDENCY ANALYSIS OF PRESENT IT SERVICES<br />
STATUS<br />
Although late beginning, with the development <strong>of</strong> IT<br />
and its wide use in various fields, China’s IT services<br />
industry now goes into the rapid growth stage, having a<br />
certain scale and achieving substantial results. The scale<br />
<strong>of</strong> s<strong>of</strong>tware industry continues to grow. IT services,<br />
represented by communications industry, has been<br />
developing rapidly, the development and application <strong>of</strong><br />
information resources having made positive progress,<br />
outsourcing service becoming a new highlight, some<br />
excellent digital content products having been created,<br />
information services promoting information technology<br />
having preliminary breakthrough and leading to upgrade<br />
<strong>of</strong> optimization. But, compared with the international<br />
advanced level, China’s information technology industry<br />
is still in lower level. Specific development status and<br />
future trends show as follows.<br />
A. Analysis <strong>of</strong> Information Technology Services<br />
(a) Unbalance <strong>of</strong> the Development Shown from Market<br />
Structure, Low Level <strong>of</strong> Market Development<br />
Unbalanced development <strong>of</strong> information technology<br />
services is mainly reflected in the unbalance <strong>of</strong> balance<br />
industrial structure and the regional development. In<br />
recent years, the information service level <strong>of</strong> information<br />
service industry has been obviously improved, but, the<br />
entire IT services market presents low-level trend. At the<br />
same time, the development <strong>of</strong> information services level<br />
is vigorously unbalanced. The information service<br />
industry level <strong>of</strong> cities like Shanghai, Shenzhen,<br />
Guangzhou is far above the China’s average level. As for<br />
the present, the communication industry in China<br />
accounts for over half <strong>of</strong> IT services industry, but, the<br />
s<strong>of</strong>tware, technical services and information content<br />
services accounts for miner proportion. The development<br />
is unbalanced. In addition, due to the influence <strong>of</strong> many<br />
aspects such as its own conditions and external<br />
environment, the overall strength <strong>of</strong> IT services industry<br />
in China is low and small. The large groups with<br />
internationally competitiveness are too limited.<br />
(b) Developing Fast Scale, a Big Gap to the<br />
International Advanced Level, Belonging to Growth<br />
Stage<br />
Viewing from the revenue <strong>of</strong> the whole industry, the<br />
growth tendency <strong>of</strong> China’s IT services is great. Although<br />
China’s IT services industry, especially, modern IT<br />
© 2011 ACADEMY PUBLISHER<br />
services industry has made great progress, the proportion<br />
in the national economy and services industry is still low.<br />
Compared to the developed countries the overall level<br />
falls behind twenty to thirty years. In the global share <strong>of</strong><br />
modern IT services market, IT services market scale <strong>of</strong><br />
the United States, Western Europe and Japan altogether<br />
accounts for 90 percent <strong>of</strong> the global market, while<br />
China’s IT services market share accounts for only less<br />
than 10%. The gap is obvious.<br />
(c) Falling Behind Relatively in Information Resources<br />
Construction, the Information Infrastructure Not Being<br />
Sound Enough<br />
Not enough is the input <strong>of</strong> China’s database<br />
information systems infrastructure. Nor is the<br />
development strength <strong>of</strong> information resources. For<br />
example, there are many shortcomings in database<br />
construction: small quantity <strong>of</strong> database construction,<br />
small capacity, lack <strong>of</strong> database product with high<br />
quality, present database information resources focusing<br />
on hardware neglecting s<strong>of</strong>tware, slow updating speed,<br />
contents focusing on science and technology and difficult<br />
to meet the market requirements.<br />
B. Analysis on Future Development Trend <strong>of</strong> IT Services<br />
Industry<br />
(a) The Traditional IT Hardware Manufacturers<br />
Transformation to S<strong>of</strong>tware and Services<br />
Many well-known enterprises <strong>of</strong> today’s IT<br />
manufacturing fields in the whole world see IT services<br />
as a key development. Some manufacturers like<br />
internationally renowned electronic data processing<br />
companies successfully realize transformation from the<br />
traditional hardware manufacturers to s<strong>of</strong>tware and<br />
hardware manufacturers and service. In the Unite States,<br />
the largest feature <strong>of</strong> IT services industry is meanwhile<br />
providing the integrated IT services <strong>of</strong> hardware and<br />
s<strong>of</strong>tware. The largest project in Japan’s IT services<br />
industry is s<strong>of</strong>tware development services, which<br />
accounts for 60% <strong>of</strong> total sales. While, now in China, IT<br />
services industry should do better to prepare for strategic<br />
transformation, leading it to transformation <strong>of</strong> s<strong>of</strong>tware<br />
and services in developing hardware manufacturing<br />
industry and realizing IT service industry’s convergence<br />
with international standards.<br />
(b) Adaping International Trends, Realizing<br />
Outsourcing <strong>of</strong> S<strong>of</strong>tware and Services<br />
The local industrial cluster formed and developed<br />
under the global background is the carrier <strong>of</strong> the regional<br />
and global economy. The IT development makes the<br />
industrial cluster theory take on a revolutionary change.<br />
The border region and national boundary and therefore,<br />
conduct development, production and sales activities in<br />
different areas. S<strong>of</strong>tware outsourcing, compared with the<br />
previous one, in terms <strong>of</strong> scale or in the related content,<br />
has presented unprecedented differentiation and<br />
development. Shown as follows, first, scale enlarging and<br />
energy degree increasing. To ensure the quality and<br />
speed, the contractor establishes more and more large<br />
development centers where the party contracting locates.<br />
Second, diversification <strong>of</strong> outsourcing situation, there are<br />
two types: direct and indirect sub-contracting. The United
1808 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
States usually adopts direct contracting. Japanese<br />
s<strong>of</strong>tware enterprises adopt indirect subcontracting. In the<br />
initial stage <strong>of</strong> IT services development, China has<br />
become contractors <strong>of</strong> developed countries.<br />
(c) Highlighting IT Services Development <strong>of</strong> key<br />
Areas, Emphasizing Its Services to Producers<br />
All countries emphasize highlighting IT services<br />
development <strong>of</strong> key areas and put forward specific plans.<br />
The United States puts the emphasis <strong>of</strong> IT services<br />
industry development on s<strong>of</strong>tware industry. Japan puts<br />
the emphasis <strong>of</strong> IT services industry development on<br />
s<strong>of</strong>tware industry on e-government and e-commerce.<br />
While China should combine its own characteristics and<br />
simultaneously develop IT services industry by drawing<br />
on foreign development mode. When the manufacturing<br />
industry developing to a certain stage, it is bound to<br />
transfer to the productive services industry <strong>of</strong> high pr<strong>of</strong>it<br />
and high value-added such as research and development;<br />
consulting and design. It is bound to generate the demand<br />
for the productive IT services. At the same time, IT<br />
services industry gradually takes on the developing trends<br />
<strong>of</strong> the “outside” and “pr<strong>of</strong>essional”. Now it becomes<br />
necessary requirements <strong>of</strong> China’s economic<br />
development to a certain period how to adapt to the new<br />
situation <strong>of</strong> economic development and to accelerate IT<br />
services industry.<br />
(d) Realizing the Integration <strong>of</strong> Consulting and IT<br />
Services Industry, Providing Pr<strong>of</strong>essional Value-Added<br />
IT Services<br />
The integration <strong>of</strong> consulting and IT services industry<br />
not only promotes revolutionary breakthrough <strong>of</strong><br />
consulting industry but also the development <strong>of</strong><br />
information services industry. Information value-added<br />
services is a service mode that uses market means to<br />
increase the asset value <strong>of</strong> the information service by<br />
classifying, processing, arranging and analyzing a large<br />
number <strong>of</strong> original information, aiming at different<br />
customer’s demands and features. Information valueadded<br />
service is one <strong>of</strong> the developing trends <strong>of</strong><br />
information services industry. Today, in such a colorful<br />
information contents, only to make good use <strong>of</strong><br />
information resources to improve the technology <strong>of</strong> using<br />
information, to satisfy customer’s various demands and<br />
keep competitive forces. Moreover, the added value<br />
services <strong>of</strong> production largely reflects the pr<strong>of</strong>essional<br />
level <strong>of</strong> service, so, the needs for the pr<strong>of</strong>essional<br />
information service is IT services industry developing to<br />
a certain stage.<br />
V. PROBLEMS EXISTED IN CHINA’S IT SERVICES<br />
INDUSTRY DEVELOPMENT<br />
Because <strong>of</strong> late start, there is a big gap compared to the<br />
international advanced level. IT services industry is<br />
developing very rapidly in recent years, but, whether<br />
from the external macro environment <strong>of</strong> IT services<br />
industry or from the internal factors <strong>of</strong> micro aspects,<br />
there are some problems and shortcoming, affecting IT<br />
services industry to develop better.<br />
© 2011 ACADEMY PUBLISHER<br />
A. Macro Aspects<br />
(a) Imperfect Legal System <strong>of</strong> IT Services Industry<br />
At present, in the rapid development <strong>of</strong> IT services<br />
industry, there is no corresponding laws and regulations<br />
<strong>of</strong> IT services industry. The serious delay <strong>of</strong> policy and<br />
regulation leads to the irregularities <strong>of</strong> information<br />
market operation. With the development <strong>of</strong> the times and<br />
IT, China’s laws and regulations <strong>of</strong> promoting consulting<br />
industry development made in the eighties and ninties <strong>of</strong><br />
last century were not suited to the demands <strong>of</strong> the Internet<br />
times and must accelerate replacement pace. The<br />
legislation <strong>of</strong> IT services industry in China has long been<br />
lacking a commanding basic law with a higher status.<br />
China’s legal system <strong>of</strong> existing information services<br />
technology is still in its infancy, lacking comprehensive<br />
legislation system and the clear legislative goals. The<br />
existing legislation is largely for the sector and local one<br />
with imcomplete system, low grade<br />
(b) Imperfect Management and Organization<br />
Coordination Mechanisms <strong>of</strong> IT Service Industry<br />
China’s information market, whether business income<br />
or employment maintains a fairly large growth every<br />
year, but lacks effective and integrated management. The<br />
structural integrity <strong>of</strong> information market system and the<br />
integrated development policy and planning depends on<br />
sound management system and organizational<br />
coordination mechanisms. As for our country, there aren’t<br />
any comprehensive, integrated and centralized leading<br />
departments. Such a new industrial service institution as<br />
IT services industry scatters various administrative<br />
departments <strong>of</strong> various fields and brings great difficulties<br />
to the organizational management. Meanwhile, the<br />
chaotic situation <strong>of</strong> higher-level managing departments<br />
and higher authorities fragmented many leaders and<br />
management disorder reduces the government’s control<br />
and operational efficiency. It is not conducive to the rapid<br />
development <strong>of</strong> the entire IT services industry.<br />
(c) Imperfect Market Operating Mechanism <strong>of</strong> IT<br />
Services Industry<br />
Information market operation mechanism based on IT<br />
services industry mainly consists <strong>of</strong> the price<br />
mechanism, competition mechanism and supply and<br />
demand mechanism. Now, information market <strong>of</strong> IT<br />
services industry in China is still at an early stage <strong>of</strong><br />
development. There exists a great defect in the operation<br />
mechanism <strong>of</strong> information market. First, it is difficult to<br />
determine reasonable price <strong>of</strong> information products and<br />
there exists a large subjective arbitrariness in the price<br />
determination, causing some disorder to the information<br />
product price <strong>of</strong> IT services industry. So, it is difficult to<br />
form reasonable price mechanism. Second, duo to the<br />
asymmetry <strong>of</strong> social information, resulting in unbalanced<br />
contradiction <strong>of</strong> information product supply and demand<br />
in IT services industry, there <strong>of</strong>ten appears the<br />
phenomena <strong>of</strong> some information products in short supply<br />
or oversupply. The contrast between supply and demand<br />
information is fairly larger. The supply and demand<br />
mechanism <strong>of</strong> IT services industry is still not mature.<br />
Third, the Information market <strong>of</strong> IT services industry<br />
lacks effective competitive mechanism. At present, IT
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1809<br />
competitive conditions between the services sectors are<br />
not mature. There exist many unequal competitive<br />
phenomena, lacking standardized competitive behavior in<br />
information services departments. Furthermore, IT<br />
services agencies themselves do not establish a sound<br />
internal operation mechanism.<br />
B. Micro Aspects<br />
(a) Lacking the Advanced Technology Management<br />
Personnel <strong>of</strong> Modern IT Services<br />
The rapid development <strong>of</strong> IT services industry depends<br />
on high-quality technical personnel and management<br />
personnel. This is a knowledge-intensive industry.<br />
Although there are a large number <strong>of</strong> qualified personnel,<br />
some <strong>of</strong> which is very excellent, generally speaking, the<br />
quality is universally low. The reason for this<br />
phenomenon is duo to the rapid expansion <strong>of</strong> China’s IT<br />
services market and the market is still in its infancy with<br />
a serious shortage <strong>of</strong> supply <strong>of</strong> pr<strong>of</strong>essionals, (especially<br />
lacking the senior personnel <strong>of</strong> market management and<br />
international management.) Moreover, the staff currently<br />
has no unified system <strong>of</strong> accreditation and assessment<br />
with the uneven quality. There universally exists the<br />
phenomenon that expertise is obvious, but they lack the<br />
awareness and experience <strong>of</strong> the modern <strong>of</strong> consultation.<br />
At the same time, many excellent personnel <strong>of</strong><br />
information services flows to developed countries and<br />
domestic foreign-funded enterprises, especially high-level<br />
brain drain is serious.<br />
(b) Information Products and Service Standards Are<br />
Not Unified, Producer Services and Information Value-<br />
Added Services Expecting Further Development<br />
It is difficult to form a united definition to the standard<br />
that information services s<strong>of</strong>tware products provided by<br />
IT services industry to intangible products, producing the<br />
result that there are no uniform rules and standards in the<br />
products and services provided by IT service in our<br />
country. But, with the development <strong>of</strong> IT, the productive<br />
services raise the requirements <strong>of</strong> standardization and<br />
specialization, the degree <strong>of</strong> outside gradually increasing.<br />
At the same time, with the “services” <strong>of</strong> manufacturing<br />
industry and the adjustment <strong>of</strong> pr<strong>of</strong>it nodes in industrial<br />
chain, the world’s manufacturing industry is conducting a<br />
big strategic shift to China, causing the growing demand<br />
for the information services <strong>of</strong> production and value<br />
added services.<br />
(c) Information Language Not Compatible with<br />
Technology<br />
That information technology is not compatible with the<br />
language makes all the system difficult to communicate<br />
and link, resulting in language barriers between systems.<br />
Duplication <strong>of</strong> existing databases and separation from the<br />
market demand is mainly due to the reason that many<br />
information industry sectors haven’t implemented the<br />
national standards and international standards, causing<br />
irregular organization <strong>of</strong> information resources and<br />
seriously hindering the progress <strong>of</strong> information industry<br />
and network.<br />
(d) Poor Information Resources and Poor Service<br />
Chinese information in China’s online is too small and<br />
the information resources are poor so that it is difficult to<br />
© 2011 ACADEMY PUBLISHER<br />
meet the needs <strong>of</strong> many customers. Many information<br />
technology services agencies have financial problem,<br />
most <strong>of</strong> which only conduct a simple collection and<br />
accumulation and don’t deeply analyze or evaluate the<br />
information. So, the short-term behavior <strong>of</strong> research leads<br />
to the poor information services.<br />
(e) Insufficient Capacity <strong>of</strong> Independent Innovation<br />
At present, the innovation capacity China’s<br />
information service enterprise has been considerably<br />
improved, but there is still a big gap compared to<br />
developed countries, which is mainly reflected in the<br />
inadequate innovation capacity such as core technology,<br />
business development, products development technology<br />
and service model. Because <strong>of</strong> the basic institutions and<br />
core technology lying in the hands <strong>of</strong> international<br />
companies, controlled by others in core technology; due<br />
to inadequate protection <strong>of</strong> business ideas in the<br />
innovation <strong>of</strong> the service model, the copying<br />
phenomenon <strong>of</strong> information services being universally<br />
serious; difficult to grasp customer needs in business<br />
innovation, making few original innovation <strong>of</strong><br />
information service products, the most being imitation<br />
and lead innovation; the homogenization phenomenon <strong>of</strong><br />
business and application being serious in products<br />
development.<br />
VI. RELATED COUNTERMEASURES ON CHINA’S<br />
DEVELOPMENT OF MODERN IT SERVICES<br />
INDUSTRY<br />
A. Macro Aspects<br />
(a) Optimize the Policy Environment <strong>of</strong> Information<br />
Services Development, Further Improving the Legal<br />
System<br />
Some key areas <strong>of</strong> China’s IT services are still in the<br />
early stages <strong>of</strong> development. The information industry<br />
must rely on the state’s information policy and the<br />
corresponding legal protection and need the strong<br />
support <strong>of</strong> government and the society. Government<br />
departments should improve the legislative and legal<br />
system <strong>of</strong> IT services. The first is to make and improve<br />
the legislative and legal system closely related to IT<br />
services industry such as “Information Services<br />
Management Ordinance”, “Information Law”, “database<br />
revitalization Law”, “Government Information Resources<br />
Management Regulations”, and s<strong>of</strong>tware standards,<br />
information standards, network technology standards and<br />
those related to IT standards and regulations <strong>of</strong> IT<br />
services market. The second is to make the public law <strong>of</strong><br />
government information. Because <strong>of</strong> no reliable law,<br />
there are the contradictory relationships between<br />
information resource sharing and safety security, the<br />
public law <strong>of</strong> government development and use <strong>of</strong> no<br />
confidential information resources. The last is to build<br />
and improve the related market-oriented industrial policy,<br />
optimizing the external environment <strong>of</strong> IT services<br />
development such as financial, fiscal and tax policy.<br />
(b) To Further Establish and Improve a Strong<br />
Administration <strong>of</strong> Information Technology Services
1810 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
IT services development goes deep into all fields <strong>of</strong><br />
society, to change the present situation <strong>of</strong> no unified<br />
industrial management developments against managing<br />
disorder and many leaders, it is necessary to establish and<br />
perfect a strong national administration <strong>of</strong> information<br />
services so as to implement effective unified management<br />
to the IT services industry and coordinate the business<br />
between the competent departments, to formulate<br />
development strategy <strong>of</strong> IT services industry. At the same<br />
time, it is necessary for the industrial institutions to give<br />
play to the function <strong>of</strong> standard and self-discipline on IT<br />
services industry, learning form the U.S. management<br />
model <strong>of</strong> IT services industry and gradually forming a<br />
line with the management system <strong>of</strong> socialist market<br />
economic system.<br />
(c) Perfecting Operational Mechanism <strong>of</strong> Information<br />
Market<br />
Information, as a social resource, depends on the<br />
information market for allocate. The market <strong>of</strong><br />
information transaction and the commercialization <strong>of</strong><br />
information is very important way to promote the<br />
development <strong>of</strong> IT services industry. Only by establishing<br />
a unified, open, competitive and orderly information<br />
market system, constantly improving the operational<br />
mechanism <strong>of</strong> the information services market, can the<br />
sound development <strong>of</strong> information technology services<br />
industry be promoted. The main areas are as follows:<br />
first, using the market-oriented mode <strong>of</strong> operation to<br />
perfect supply and demand mechanism, putting the<br />
production and sales process <strong>of</strong> IT services industry into<br />
the orbit <strong>of</strong> market operation; second, by the<br />
improvement <strong>of</strong> the price mechanism, playing its full role<br />
in the adjustment <strong>of</strong> the market, balancing supply and<br />
demand relationship <strong>of</strong> information products in the<br />
operation <strong>of</strong> information services markets; Finally, by<br />
improving the competitive mechanism, having each<br />
sector <strong>of</strong> information services form survival <strong>of</strong> the fittest<br />
and a mutual competition in the information services<br />
market, forming a prosperous market <strong>of</strong> IT services.<br />
B. Micro Aspects<br />
(a) Raising Pr<strong>of</strong>essional Service Level, Training and<br />
Introducing Pr<strong>of</strong>essional Personnel on Information<br />
Services<br />
IT services industry belongs to knowledge-intensive<br />
industries and the pr<strong>of</strong>essional service level it provides<br />
depends on the practitioners’ knowledge reserve and the<br />
pr<strong>of</strong>essional level. At present, our country lacks the<br />
pr<strong>of</strong>essional personnel <strong>of</strong> IT services, particularly in IT<br />
services industry and other industries <strong>of</strong> mutual<br />
penetration. The market needs a large number <strong>of</strong><br />
pr<strong>of</strong>essional personnel and puts forward higher<br />
requirements to the personnel. Therefore, it is extremely<br />
important to introduce and train the qualified pr<strong>of</strong>essional<br />
<strong>of</strong> IT services by variety <strong>of</strong> measures improve the<br />
gathering space <strong>of</strong> high-end qualified personnel <strong>of</strong><br />
information services, to raise the pr<strong>of</strong>essional level and to<br />
promote the development <strong>of</strong> information technology<br />
services. First, to pay attention to “talent cultivation”<br />
strategically, to implement national strategy planning <strong>of</strong><br />
the qualified personnel <strong>of</strong> IT services industry and strive<br />
© 2011 ACADEMY PUBLISHER<br />
to create a large number <strong>of</strong> pr<strong>of</strong>essional personnel <strong>of</strong> high<br />
skill and high-level, familiar with modern IT services<br />
industry <strong>of</strong> international regulation and management.<br />
Second, full implementation vocational qualification<br />
certificate system, strengthening vocational job training,<br />
improving the basic quality <strong>of</strong> information services<br />
practitioners. Third, introducing domestic and<br />
international qualified personnel by many forms and<br />
channels and retaining them by establishing effective<br />
incentives and compensation mechanisms. Fourth,<br />
training high-intermediate information services personnel<br />
strengthening subject construction related to information<br />
services in colleges and universities so as to adapt to the<br />
needs <strong>of</strong> the rapid development <strong>of</strong> modern information<br />
service.<br />
(b) Full Developing and Using Information Resources<br />
Speeding Up Construction <strong>of</strong> Various Types <strong>of</strong> Databases<br />
Information service providers should focus on the<br />
development and utilization <strong>of</strong> information resources.<br />
Therefore, to fully develop and use information<br />
resources, it is necessary to build basically complete<br />
market demand-oriented databases and databases with<br />
local features, to mainly develop all kinds <strong>of</strong> the public<br />
and commercial databases serving the society, to adopt<br />
the principle <strong>of</strong> unified management, building databases<br />
in different places, resource sharing and multi-service,<br />
gradually forming a large practical database network.<br />
(c) Forming New Economic Growth Points, Grasping<br />
Innovation Project<br />
Innovation is the soul <strong>of</strong> information services<br />
development and the theme <strong>of</strong> a new era. In today’s<br />
network environment such as information visualization,<br />
digital, commercialization and globalization <strong>of</strong><br />
information, it is imperative to use modern information<br />
technology to establish IT service innovation projects<br />
with high-speed information network as its main body.<br />
Therefore, innovation information management <strong>of</strong><br />
system must be carried out well, information<br />
management, information systems, information services<br />
and information application technology to further<br />
promote better development <strong>of</strong> IT services industry.<br />
REFERENCES<br />
[1] ZHAN Jing. Innovation and Development <strong>of</strong> 21st<br />
Information Service Industries in Our County. Modern<br />
Library and Information Technology, 2002.<br />
[2] LI Jian-Ge. Developing Strategies <strong>of</strong> Modern Information<br />
Service Industries. China Information Fields, 2007.<br />
[3] WANG You-Gang. Research on Development Mode <strong>of</strong><br />
Information Service Industries. Industry Analysis, 2005.<br />
[4] GUO Dong-Qiang. Current Situation and Solution <strong>of</strong><br />
Development Information Service Industries in Our<br />
County. Market Weekly, 2008.<br />
[5] CAO Kuan-Zeng. Research on Development strategy <strong>of</strong><br />
information Service Industries in Our County. Practice<br />
Research, 2003.<br />
[6] LI Jian-Ge. Development strategy <strong>of</strong> Modern Information<br />
Service Industries. China Information Industry, 2007.<br />
[7] KUANG Pei-Yuan. Information Services: Definition and<br />
Statistical framework. Statistics Education, 2009.
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1811<br />
[8] WANG Xin. Development Mechanism and Measurement<br />
Theory <strong>of</strong> Information Industries and Methods. Jilin<br />
University PhD thesis. 2008.<br />
[9] YANG Xiang-Ming. Some Thoughts on Development <strong>of</strong><br />
Information Service Industries in China. Library Theory<br />
and Practice, 2007.<br />
[10] HOU Fu-Li. Research on Current Situation and Solution <strong>of</strong><br />
Modern Information Service Industries Development.<br />
Group Economic Research, 2007.<br />
© 2011 ACADEMY PUBLISHER<br />
Wei Gao (1964- ). Female, Han Dynasty, Nantong Jiang Su,<br />
College <strong>of</strong> Statistics and Applied Mathematics, Anhui<br />
University <strong>of</strong> Finance and Economics Vice Pr<strong>of</strong>essor. Research<br />
Field: Quantity Economics.<br />
E-mail: gaowei.64@163.com, Mobile Phone: 15805525532.<br />
Feng Wang (1962- ). Male, Han Dynasty. Nantong Jiang Su,<br />
Anhui University <strong>of</strong> Finance and Economics. Vice Pr<strong>of</strong>essor.<br />
Research Field: National Economiy<br />
Li Wang (1984- ) Male, Handan, Bank <strong>of</strong> Communication<br />
Hebei. Research Field: Quantity Economics.
1812 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
A Web Survey Program Based on Computer<br />
Technology and Its Application to Evaluation<br />
Model about Youth Self-organizations in China<br />
Ma-lin Song<br />
School <strong>of</strong> Statistics and Applied Mathematics, Anhui University <strong>of</strong> Finance and Economics, Anhui Bengbu, China<br />
Email: songmartin@163.com<br />
Tong Yang and Ya-qing Song<br />
Anhui University <strong>of</strong> Finance and Economics, Anhui Bengbu, China<br />
Abstract—The network has become the second space for<br />
people in China, and network and youth self-organizations<br />
based on web-platform have influenced young people more<br />
than ever before. From the viewpoint <strong>of</strong> the overall<br />
development <strong>of</strong> youth and building a harmonious society,<br />
it’s an important thing to reduce the negative influence <strong>of</strong><br />
the network and strengthen the sustainable development <strong>of</strong><br />
media ecology. The paper forecasts the developmental trend<br />
<strong>of</strong> adolescents by analyzing their current situation in China<br />
and builds the evolution model for youth self-organizations.<br />
This web survey program uses the IIS web server +<br />
ASP.NET service + SQL Server database. Survey.aspx<br />
could be generated in the server dynamically, so the web<br />
survey program can be achieved by computer. Finally, the<br />
paper suggests some advices to eliminate the negative effects<br />
<strong>of</strong> internet and to strengthen youth self-organizations.<br />
Index Terms—youth self-organizations; internet media; grey<br />
forecasting model; analytic hierarchy process; Web Survey<br />
program<br />
I. INTRODUCTION<br />
Three decades after reform and opening up, China has<br />
undergone enormous changes. Amateur live <strong>of</strong> Chinese<br />
people, young people in particular, has increased more<br />
rich and varied. In recent years, with progress in science<br />
and technology, communications, and popularization <strong>of</strong><br />
the Internet, the network has become the second largest<br />
human space, by which the impact <strong>of</strong> it on youth is<br />
growing. Although network have brought great<br />
convenience for our times, the deterioration <strong>of</strong> its<br />
environment, such as network information pollution,<br />
network security crisis, private space crisis <strong>of</strong> network,<br />
the shortage and expansion <strong>of</strong> networks information, also<br />
seriously endangers the physical and mental health <strong>of</strong><br />
youth. Therefore, from the perspective <strong>of</strong> the overall<br />
development <strong>of</strong> young people themselves, or for building<br />
a harmonious society, to consider how to reduce and<br />
eliminate the negative impact <strong>of</strong> network on youth and<br />
strengthen ecological civilization construction <strong>of</strong><br />
network, has increasingly become an important issue to<br />
be settled urgently.<br />
© 2011 ACADEMY PUBLISHER<br />
doi:10.4304/jcp.6.9.1812-1818<br />
II. LITERATURE REVIEW<br />
An Guoqi, Deng xiquan and Cao Kai (2006) pointed<br />
out that the Government has to face up to<br />
non-governmental organizations and the role <strong>of</strong> youth,<br />
and <strong>of</strong>ficial organizations ought to take positive measures<br />
to guide and monitor the non-governmental organizations<br />
and effective role <strong>of</strong> youth [1]. Ma Chunlei (2007)<br />
considered that self-organizing system is still beyond our<br />
traditional work, by which the formation <strong>of</strong> its social<br />
forces deserves our attention and research particularly<br />
[2]. Shi Guoliang (2007) thought that youth<br />
organizations, especially the informal youth<br />
organizations, are increasingly becoming a social<br />
organization that have rapid development, strong vitality,<br />
increasing cohesiveness, and influence [3]. Xu Rong and<br />
Zheng Chen (2007) suggested an educational<br />
management method that the active roles <strong>of</strong> informal<br />
organizations in students ought to be played and their<br />
negative effects should be controlled [4].<br />
Zeng hong considered we should concern with the<br />
composition and behavior characteristics <strong>of</strong> Internet users<br />
primarily for how to design web survey program.<br />
According to the Chinese Internet Network Development<br />
survey data in Chinese Internet Network Information<br />
Center (CNN IC), he made a quantitative analysis in the<br />
composition and behavior characteristics <strong>of</strong> Chinese<br />
Internet users, and then discusses the network survey<br />
design effects [5].<br />
The paper achieves the web survey program through<br />
the IIS web server + ASP.NET service+ SQL Server<br />
database. Study on China’s youth self-organization based<br />
on media ecology perspective, this article suggests that<br />
ecological construction <strong>of</strong> young self-organization need<br />
to be strengthened and ecological environment <strong>of</strong> China’s<br />
internet media should be optimized to promote diversity,<br />
rationalization and ecology distribution to strengthen the<br />
full development <strong>of</strong> youth and harmonious society<br />
building.
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1813<br />
III. ANALYSIS OF INTERNET YOUTH USERS IN CHINA<br />
A. The status quo <strong>of</strong> internet Youth users in China<br />
With the advent <strong>of</strong> the information age, improvement<br />
<strong>of</strong> communication facilities and increasing <strong>of</strong> people's<br />
income level, the Internet is getting into millions <strong>of</strong><br />
households.<br />
The scale <strong>of</strong> Chinese internet user has showed the<br />
trend <strong>of</strong> sustained and rapid development, In June 2008,<br />
the number <strong>of</strong> Chinese Internet users is 4.52 times than it<br />
in June 2002. In June 2008, the number reached 25.3<br />
million, ranked first in the world. In June 2008, the<br />
number <strong>of</strong> Internet users under 24 years old is 4.15 times<br />
than it in June 2002. An increase from 18 to 24 year old<br />
Chinese Internet users is 3.50 times than it in June 2002.<br />
B. Trends forecast Youth Internet users<br />
In this paper, GM (1, 1) model is used to predict the<br />
size <strong>of</strong> China's young Internet users and Internet users.<br />
The gray system theory is proposed by Pr<strong>of</strong>essor Deng<br />
Julong, a China scholar, in the 1980's, which is used to<br />
control and prediction and is widely applied in<br />
agriculture, socio-economic and other fields [6]. In this<br />
paper, GM (1,1) model is used to forecast China's total<br />
Internet users and its change in the trend. The simulation<br />
model and the residual difference are shown in table I. As<br />
a result <strong>of</strong> p = 1.0000, c = 0.1942, the current model is in<br />
a very good level <strong>of</strong> prediction.<br />
TABLE 1 CHANGES OF THE TRENDS IN THE TOTAL NUMBER OF<br />
CHINESE NETIZENS<br />
Sequences<br />
Original<br />
value<br />
(0)<br />
x () i<br />
Predictive<br />
value<br />
(0)<br />
xˆ () i<br />
Residual<br />
errors<br />
(0)<br />
ε () i<br />
Relative<br />
errors<br />
(%)<br />
X(2) 5910.0000 5127.3129 782.6871 13.2434<br />
X(3) 6800.0000 5868.6508 931.3492 13.6963<br />
X(4) 7950.0000 6717.1758 1232.8242 15.5072<br />
X(5) 8700.0000 7688.3857 1011.6143 11.6278<br />
X(6) 9400.0000 8800.0191 599.9809 6.3828<br />
X(7) 10300.0000 10072.3790 227.6210 2.2099<br />
X(8) 11100.0000 11528.7046 -428.7046 -3.8622<br />
X(9) 12300.0000 13195.5944 -895.5944 -7.2813<br />
X(10) 13700.0000 15103.4933 -1403.4933 -10.2445<br />
X(11) 16200.0000 17287.2477 -1087.2477 -6.7114<br />
X(12) 21000.0000 19786.7426 1213.2574 5.7774<br />
X(13) 25300.0000 22647.6296 2652.3704 10.4837<br />
The predicted results <strong>of</strong> other variables (18 ~ 24 years<br />
<strong>of</strong> age the number <strong>of</strong> users) is available similarly, <strong>of</strong><br />
which the proportion <strong>of</strong> 18 to 24 years old <strong>of</strong> Internet<br />
users is get through each stage netizens divides total<br />
number.<br />
As can be seen through the forecast, the next three<br />
years the total number <strong>of</strong> Chinese Internet users and the<br />
number <strong>of</strong> Internet users <strong>of</strong> 18 - 24 years old will<br />
© 2011 ACADEMY PUBLISHER<br />
continue to increase. In June 2011, the number reached<br />
13.893 million, the Chinese youth will account for about<br />
27.24 percent <strong>of</strong> China's total Internet users.<br />
It is generally believed that China's rapid development<br />
<strong>of</strong> Internet network bring about opportunities for the<br />
youth self-organizations’ flourish. The Internet goes into<br />
millions <strong>of</strong> households, in which its fashion and<br />
convenience attract a lot <strong>of</strong> young people involved.<br />
Internet provides equality, freedom, easy platform<br />
exchange for young people’s activities, by which it<br />
brought more opportunities for the formation <strong>of</strong><br />
self-organizations. Low-cost <strong>of</strong> Internet resources’<br />
network configuration also carries out facilitations for the<br />
formation <strong>of</strong> self-organizations’ establishment,<br />
management and activities.<br />
IV. INVESTIGATION DESIGNATION OF YOUTH<br />
SELF-ORGANIZATIONS BASED ON INTERNET MEDIA<br />
PROSPECTIVE<br />
A. Index system for youth self-organizations evaluation<br />
and its quantification<br />
In this paper, four-level index system is used to<br />
evaluate youth self-organizations, in which the target<br />
level is youth self-organizations indicators index, criteria<br />
level includes eight indicators used to measure the<br />
members’ feeling <strong>of</strong> youth self-organizations, indicator<br />
level includes a total <strong>of</strong> 28 indicators and the last level<br />
mainly includes questionnaire design for indicator level.<br />
B. Determination Indicators’ Weights<br />
In this paper, evaluation system <strong>of</strong> indicators <strong>of</strong> youth<br />
self-organization is composed <strong>of</strong> the multi-level index<br />
cluster. It constructs judgment matrix structure after<br />
seeking the advices from experts and determines weigh<br />
by mathematical treatment in some forms. Therefore, this<br />
article will make it more scientific to combine qualitative<br />
and quantitative weigh determination by Analytic<br />
Hierarchy Process (AHP) [7].<br />
Analysis Hierarchy Structure are Constructed with<br />
Index System for the Calculation, which includes object<br />
layer A; rule hierarchy B1 ~ B8; individual indicators are<br />
just index hierarchy. After using “1 to 9 scales”,<br />
judgment matrix <strong>of</strong> the criteria to the objective is<br />
constructed. It is important to carry out Consistency test<br />
<strong>of</strong> judgment matrix and level-ranking, which can be seen<br />
in table II.<br />
TABLE 2 CONSISTENCY TEST OF JUDGMENT MATRIXES<br />
CI CR<br />
A 0.0721 0.0511<br />
B1 0.0018 0.0032<br />
B2 0 0<br />
B3 0.0652 0.0693<br />
B4 0.0853 0.0761<br />
B5 0.0193 0.0332<br />
B6 0.0198 0.0220<br />
B7 0.0193 0.0332<br />
B8 0.0046 0.0079
1814 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
Under such circumstances, the judgment matrixes <strong>of</strong><br />
the CR are less than 0.10, which can be considered sort <strong>of</strong><br />
single-level structure with consistency. As results,<br />
Hierarchy general ranking results are as follows:<br />
W ′ B1<br />
=[ 0.1427 0.0269 0.0506]′<br />
(1)<br />
W ′ B2<br />
= [ 0.0075 0.0075 0.0226]′<br />
(2)<br />
W ′ B3<br />
= [ 0.0041 0.0095 0.0160 0.0294]′<br />
(3)<br />
W ′ B4<br />
= [ 0.0480 0.0032 0.0060 0.0151 0.0205]′<br />
(4)<br />
W ′ B5<br />
=[ 0.0158 0.0390 0.0962]′<br />
(5)<br />
W ′ B6<br />
= [ 0.1904 0.0692 0.0316 0.1073]′<br />
(6)<br />
W ′ B7<br />
= [ 0.0107 0.0018 0.0043]′<br />
(7)<br />
W ′ B8<br />
=[ 0.0134 0.0041 0.0074]′<br />
(8)<br />
These matrixes from (1) to (8) are the corresponding<br />
weights <strong>of</strong> single indicators.<br />
V. EMPIRICAL ANALYSIS OF PROGRAM DESIGN<br />
A series <strong>of</strong> investigations are carried out in a university<br />
surrounding schools via the design <strong>of</strong> the questionnaire <strong>of</strong><br />
the authors, by which some youth self-organizations are<br />
known, 15 self-organizations being more influential and<br />
Internet-based, can be selected to be conducted a<br />
questionnaire survey on. The specific names <strong>of</strong><br />
self-organization are as follows: Economic Research<br />
Institute (Y01), Employment and Entrepreneurial<br />
Associations <strong>of</strong> University Students (Y02), Mutual<br />
Assistance Center <strong>of</strong> college students (Y03), Students<br />
Association <strong>of</strong> Financial Investment (Y04), Association<br />
<strong>of</strong> popular science <strong>of</strong> Students (Y05), Computer<br />
Association (Y06), Basketball Association (Y07), Table<br />
Tennis Union (Y08), Management Institute (Y09), Art<br />
Troupe <strong>of</strong> university students (Y10), Advertising Art<br />
Association (Y11), English Society (Y12), Green IN<br />
Society (Y13), Association <strong>of</strong> Wushu Enthusiasts (Y14)<br />
and Press Corps <strong>of</strong> university students (Y15). According<br />
to tests in among small proportion <strong>of</strong> their numbers, the<br />
revised final version, including 28 issues, is concluded.<br />
Because <strong>of</strong> limited space, the programs do not list; and it<br />
can be obtained from the author if necessary.<br />
Because <strong>of</strong> the difficulty in implement stabile retest<br />
reliability, most questionnaires use consistency reliability<br />
testing generally, in which reliability coefficient α is the<br />
most commonly used method. Cronbach α, being<br />
reliability coefficient, can be used for test <strong>of</strong> consistency.<br />
In general, α may be accepted if it is larger than 0.5. If the<br />
reliability coefficient is greater than 0.7, it means a very<br />
high reliability; when the range between 0.7 and 0.35, it<br />
means so-so; if it is less than 0.35, it means low<br />
reliability. Web survey can be carried out through emails,<br />
by which emails with questionnaire send.<br />
© 2011 ACADEMY PUBLISHER<br />
VI. ACHIEVEMENT OF INTERNET SURVEY PROGRAM<br />
A. Introduction<br />
This web survey program uses the IIS web server +<br />
ASP.NET service+ SQL Server database. Survey.aspx<br />
could be generated in the server dynamically.<br />
The entire program is divided into three layers: client<br />
layer, service layer and data layer. Client base in the<br />
surveyed users computers. The user could request<br />
survey.aspx page by IE browser. Service layer in the IIS<br />
web server, survey.aspx is generated in the IIS server<br />
dynamically and passed to the customer's IE browser.<br />
Data layer in the SQL Server database; all the issues<br />
involved in the web survey, survey and the user<br />
participated in the investigation are stored in the database.<br />
The hierarchical structure as shown in Figure 1:<br />
IIS 网页服务器 SQL<br />
Server<br />
数据库<br />
Figure 1 hierarchical structure chart<br />
B. Design<br />
For this program, the design includes class design in<br />
server and database design in data table.<br />
1. class design<br />
Mainly survey.aspx web pages generated dynamically.<br />
Survey.aspx initially only contains the user table<br />
information and the submit button, the specific details are<br />
shown in survey.aspx file. For different request survey<br />
class handle survey web pages and the result web pages<br />
dynamically generated.<br />
Survey class view is as follows:<br />
Figure 2 Survey class view<br />
In the class, Page_Load () function is executed for<br />
each user apply for survey.aspx page, according to<br />
different application parameters sub-function generated<br />
different pages dynamically. In order to build the survey<br />
and result web page in the web survey, we design two<br />
different classes: Submit Page Creator and Result Page<br />
Creator. Submit Page Creator class is used to generate the
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1815<br />
survey website; Result Page Creator class is used to<br />
generate the result page. Their class diagram as follows:<br />
Figure 3 Submit Page Creator and Result Page Creator class chart<br />
In Survey class, Submit_Click () function is executed<br />
when the user clicks on the submit button each time. This<br />
function collects the information in survey and user<br />
tables, to keep the information to the database.<br />
2. The design <strong>of</strong> the data table structure<br />
The network survey program involves 7 data table:<br />
ref_Organization, ref_QuestionType, ref_Question,<br />
ref_SurveyCatalog, Survey, Survey_Detail and User<br />
table.<br />
1) Ref_Organization table<br />
The table saves the information <strong>of</strong> survey; the table<br />
structure is defined as:<br />
Figure 4 Ref_Organization table structure chart<br />
ID is used to identify each organization; Name is the<br />
organization's name; Description gives a brief description<br />
<strong>of</strong> the organization.<br />
Ref_Question Type table<br />
The table holds the type <strong>of</strong> survey questions. The table<br />
structure is defined as:<br />
Figure 5 Ref_Question Type table structure chart<br />
Type is used to identify each problem type; Choices is<br />
all the alternative answers for the kind <strong>of</strong> problem<br />
(separated different answers by semicolon); Weights is<br />
scores for the corresponding optional answer (separated<br />
© 2011 ACADEMY PUBLISHER<br />
different scores by semicolon); Description is a brief<br />
description <strong>of</strong> such problems.<br />
Ref_Question table<br />
The table holds all the survey questions. The table<br />
structure is defined as:<br />
ID is used to represent each problem; Name gives the<br />
content <strong>of</strong> issue; Type specifies the type <strong>of</strong> problem;<br />
Description gives a brief description <strong>of</strong> the problem.<br />
Ref_Survey Catalog table<br />
The table holds all the web survey by the system<br />
launched. The table structure is defined as:<br />
CatalogID is used to distinguish different surveys;<br />
Questions gives all the problems in the survey (separated<br />
different issues by semicolon); Organizations gives all the<br />
surveyed organizations involved in the survey (separated<br />
different organizations by semicolon); Valid Survey<br />
Count represents the total number <strong>of</strong> all valid<br />
questionnaire in the survey; Total Survey Count<br />
represents the total number <strong>of</strong> all submitted questionnaire<br />
in the survey; Descriptions gives a brief description for<br />
this questionnaire.<br />
2) Survey table<br />
Figure 6 Ref_Question table structure chart<br />
Figure 7 Ref_Survey Catalog table structure chart<br />
The table participates in the questionnaire submitted by<br />
the user each time through save the system. The table<br />
structure is defined as:<br />
Figure 8 Survey table structure chart
1816 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
ID is used to identify this survey questionnaire;<br />
CatalogID represents the questionnaire belongs to which<br />
network survey; UserID represents the user ID who<br />
submitted the questionnaire; Time represents the<br />
submitted time <strong>of</strong> questionnaire; Valid indicates the<br />
validity <strong>of</strong> this questionnaire; SourceIP represents the IP<br />
address submitted to the client in the questionnaire.<br />
3) Survey_Detail table<br />
The table holds each question and participates in the<br />
user's choice in the entire questionnaire. The table<br />
structure is defined as:<br />
SurveID identifies each specific survey item in the<br />
questionnaire. QuestionID gives the question identity<br />
involved in this investigation; OrganizationID gives the<br />
organization identity involved in this investigation;<br />
Record shows the results <strong>of</strong> the survey items.<br />
4) User table<br />
Figure 9 Survey_Detail table structure chart<br />
The table holds the detail <strong>of</strong> the involved user. The<br />
table structure is defined as:<br />
ID is used to identify each user; Gender gives the user<br />
gender; AgeRange gives the age range <strong>of</strong> users; Domain<br />
gives the industry <strong>of</strong> user; Name gives the user's name;<br />
Email gives the e-mail <strong>of</strong> users; Address gives the contact<br />
<strong>of</strong> user; Comments gives additional user information.<br />
C. Demonstrate<br />
User need to provide id parameter to apply for<br />
survey.aspx, this parameter is used to distinguish<br />
different web survey. The IE browser displays as follows:<br />
© 2011 ACADEMY PUBLISHER<br />
Figure 10 User table structure chart<br />
Figure 11 Web survey chart<br />
In the table, the users make a choice for the overall<br />
impression <strong>of</strong> 15 self-organizations, options can be<br />
divided into very satisfied, satisfied, more satisfied, in<br />
general, less satisfied, dissatisfied, very dissatisfied.<br />
However, if the users have not participated in the<br />
self-organization, some <strong>of</strong> the problems they are not<br />
interested or do not know the answer, please do not<br />
answer.<br />
Finally, the user provides some necessary personal<br />
information.<br />
Figure 12 Personal information chart<br />
In the table, the users need to select the relevant<br />
information, including gender, age range, and<br />
pr<strong>of</strong>essional. If the users need the results <strong>of</strong> this survey,<br />
please provide name, address, zip code, and Email.<br />
After the user clicks the submit button, according to<br />
different user information, it will show different findings<br />
slightly. In the result page <strong>of</strong> web survey, for each<br />
question, page provides to the total users number <strong>of</strong><br />
answering the question and the current result. The result<br />
represents by the color section, the shorter color section,<br />
the closer to green, and the result is more close to the left<br />
<strong>of</strong> alternative answers in the list.
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1817<br />
When the user to provide personal contact, the result<br />
page as shown below:<br />
In the table, we can see the web survey results for the<br />
overall impression <strong>of</strong> 15 self-organizations; the second<br />
column shows the number <strong>of</strong> result options.<br />
After the user provides personal contact, the results as<br />
shown below:<br />
The table shows that the latest survey results will be<br />
sent to the user’s e-mail or postal address.<br />
VII. CONCLUSIONS<br />
Different types <strong>of</strong> self-organizations have different<br />
impacts on the growth <strong>of</strong> young people, while young<br />
people also have their own objective assessment <strong>of</strong><br />
different self-organizations. The diversification,<br />
co-existence and symbiosis <strong>of</strong> self-organizations will<br />
enrich the lives <strong>of</strong> young people and promote the<br />
comprehensive development <strong>of</strong> youth. At present,<br />
network environment should be optimized to provide<br />
good platforms for young and healthy development <strong>of</strong><br />
self-organization.<br />
© 2011 ACADEMY PUBLISHER<br />
Figure 13 Web survey results chart<br />
Figure 14 Web survey results chart<br />
The development <strong>of</strong> Networks is breaking down the<br />
temporal and spatial boundaries <strong>of</strong> ideological and<br />
political education and provides new opportunities for<br />
further strengthening <strong>of</strong> educational influence.<br />
Through the building <strong>of</strong> network platform, new moral<br />
space could be opened up. Strengthening the building <strong>of</strong><br />
communication channels can help students explore a<br />
variety <strong>of</strong> ideological confusion or communication issues<br />
freely with their teachers, parents and students and keep<br />
abreast <strong>of</strong> all kinds <strong>of</strong> information in society, by which<br />
expectations <strong>of</strong> communities, school and parents are<br />
conducted together through the network society.<br />
These will cherish the transfer for students to increase<br />
the original space <strong>of</strong> narrow education into whole society<br />
and develop ideological and political education, which<br />
make the original lag content <strong>of</strong> ideological and political<br />
education into a more forward-looking for students. The<br />
timeliness <strong>of</strong> content makes room for the ideological<br />
and political education that be extended to the entire<br />
network. Publicity through the network, young people<br />
may know that indulging in online games is dangerous.<br />
Understanding, proper use <strong>of</strong> the Internet and ability to<br />
selection <strong>of</strong> useful information will enhance their own<br />
ability to resist information pollution.<br />
It is believed that through the integration <strong>of</strong> network<br />
information and practical resources, combining with the<br />
establishment <strong>of</strong> a good cultural atmosphere <strong>of</strong> the<br />
network, various types <strong>of</strong> self-organization will<br />
strengthen exchanges and cooperation among them.<br />
Meanwhile, vigorous, healthy and civilized organization<br />
activities could enrich the lives <strong>of</strong> amateurs <strong>of</strong> youth.<br />
ACKNOWLEDGMENT<br />
The authors wish to thank Yang Jie, from School <strong>of</strong><br />
Adult Education, Anhui University <strong>of</strong> Finance and<br />
Economics, Anhui Bengbu, China, for his help <strong>of</strong> the<br />
finish <strong>of</strong> this paper. This paper is supported by Supported<br />
by National Natural Science Funds <strong>of</strong> China for<br />
Innovative Research Groups (70821001), National<br />
Natural Science Foundation <strong>of</strong> China (70901069), Social<br />
Science Foundation <strong>of</strong> Ministry <strong>of</strong> Education <strong>of</strong> China<br />
(10YJC630208), Social Science Foundation <strong>of</strong> Anhui,<br />
China (AHSK07-08D25, AHSKF09-10D116,<br />
AHSK09-10D14), and Anhui Provincial Natural Science<br />
Research Project for Universities (KJ2011A001).<br />
REFERENCES<br />
[1] An Guoqi, Deng xiquan and Cao Kai. Research on the<br />
roles and development trends <strong>of</strong> contemporary youth in<br />
non-governmental organizations [J]. Youth Studies,<br />
2006(5): 3-5. (In Chinese)<br />
[2] Ma Chunlei. Status quo <strong>of</strong> youth non-governmental<br />
organizations and theirs guides [J]. China Youth Study,<br />
2007(11): 38-39. (In Chinese)<br />
[3] Shi Guoliang. Analysis <strong>of</strong> development trends <strong>of</strong> youth<br />
organizations in today's world [J]. China Youth Study,<br />
2007(12): 22-24. (In Chinese)<br />
[4] Xu Rong & Zheng Chen. College students' informal<br />
organizations and educational management<br />
countermeasures [J]. <strong>Journal</strong> <strong>of</strong> Ningbo Radio & TV<br />
University, 2007(2): 91-93. (In Chinese)
1818 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
[5] Zeng hong. See the program design <strong>of</strong> web survey from<br />
Chinese internet users features [J].Economic<br />
Issues,2006(1): 145-147.(In Chinese)<br />
[6] Deng Julong. Basic methods <strong>of</strong> gray system [M].Wuhan:<br />
Publishing House <strong>of</strong> Huazhong University <strong>of</strong> Science and<br />
Technology, 2006. (In Chinese)<br />
[7] Xiong li, liang Liang and Wang Guo-hua. Method research<br />
on selection and valuation <strong>of</strong> numeric scale in analytic<br />
hierarchy process [J]. Systems Engineering-theory &<br />
Practice, 2005(3): 72-79. (In Chinese)<br />
© 2011 ACADEMY PUBLISHER<br />
Malin Song, corresponding author, is a teacher in School <strong>of</strong><br />
Statistics and Applied Mathematics, Anhui University <strong>of</strong><br />
Finance and Economics, Bengbu, Anhui, China. His major field<br />
<strong>of</strong> study includes management <strong>of</strong> computer manufacturing<br />
enterprise, credit risk, strategic alliance and eco-industrial park<br />
(E-mail: songmartin@163.com). He is the corresponding author<br />
<strong>of</strong> this article.
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1819<br />
The Research on the Influencing Factors <strong>of</strong><br />
Financing Strategy <strong>of</strong> Woman Entrepreneurs in<br />
China<br />
Xiong Xiong<br />
College <strong>of</strong> Management and Economics, Tianjin University ,Tianjin , China<br />
Email: xxpeter@tju.edu.cn<br />
Rong Fu, Wei Zhang, Yongjie Zhang<br />
College <strong>of</strong> Management and Economics, Tianjin University,Tianjin, China<br />
Email: diana1228cn@yahoo.com.cn, weiz@tju.edu.cn, yjz@tju.edu.cn<br />
Abstract—Based on the data from the nationwide surveys <strong>of</strong><br />
SMEs in "China's Private Economic Research in 2002, this<br />
paper examines gender differences among Chinese<br />
entrepreneurs seeking financing pattern, including external<br />
and internal financing, and studies on the factors those<br />
affect women entrepreneurs’ financing strategies through<br />
theoretical analysis and model validation from the human<br />
capital and social capital perspective. We find that human<br />
capital and social capital have positive influence on seeking<br />
external financing. There is also some evidence that the<br />
impact in Administrative system may promote external<br />
financing in China.<br />
Index Terms—Women entrepreneurship, Human capital,<br />
Social capital, financing strategy<br />
I. INTRODUCTION<br />
With the development <strong>of</strong> economy, increasing women<br />
begin to start their own business, no matter in developed<br />
or developing countries. Although female<br />
entrepreneurship has drawn wide attention only in recent<br />
20 years, the female entrepreneurship developed rapidly,<br />
and women enterprise has become an important driving<br />
force for the global economic growth. According to the<br />
GEM report (2005), female entrepreneurship is booming<br />
worldwide, and more than a third <strong>of</strong> entrepreneurs are<br />
woman. According to GEM 2007, the Chinese women's<br />
entrepreneurial activity index was as high as 11.16%,<br />
higher than the global average. However, contrast to the<br />
entrepreneurial enthusiasm, woman enterprises rely more<br />
on self-accumulation <strong>of</strong> capital and develop relatively<br />
slow.<br />
For the research on women entrepreneurs’ financing<br />
strategy, many scholars believe that there are some<br />
Corresponding author: Yongjie Zhang.<br />
© 2011 ACADEMY PUBLISHER<br />
doi:10.4304/jcp.6.9.1819-1824<br />
Lin Xiong<br />
The Robert Gordon University ,Scotland, Aberdeen ,UK<br />
Email: l.xiong@rgu.ac.uk<br />
differences between male entrepreneurs and women<br />
entrepreneurs in financing patterns. Women<br />
entrepreneurs face more difficult in obtaining financing,<br />
and seem to have some specific financing [1]. Women<br />
face significant difficulties in external financing,<br />
particularly bank loans, venture financing. In the past 40<br />
years, the United States about 40% <strong>of</strong> the enterprises are<br />
owned or managed by women, but less than 5% <strong>of</strong> the<br />
venture capital invest in the women-led enterprises [2].<br />
In order to analyze this subject, scholars give various<br />
explanations from different points <strong>of</strong> view. On the supply<br />
side (behaviors <strong>of</strong> bankers and public funders), some<br />
scholars believe that women entrepreneurs encountered<br />
with some credit discrimination when seeking to external<br />
financing. Women were required higher interest rates and<br />
more additional conditions when applying for loan [3].<br />
Using the methods <strong>of</strong> experimental and qualitative<br />
analyze, Sara Carter et al [4] found that the loans lenders<br />
assess different conditions when dealing with the<br />
application <strong>of</strong> loan from male and female appliers. When<br />
the variables such as business industry, credit market<br />
structure are controlled, women business owners still<br />
have to pay more for the loan, and there is no evidence to<br />
prove that women enterprises are greater risk than men’s.<br />
On the demand side (behaviors <strong>of</strong> women<br />
entrepreneurs), Orser [5] believe that women<br />
entrepreneurs are not that eager for business growth, and<br />
usually consider more on risks. In the process <strong>of</strong> applying<br />
for bank loans, women business owners are more<br />
negative, although their applications will not be easier to<br />
reject by banks. Some scholars also attempted to explain<br />
this phenomenon from human capital and social capital<br />
perspective. Nancy [2] find that only education level <strong>of</strong><br />
women business owners has a significant impact on the<br />
choice <strong>of</strong> external equity financing strategy, while social<br />
capital is not directly affected.
1820 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
II. LITERATURE REVIEW<br />
Women entrepreneurs’ financing pattern has been<br />
noticed by many scholars recently. Some <strong>of</strong> them tried to<br />
study the issue from human capital perspective. By<br />
studying the Finnish company, Cooper et al. [6] found<br />
entrepreneurs’ education level have a significant impact<br />
on the enterprise's survival and development. Loscocco et<br />
al. [7] believe one <strong>of</strong> the key factors leading to the<br />
success <strong>of</strong> small businesses is the relevant industry<br />
experience. They found that women has disadvantage in<br />
this area, because the female entrepreneurs than male<br />
entrepreneurs usually have less relevant experience in<br />
specific industry. Bosma et al. [8] studied more than<br />
1,000 Dutch companies, and confirmed the industry<br />
experience prior to starting one’s own business has<br />
played an important role in pr<strong>of</strong>itability and growth <strong>of</strong><br />
small business. Bates [9] found that college-educated<br />
entrepreneurs has a higher rate <strong>of</strong> success <strong>of</strong> starting<br />
one’s own company than less educated entrepreneurs, and<br />
they apply for a loan from commercial banks easier.<br />
Fabowale et al [10] found that banks increased rejection<br />
rate <strong>of</strong> lending loan when women business owners had<br />
few management experience. Boden and Nucci [11]<br />
believe that women has less opportunity to accumulate<br />
human capital because <strong>of</strong> their lower payment and less<br />
management experience.<br />
Some <strong>of</strong> scholars also tried to carry out the research<br />
from social capital perspective. Uzzi’s [12] research<br />
shows that the strong and weak links <strong>of</strong> the network<br />
between companies and banks are favorable for applying<br />
loan, and obtaining lower interest rates. Higgins and<br />
Gulati [13] found that the more extensive networks<br />
business owners have, the more excellent investment<br />
banks willing to underwrite the company’s IPO. Shane<br />
and Stuart [14] found that venture capitalists seem to be<br />
more willing to invest in emerging companies those they<br />
are familiar with, especially the entrepreneurs who had<br />
previously sponsored or had close contacted before.<br />
Priscilla chu[15] and other scholars studied the Chinese<br />
entrepreneurs in Hong Kong and Canada, and found that<br />
social capital <strong>of</strong> entrepreneurs could provide an access to<br />
critical resources for enterprises’ development, such as<br />
market, technology, capital, knowledge. In addition,<br />
Tjosvold [16] also believe social capital could be used in<br />
getting support from government. Pearce and Robinson<br />
[17] also pointed out that in China, enterprises leaders<br />
usually set up long term relationship among political<br />
parties, administrative leaders and other business<br />
executives, and the relationship with government <strong>of</strong>ficials<br />
is <strong>of</strong>ten essential for industrial and commercial<br />
enterprises on business success. David [18] found that at<br />
present stage, both private entrepreneurs and managers <strong>of</strong><br />
state-owned enterprises in China all maintain close and<br />
good social relations with the government and the Party,<br />
which is one <strong>of</strong> important ways to obtain economic<br />
resources. On the other hand, Renzulli, Aldrich, and<br />
Moody [19] found that compared with men’s, most<br />
women’s networks are lack <strong>of</strong> diversity, which will<br />
hinder women entrepreneurs to identify entrepreneurial<br />
opportunities as well as search for scare resources. Just as<br />
© 2011 ACADEMY PUBLISHER<br />
Nancy [20] mentioned, the lack <strong>of</strong> diversity <strong>of</strong> female’s<br />
networks results from the lack <strong>of</strong> contact with those who<br />
control key resources and critical introducers in their past<br />
experience.<br />
III. RESEARCH QUESTIONS AND METHOD<br />
A.Research questions<br />
Several specific financing patterns <strong>of</strong> women-owned<br />
businesses and relative explanations have been identified<br />
in the previous section. In this review, we try to verify<br />
several hypothesis and interpret the reasons <strong>of</strong> women<br />
entrepreneurs’ financing strategies from human capital<br />
and social capital perspective.<br />
To identify human capital, we use 2 variables. ① the<br />
result <strong>of</strong> formal education – education level; ②<br />
knowledge gained from work experience and practice-<br />
years <strong>of</strong> work experience.<br />
To quantify social capital, we design 3 categories <strong>of</strong><br />
variables, including the type <strong>of</strong> social network, the scale<br />
<strong>of</strong> social network, and the intensity <strong>of</strong> social network.<br />
The network we surveyed are following 6 types <strong>of</strong><br />
network, the member <strong>of</strong> Deputy to People’s Congress,<br />
CPPCC National Committee member, the Federation <strong>of</strong><br />
Industry member, member <strong>of</strong> the Individual and Private<br />
Entrepreneur Association, Chinese Communist Party, and<br />
Democratic Party. We also investigated the highest duties<br />
and the levels <strong>of</strong> network in the first 4 networks. The<br />
scale <strong>of</strong> network is measured by the number <strong>of</strong><br />
membership. The greater the number <strong>of</strong> membership is,<br />
the larger network is. The intensity <strong>of</strong> network is<br />
measured by the following indicators: ① the cost spends<br />
on social activities in that year; ② average time spend on<br />
social activities weekly. The greater the value <strong>of</strong> these<br />
two indicators shows, the greater the intensity <strong>of</strong> the<br />
network is.<br />
Meanwhile, we have three hypotheses to verify in<br />
Chinese market.<br />
H1: There are gender differences when entrepreneurs<br />
choose financing strategy.<br />
H2: More human capital female entrepreneurs have,<br />
external financing are more likely to be used.<br />
H3: More social capital female entrepreneurs have,<br />
external financing are more likely to be used.<br />
The conceptual model could be established in<br />
following way:<br />
Figure.1 conceptual model <strong>of</strong> analysis
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1821<br />
B.The data and research method<br />
This study is based on the data from the 2002 China<br />
Private Economy Research survey, a national wide<br />
survey conducted by the CPC Central Committee United<br />
Front Work Department, China Federation <strong>of</strong> Industry<br />
and State Administration for Industry. To ensure data’s<br />
integrity, reliability, we use Micros<strong>of</strong>t Excel to eliminate<br />
invalid original survey for the preliminary screening.<br />
After deleting the data those are incomplete or<br />
obviously invalid data, we carry out mean value analysis<br />
According to above analysis, we obviously notice that<br />
men and women business owners have significant<br />
difference when choosing whether using external<br />
financing (Sig = 0.001), and averagely men prefers to<br />
external financing than women (Mean Male > Mean<br />
Female). However, the proportion <strong>of</strong> start-up capital in<br />
individual investors does not appear significant difference<br />
in genders.<br />
To identify the number <strong>of</strong> factors, and the correlation<br />
between the observed variables, we carried correlation<br />
between the observed variables, we carried out<br />
TABLE 4-1 SAMPLES INFORMATION<br />
Then we use SPSS17.0 statistics s<strong>of</strong>tware to process the<br />
data, and the main analysis methods include reliability<br />
and validity analysis, descriptive statistics analysis,<br />
ANOVA, exploratory factor analysis, Logistic regression.<br />
IV. EMPIRICAL ANALYSIS<br />
Before starting our analysis, we summarize basic<br />
information <strong>of</strong> our samples as follows, shown in table 4-1:<br />
Gender Education Age<br />
Type: Male<br />
No:2328<br />
89.9%<br />
Type: Female<br />
No:261<br />
10.1%<br />
Type No. % Type No. %<br />
Primary 49 2.1 Under 25 11 0.5<br />
Junior 434 18.6 26~35 301 12.9<br />
Senior 983 42.2 36-45 1024 44.0<br />
College 764 32.8 46-55 798 34.3<br />
Master 98 4.2 Above 55 194 8.3<br />
Primary 7 2.7 Under 25 0 0<br />
Junior 31 11.9 26~35 44 16.9<br />
Senior 117 44.8 36-45 122 46.7<br />
College 90 34.5 46-55 76 29.1<br />
Master 16 6.1 Above 55 19 7.3<br />
Table 4-2 ANOVA Result<br />
Gender Work experience Industry<br />
and ANOVA based on gender differences. The results are<br />
shown in Table 4-2:<br />
Type No. Percent Type No. Percent<br />
Male Less than 1 year 37 1.6 Manufacturing category ( Agriculture,<br />
Mining, Manufacturing, Geology,<br />
Construction, Electricity & Gas)<br />
1306 56.1<br />
2-5 years 142 6.1 Service category(Food service, Finance,<br />
Insurance, Real State, Social Services,<br />
Education, Scientific Research, Health)<br />
806 34.6<br />
6-10 years 339 14.6 Others 216 9.3<br />
More than 10<br />
years<br />
1810 77.7<br />
Female Less than 1 year 7 2.7 Manufacturing category ( Agriculture,<br />
Mining, Manufacturing, Geology,<br />
Construction, Electricity & Gas)<br />
102 39.1<br />
2-5 years 14 5.4 Service category(Food service, Finance,<br />
Insurance, Real State, Social Services,<br />
Education, Scientific Research, Health)<br />
136 52.1<br />
6-10 years 34 13.0 Others 23 8.8<br />
More than 10<br />
years<br />
206 78.9<br />
© 2011 ACADEMY PUBLISHER<br />
exploratory factor analysis by using formal questionnaire<br />
date. The data’s KMO = 0.660, and Bartlett's test <strong>of</strong> the<br />
value <strong>of</strong> spherical is 6678.330, whose Sig =.000
1822 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
TABLE 4-3 FACTORS MATRIX AFTER ROTATION<br />
Component<br />
1 2 3 4 5<br />
F10 Post in Private<br />
Entrepreneur Association<br />
.841<br />
F11 Level in Private<br />
Entrepreneur Association<br />
.815<br />
F3 No. <strong>of</strong> social networks .638<br />
F8 Industry and Commerce<br />
Level<br />
.511<br />
F9 Industry and Commerce Post .444<br />
F6 CPPCC Post .897<br />
F7 CPPCC Level .873<br />
F4 NPC Post .887<br />
F5 NPC Level .882<br />
F2 Work experience -.763<br />
F1 Education .761<br />
F12 Cost for social activities .681<br />
F13 Time spends on social<br />
activities<br />
.530<br />
To further study the factors those impact women<br />
entrepreneurs’ financing strategy, we tried to establish a<br />
Logistic Regression Model. The results are as follows,<br />
shown in table 4-4, 4-5, and 4-6:<br />
Table 4-4 Hosmer and Lemeshow Test<br />
Step Chi-square df Sig.<br />
1 3.798 5 .579<br />
2 2.955 6 .814<br />
3 1.076 6 .983<br />
4 6.893 7 .440<br />
According to above tables, we can conclude our<br />
equation:<br />
P=1/(1+e-z)<br />
And z=-3.383 +0.504*Work experience+0.057*Cost<br />
for social activities+0. 774*Industry and Commerce+0.<br />
.375*Level in Private Entrepreneur<br />
Association+21.669*NPC Post.<br />
We may notice that the coefficients <strong>of</strong> work<br />
experience, cost for social activities, whether to join the<br />
Industry and Commerce, Level in Private Entrepreneur<br />
Association and the NPC Post, which are as measures <strong>of</strong><br />
human capital and social capital, are positively correlated<br />
with external financing. Thus, this result confirmed<br />
hypothesis H2 and H3, that is the more human capital and<br />
social capital women entrepreneurs have, the more they<br />
tend to use external financing.<br />
We can explain these results from the previous<br />
literature and the social status.<br />
(a) human capital: The work experience, as an<br />
indicator <strong>of</strong> human capital, may promote using external<br />
financing. From the demand side, it may enhance the<br />
relevant skills and accumulate <strong>of</strong> the managerial<br />
experience. From the investors’ points <strong>of</strong> view, whether<br />
having relative work experience is considered an<br />
© 2011 ACADEMY PUBLISHER<br />
important measure, when applying bank’s loan and<br />
attracting venture investment, which has also been<br />
confirmed in previous research. However, another<br />
measure <strong>of</strong> human capital, education background, was not<br />
evolved in our final model equation. Therefore, the<br />
higher education <strong>of</strong> women entrepreneurs does not means<br />
they would tend to use external financing. But this result<br />
is in accordance with Hu Huaimin's finding, that is the<br />
number <strong>of</strong> women entrepreneurs and their education level<br />
distributed in inverted “U ", that is to say, both less<br />
educated and higher educated women are not that<br />
interested in starting their own business.<br />
(b) the intensity <strong>of</strong> network: Cost for social activities,<br />
as an indicator <strong>of</strong> the intensity <strong>of</strong> network, may reflect the<br />
maintenance <strong>of</strong> social networks by women entrepreneurs<br />
to some extent. Scholars have found that there are<br />
triangular interactions among emotions, resources and<br />
interactive reciprocal relationship, which is the more the<br />
interaction between individuals; the more likely they are<br />
to participate in group activities to share feelings, the<br />
more likely to exchange resources. Therefore, spending<br />
more on social activities and participating in more<br />
interactive activities, women entrepreneurs may obtain<br />
external financing resources easier.<br />
(c) networks between enterprises and industries:<br />
Although Chinese people values family relationship<br />
deeply, family network still cannot meet the needs <strong>of</strong><br />
enterprises’ development. Therefore, joining in the<br />
Industry and Commerce, Association <strong>of</strong> Private<br />
Entrepreneurs may expand women entrepreneurs’ social<br />
circle and benefit their career. On one hand, among these<br />
networks between enterprises and industries, network<br />
members are engaged in similar activities, and exchange<br />
information <strong>of</strong> different market, related technical advice<br />
and financing with women entrepreneurs. On the other<br />
hand, female entrepreneurs have more opportunity to the<br />
key figures mastering scarce financial resources by these<br />
networks.<br />
(d) Administrative impact:The coefficient <strong>of</strong> NPC<br />
Post, a measure <strong>of</strong> administrative impact, is much<br />
greater than the other three, which explains this factor has<br />
larger impact <strong>of</strong> using external financing. In China, the<br />
people's congress is China's highest authority. Deputies<br />
generally have high social reputation in social life, and<br />
also have some administrative influence in the<br />
administrative system. Therefore, when women<br />
entrepreneurs take some deputies duties, such as director<br />
or deputy director <strong>of</strong> the Standing Committee, it will<br />
increase their personal reputation to a large extent.<br />
Because <strong>of</strong> financial market imperfections and lack <strong>of</strong><br />
policy stability, informal constraints in the economy<br />
during the transition period has played an important role<br />
in society, and network and people's trust has become<br />
extremely important. Thus, taking some duties in NPC is<br />
beneficial to attract external financing. Besides, based on<br />
previous studies, women, lacking <strong>of</strong> access to critical<br />
resources to grasp chances, are <strong>of</strong>ten disadvantaged in the<br />
network status. In China, about 75% <strong>of</strong> the deputies to the<br />
NPC are <strong>of</strong>ficials. Therefore, by serving as a certain NPC<br />
Post, women entrepreneurs will have an access to the
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1823<br />
some scarce financial resources and optimize their<br />
network infrastructure.<br />
According to above analysis, we establish our model as<br />
follows.<br />
Figure.2 Women entrepreneur’s financing strategy<br />
V. CONCLUSION<br />
The main finding <strong>of</strong> this paper including:<br />
(a) From the result <strong>of</strong> ANOVA, we may notice that<br />
during the development <strong>of</strong> enterprise, men and women<br />
Table 4-5 Variable in equation<br />
entrepreneurs have a significant difference in using<br />
internal or external financing.<br />
(b) Female entrepreneurs who have more human<br />
capital are more likely to use external financing. Work<br />
experience <strong>of</strong> women entrepreneurs has positive<br />
correlation to external financing, while education<br />
background is not significant in China.<br />
(c) Social capital on entrepreneurial financing<br />
strategies was significant. From the logistic regression<br />
equation, we find that Cost for social activities, Industry<br />
and Commerce, Association <strong>of</strong> Private Entrepreneurs and<br />
NPC Post are the key factors <strong>of</strong> promoting using external<br />
capital for Chinese women entrepreneurs.<br />
ACKNOWLEDGMENT<br />
This research is supported by NSFC (Project<br />
70603021) and Royal Society <strong>of</strong> Edinburgh and National<br />
Natural Science Foundation <strong>of</strong> China for financial<br />
support (Project 70911130020).<br />
B S.E. Wald df Sig. Exp(B)<br />
Step 1 a Cost for social<br />
activities<br />
.062 .023 7.664 1 .006 1.064<br />
Constant -.707 .152 21.555 1 .000 .493<br />
Step 2 b Cost for social<br />
activities<br />
.062 .023 7.052 1 .008 1.064<br />
Level in Private<br />
Entrepreneur<br />
Association<br />
.428 .159 7.216 1 .007 1.534<br />
Constant -.844 .164 26.611 1 .000 .430<br />
Step 3 c Cost for social<br />
activities<br />
.061 .023 6.966 1 .008 1.062<br />
NPC Post 21.636 19923.521 .000 1 .999 2.491E9<br />
Level in Private<br />
Entrepreneur<br />
Association<br />
.422 .160 6.948 1 .008 1.525<br />
Constant -.881 .165 28.519 1 .000 .415<br />
Step 4 d Work experience .495 .242 4.185 1 .041 1.641<br />
Cost for social<br />
activities<br />
.060 .023 6.758 1 .009 1.061<br />
NPC Post 21.502 19939.319 .000 1 .999 2.178E9<br />
Level in Private<br />
Entrepreneur<br />
Association<br />
.409 .162 6.423 1 .011 1.506<br />
Constant -2.719 .928 8.591 1 .003 .066<br />
Step 5 e Work experience .504 .245 4.241 1 .039 1.655<br />
Cost for social<br />
activities<br />
.057 .022 6.625 1 .010 1.059<br />
Industry and<br />
Commerce<br />
.774 .401 3.730 1 .053 2.169<br />
NPC Post 21.669 19544.921 .000 1 .999 2.575E9<br />
Level in Private<br />
Entrepreneur<br />
Association<br />
.375 .164 5.259 1 .022 1.455<br />
Constant -3.383 1.004 11.359 1 .001 .034<br />
a. Variable(s) entered on step 1: Cost for social activities. b. Variable(s) entered on step 2: Level in Private Entrepreneur Association.c.<br />
Variable(s) entered on step 3: NPC Post. d. Variable(s) entered on step 4: Work experience. e. Variable(s) entered on step 5: Industry and<br />
Commerce.<br />
© 2011 ACADEMY PUBLISHER
1824 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
REFERENCES<br />
[1] Nancy Carter, Candida Brush, Patricia Greene, Elizabeth<br />
Gatewood and Myra Hart, “Financing High-Growth<br />
Enterprise: Is Gender an Issue?”, J. Women in Business:<br />
Access and Success, 2003: 45-52.<br />
[2] Candida G. Brush, Nancy M. Carter, Patricia G. Greene,<br />
Myra M. Hart, “The role <strong>of</strong> social capital and gender in<br />
linking financial suppliers and entrepreneurial firms: a<br />
framework for future research”, J. Venture Capital, 2002,<br />
Vol. 4, Issue 4:305 – 323.<br />
[3] Coleman S, “Access to capital and terms <strong>of</strong> credit: A<br />
comparison <strong>of</strong> men-and women owned small businesses”,<br />
J. <strong>Journal</strong> <strong>of</strong> Small Business Management, 2000, 38<br />
(3):37-5.<br />
[4] Sara Carter, Eleanor Shaw, Wing Lam and Fiona Wilson,<br />
“Gender, Entrepreneurship and Bank Lending: The<br />
Criteria and Processes Used by Bank Loan Officers in<br />
Assessing Applications”, J. Entrepreneurship Theory and<br />
Practice, 2007, 31(3):427 – 444.<br />
[5] Barbara Orser, Sandra Hogarth-Scott, “Opting for Growth:<br />
Gender Dimensions <strong>of</strong> Choosing Enterprise Development”,<br />
Canadian <strong>Journal</strong> <strong>of</strong> Administrative Sciences, 2002,<br />
19(3): 284 – 300.<br />
[6] Cooper, A. C, F. J. Gimeno-Gascon and C. Y. Woo,<br />
“Initial Human and Financial Capital as Predictors <strong>of</strong> New<br />
Venture Performance”, <strong>Journal</strong> <strong>of</strong> Business Venturing,<br />
1994, 9: 371–395.<br />
[7] Loscocco, K. A., J. Robinson, R. H. Hall and J. K. Allen,<br />
“Gender and Small Business Success: An Inquiry into<br />
Women’s Relative Disadvantage”, J. Social Forces 1991,<br />
70(1): 65–85.<br />
[8] Bosma, N., M. van Praag, R. Thurik and G. de Wit, “The<br />
Value <strong>of</strong> Human and Social Capital Investments for the<br />
TABLE 4-6 THE MODEL AFTER REMOVING SOME OF THE VARIABLES<br />
Variable Model Log Likelihood<br />
Change in -2 Log<br />
Likelihood df Sig. <strong>of</strong> the Change<br />
Step 1 Cost for social activities -171.306 11.290 1 .001<br />
Step 2 Cost for social activities -167.047 10.341 1 .001<br />
Level in Private<br />
Entrepreneur Association<br />
-165.660 7.569 1 .006<br />
Step 3 Cost for social activities -163.290 10.160 1 .001<br />
NPC Post -161.876 7.332 1 .007<br />
Level in Private<br />
Entrepreneur Association<br />
-161.830 7.239 1 .007<br />
Step 4 Work experience -158.210 4.916 1 .027<br />
Cost for social activities -160.773 10.042 1 .002<br />
NPC Post -159.101 6.697 1 .010<br />
Level in Private<br />
Entrepreneur Association<br />
-159.101 6.698 1 .010<br />
Step 5 Work experience -156.215 4.995 1 .025<br />
Cost for social activities -158.475 9.515 1 .002<br />
Industry and Commerce -155.752 4.069 1 .044<br />
NPC Post -157.328 7.221 1 .007<br />
Level in Private<br />
Entrepreneur Association<br />
-156.461 5.487 1 .019<br />
© 2011 ACADEMY PUBLISHER<br />
Business Performance <strong>of</strong> Startups”, J. Small Business<br />
Economics, 2004, 23(3): 227–236.<br />
[9] Bates, T, “Entrepreneur Human Capital Inputs and Small<br />
Business Longevity”, J. The Review <strong>of</strong> Economics and<br />
Statistics, 1990, 72(4): 551–559.<br />
[10] Fabowale, L., Orser, B. and Riding, A. ,”Gender,<br />
structural factors and credit terms between Canadian small<br />
businesses and financial institutions”, J. Entrepreneurship<br />
Theory and Practice, 1995,19(4): 41 – 66.<br />
[11] Boden, R. J. and Nucci, A. R., “On the survival prospects<br />
<strong>of</strong> men’s and women’s new business ventures”, J. <strong>Journal</strong><br />
<strong>of</strong> Business Venturing, 2000, 15(4): 347 – 362.<br />
[12] Uzzi B, “Embeddedness in the Making <strong>of</strong> Financial<br />
Capital: How Social Relations and Networks Benefit<br />
Firms Seeking Financing”, J. American Sociological<br />
Review, 1999, 64(4): 481-505.<br />
[13] Higgins M, Gulati R, “Getting <strong>of</strong>f to a good start: the<br />
effects <strong>of</strong> upper echelon affiliations on underwriter<br />
prestige”, J. Organization Science, 2003, 14(3): 244–263.<br />
[14] Shane S, Stuart TE, “Organizational endowments and the<br />
performance <strong>of</strong> university startups”, J. Management<br />
Science, 2002, 48(1): 154–170.<br />
[15] Priscilla Chu, “Social Network Models <strong>of</strong> Overseas<br />
Chinese Entrepreneurship: The Experience in Hong Kong<br />
and Canada”, J. Canadian <strong>Journal</strong> <strong>of</strong> Administrative<br />
Sciences, Dec.1996; 13(4): 358-365.<br />
[16] Tjosvold, Dean, Weieker, David, “Cooperative and<br />
competitive networking by entrepreneurs: A critical<br />
incident study”, J. <strong>Journal</strong> <strong>of</strong> Small Business<br />
Management. Milwaukee: 1993, 31(1):11 一 22.<br />
[17] Zhilong Tian, Yongqiang Gao et al, “Chinese Enterprises’<br />
Political Strategy and Behavior Study”, J. Management<br />
World, 2003, 12.
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1825<br />
A Spatial Econometric Analysis <strong>of</strong> China’s<br />
Manufacturing Agglomeration based on Geoda<br />
and Matlab<br />
Huayin Yu<br />
Department <strong>of</strong> Statistics, Anhui University <strong>of</strong> Financial and Economics, Bengbu, China<br />
Email: y_hyin@163.com<br />
Weiping Gu<br />
Department <strong>of</strong> Statistics, Anhui University <strong>of</strong> Financial and Economics, Bengbu, China<br />
Abstract—Industrial agglomeration has gradually become<br />
an economic focus in recent years. Scholars has done a lot <strong>of</strong><br />
research about the formation mechanism <strong>of</strong> industry<br />
agglomeration and its influencing factors, but the spatial<br />
correlation <strong>of</strong> variables has still been neglected. Firstly this<br />
paper gives a brief introduction about Geoda s<strong>of</strong>tware and<br />
Matlab neural network toolbox, then use spatial statistical<br />
methods to describe the 1999-2008 China's manufacturing<br />
industry agglomeration. Secondly this paper uses spatial<br />
econometric methods to analyze the influencing factors <strong>of</strong><br />
China’s provincial manufacturing Agglomeration. The<br />
results show that the spatial econometric model is superior<br />
to the traditional econometric models and the analysis based<br />
on spatial econometric model are more accurate. Finally, the<br />
paper also gives a brief forecast <strong>of</strong> the manufacturing<br />
Agglomeration.<br />
Index Terms—manufacturing Agglomeration, spatial<br />
correlation, spatial lag model, spatial error model, BP<br />
neural network<br />
I. INTRODUCTION<br />
Geoda is a collection <strong>of</strong> s<strong>of</strong>tware developed by Luc<br />
Anselin. It has a friendly and graphical interface that<br />
users can easily implement exploratory spatial data<br />
analysis (ESDA) with it, such as spatial autocorrelation<br />
analysis and spatial econometric analysis. The Geoda<br />
s<strong>of</strong>tware includes an interactive environment that<br />
combines maps with statistical graphics, using<br />
dynamic-linked-window technology. Its original version<br />
date back to the first contribution made to develop a<br />
bridge between ESRI’s ArcInfo GIS and the SpaceStat<br />
s<strong>of</strong>tware. The second version <strong>of</strong> Geoda made an<br />
improvement to ESRI’s ArcView 3.x GIS that it can<br />
implement linked windows and brushing. In contrast to<br />
the previous versions, the current Geoda is independent<br />
s<strong>of</strong>tware that runs under any <strong>of</strong> the Micros<strong>of</strong>t Windows<br />
operating systems without a specific GIS system.<br />
Matlab is an advanced language and interactive<br />
environment that users can implement numerical<br />
computation with it. And its operational efficiency is<br />
much higher than traditional programming languages<br />
such as C, C++, and FORTRAN due to the excellent<br />
© 2011 ACADEMY PUBLISHER<br />
doi:10.4304/jcp.6.9.1825-1831<br />
design. Matlab can perform many complex tasks such as<br />
signal and image processing, computation, control system<br />
design, test and measurement, financial modeling and<br />
analysis. There are more than 30 Toolboxes in Matlab<br />
and they can be divided into two categories: functional<br />
toolbo x and field-based toolbox. The functional toolbox<br />
is mainly used to expand symbolic computing, modeling<br />
and simulation capabilities, word processing and<br />
hardware real-time interactivity. Functional toolbox can<br />
be used in a variety <strong>of</strong> disciplines. In the Opposite, the<br />
field-based toolbox is highly pr<strong>of</strong>essional, such as the<br />
control system toolbox, signal processing toolbox and<br />
finance toolbox. And neural network toolbox is one <strong>of</strong><br />
them. It extends Matlab with tools for designing,<br />
implementing, visualizing, and simulating neural<br />
networks.<br />
II. DESCRIPTIVE STATISTICS<br />
As a branch <strong>of</strong> econometrics, spatial econometrics<br />
focuses on dealing with spatial interaction and spatial<br />
structure in cross-sectional data and panel data regression<br />
model. This area has developed rapidly in recent years.<br />
Spatial econometrics is widely used in applied economics<br />
and policy analysis, particularly in regional economics,<br />
residential economics, environmental and resource<br />
economics and development economics and other fields.<br />
Firstly, this paper made an exploratory spatial data<br />
analysis <strong>of</strong> China’s manufacturing agglomeration with<br />
Geoda. Secondly, we performed a spatial econometric<br />
analysis on influencing factors <strong>of</strong> china’s manufacturing<br />
agglomeration using spatial lag model and spatial error<br />
model. Finally, we used Matlab neural network toolbox<br />
to predict China’s manufacturing agglomeration based on<br />
the existing data. From an economic point <strong>of</strong> view, this<br />
article can also be seen as an example <strong>of</strong> spatial<br />
econometric analysis.<br />
A. Dependent Variable and Indicators<br />
We have many indicators to measure the industrial<br />
agglomeration in the actual study. In this paper, we chose
1826 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
Location Quotient (LQ) to describe China’s<br />
manufacturing agglomeration. It is defined as follows:<br />
LQ ij = ( Eij<br />
Ei<br />
) ( Ekj<br />
Ek<br />
)<br />
In the formula above, ij E indicate the employment<br />
in j<br />
industry <strong>of</strong> i district; Ei indicate the total<br />
employment <strong>of</strong> i kj district; E indicate the employment<br />
in j industry <strong>of</strong> the total district k ; Ek indicate the total<br />
employment <strong>of</strong> the total district k . It is generally believed<br />
that the greater the Location Quotient coefficient, the<br />
higher the level <strong>of</strong> the region's industry agglomeration.<br />
B. Spatial Statistical Analysis <strong>of</strong> China’s Manufacturing<br />
Agglomeration<br />
(a) Spatial distribution <strong>of</strong> China's manufacturing<br />
industry<br />
To give a better analysis <strong>of</strong> the spatial variation<br />
process <strong>of</strong> China’s manufacturing agglomeration, we<br />
mapping the spatial percentile chart (three periods:<br />
1999-2001, 2002-2005, 2006-2008) with Geoda095i,<br />
based on provincial Location Quotient coefficient<br />
(Calculated average for each period). The results is<br />
shown in chart 1.<br />
(a): 1999-2001<br />
(b): 2002-2005<br />
(c):2006-2008<br />
chart 1: Spatial percentile chart <strong>of</strong> China’s provincial<br />
manufacturing agglomeration (1999-2008, three periods)<br />
From chart 1 we can see that: From 1999 to 2001,<br />
Shanghai, China’s economic center, got the highest LQ<br />
coefficient and rank the first echelon; Beijing, Tianjin<br />
© 2011 ACADEMY PUBLISHER<br />
rank the second echelon; Liaoning, Hebei, Shandong,<br />
Jiangsu, Zhejiang, Fujian, Guangdong and some<br />
provinces ( 1<strong>of</strong><br />
) central region rank the third echelon;<br />
however, the agglomeration <strong>of</strong> manufacturing industry in<br />
Xizang, Yunnan and Hainan are still at a low level. From<br />
2002 to 2005, Shanghai still rank the first echelon;<br />
however, instead <strong>of</strong> Beijing, Guangdong came into the<br />
second echelon; both Inner Mongolia and Ningxia move<br />
forward to the third echelon; Shanxi, Gansu drop to the<br />
fourth echelon. Spatial percentile chart <strong>of</strong> 2006-2008<br />
hasn’t changed compared to the 2002-2005’s.<br />
As can be seen from the above analysis, China's<br />
manufacturing industry mainly concentrated in the<br />
southeast coastal areas. Manufacturing sector <strong>of</strong> coastal<br />
areas showed an increasing trend, but this trend is<br />
gradually slowing down.<br />
(b) Spatial autocorrelation analysis <strong>of</strong> China's<br />
manufacturing agglomeration<br />
In actual research we <strong>of</strong>ten use Moran'I index to test<br />
the existence <strong>of</strong> spatial autocorrelation, which is defined<br />
as follows:<br />
n n<br />
∑∑Wij(<br />
Yi<br />
−Y<br />
)( Yj<br />
−Y<br />
)<br />
Mora<br />
i=<br />
1 j=<br />
1<br />
n′<br />
s I =<br />
n n<br />
2<br />
S<br />
∑∑<br />
i=<br />
1 j=<br />
1<br />
W<br />
ij<br />
n<br />
= ∑ i −<br />
In the formula above,<br />
i=<br />
Y Y<br />
n<br />
2 1<br />
2 1<br />
S ( ) Y = ∑ Yi<br />
n 1 , n i=<br />
1 ,<br />
Yi is the value <strong>of</strong>i district, n is the number <strong>of</strong> district,<br />
Wij is the Contiguity Based Spatial Weights: if<br />
region i and region<br />
j ij<br />
is adjacent,<br />
W =1; otherwise,<br />
W ij =0.<br />
i = 1, 2,<br />
⋅⋅⋅,<br />
n ; j = 1,<br />
2,<br />
⋅⋅⋅,<br />
m ; m = n<br />
or n ≠ m .<br />
Moran’s I rank from -1 to 1.<br />
For Moran's I index results, we can use standardized<br />
statistic Z to test the existence <strong>of</strong> spatial autocorrelation<br />
between the regions.<br />
I −E(<br />
I)<br />
Z =<br />
( 3)<br />
VAR(<br />
I)<br />
Under the assumption <strong>of</strong> normal distribution, the<br />
expectation and variance <strong>of</strong> Moran's I can be calculated<br />
as follows:<br />
2<br />
2<br />
1 n w1<br />
+ nw2<br />
+ 3w0<br />
2<br />
E( I)<br />
= − , VAR(<br />
I)<br />
=<br />
−E<br />
( I)<br />
( 4)<br />
2 2<br />
n−1<br />
w0<br />
( n −1)<br />
In the formula above,<br />
n n<br />
n n<br />
n<br />
1<br />
2<br />
2<br />
w0<br />
= ∑∑w<br />
ij,<br />
w1<br />
= ∑∑(<br />
wij<br />
+ wji)<br />
, w2<br />
= ∑(<br />
wi.<br />
+ wj.)<br />
,<br />
i=<br />
1 j=<br />
1 2 i=<br />
1 j=<br />
1<br />
i=<br />
1<br />
wi. wj. and are the sum <strong>of</strong> row i and column<br />
j<br />
<strong>of</strong> the<br />
spatial weight matrix respectively. Both mean and<br />
variance are theoretical.<br />
We can make a significant test <strong>of</strong> the spatial<br />
H<br />
autocorrelation based on the statistic Z caculated. 0 :<br />
spatial autocorrelation between the regions does not exist.<br />
In Geoda095i, we use Monte Carlo method to test the<br />
existence <strong>of</strong> spatial autocorrelation, and the significant<br />
( 2)
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1827<br />
level is determined by the p value <strong>of</strong> the statistic Z. If p < α ,<br />
0 H is denied; otherwise,<br />
Table 1: Moran’s I value <strong>of</strong> China’s manufacturing agglomeration (1999-2008)<br />
Year Moran’s I Mean<br />
Standard dev<br />
iation<br />
p value<br />
1999 0.3191 -0.0333 0.1073 0.0060<br />
2000 0.3216 -0.0330 0.1087 0.0058<br />
2001 0.3342 -0.0333 0.1072 0.0042<br />
2002 0.3063 -0.0337 0.1081 0.0062<br />
2003 0.3147 -0.0321 0.1094 0.0053<br />
2004 0.3294 -0.0337 0.1097 0.0042<br />
2005 0.3421 -0.0345 0.1112 0.0031<br />
2006 0.3572 -0.0335 0.1131 0.0031<br />
2007 0.3377 -0.0332 0.1100 0.0032<br />
2008 0.3496 -0.0324 0.1090 0.0032<br />
From table 1 we can make a conclusion that there was<br />
a significant positive spatial autocorrelation between<br />
China’s provincial manufacturing industries. This<br />
indicates that China's manufacturing industry did not<br />
distribute randomly, and the spatial distribution <strong>of</strong><br />
manufacturing industry showed a clear concentration<br />
trend over the last decade: the provinces that have similar<br />
LQ coefficient tend to concentrate geographically.<br />
III. METHOD AND MODEL<br />
A. Research Methods<br />
(a) Spatial Lag Model<br />
Spatial Lag Model (SLM) is mainly used to discuss<br />
whether there is a spillover effect <strong>of</strong> variables in a region.<br />
The model is expressed as follows:<br />
y = ρ Wy + Xβ<br />
+ ε<br />
In the formula above, y is a dependent variable;<br />
X is a n × k matrix <strong>of</strong> exogenous explanatory<br />
variables; ρ is a spatial regression coefficient, reflecting<br />
the effect <strong>of</strong> spatial dependence in observations; W is a<br />
n× n matrix <strong>of</strong> spatial weight;<br />
Wy<br />
is a spatial lagged<br />
dependent variable; ε is a random error vector.<br />
Parameter β reveals the effect the explanatory<br />
variable X has to dependent variable y . Spatial lagged<br />
dependent variable Wy is a exogenous variable<br />
reflecting how spatial distance influence the act <strong>of</strong><br />
regions. The act <strong>of</strong> regions is strongly affected by the<br />
cultural environment and the transfer cost related to<br />
spatial distance.<br />
(b) Spatial Error Model<br />
Spatial Error Model (SEM) is expressed as follows:<br />
y = Xβ<br />
+ ε,<br />
ε = λWε<br />
+ µ<br />
In the formula above, ε is a random error vector; λ<br />
is a n × 1 spatial error matrix <strong>of</strong> dependent variable<br />
vector; µ is a random error vector in normal<br />
distribution.<br />
Parameter β reveals the effect the explanatory<br />
variable X has to dependent variable y . Parameter λ<br />
© 2011 ACADEMY PUBLISHER<br />
0 H is accepted.<br />
reflects the effect <strong>of</strong> spatial dependence in observations.<br />
The spatial dependence in random error term measures<br />
how the error impact <strong>of</strong> dependent variable <strong>of</strong><br />
neighboring areas influence the observations in this<br />
region.<br />
(c) Estimation method<br />
Considering the endogeneity <strong>of</strong> explanatory variables<br />
in spatial regression model, coefficient estimates will be<br />
biased or invalid if we use OLS method to estimate the<br />
coefficient in Spatial Lag Model and Spatial Error Model.<br />
Instead, we can use other methods to estimate, such as<br />
Instrumental Variable method, Maximum Likelihood<br />
method, Generalized Least Squares method and<br />
Generalized Method <strong>of</strong> Moments. Anselin (1998)<br />
recommended Maximum Likelihood method for<br />
estimating the coefficient in SLM and SEM.<br />
(d) ( A 5)<br />
choice between SLM, SEM and spatial<br />
autocorrelation test<br />
Anselin and Florax (1995) proposed the following<br />
criterion: We can determine that Spatial Lag Model<br />
would be more appropriate if (a) LMLAG is more<br />
significant than LMERR statistically; (b) R-LMLAG is<br />
significant but R-LMERR is not significant. In the<br />
Opposite, Spatial Error Model would be better if (a)<br />
LMERR is more significant than LMLAG; (b)<br />
R-LMERR is significant but R-LMLAG isn’t.<br />
Besides R-squared, some common criterion includes<br />
Log likelihood, Likelihood Ratio, Akaike Information<br />
Criterion (AIC) and Schwartz Criterion (SC). The higher<br />
the Log likelihood, the lower the AIC and SC, the<br />
better the model. These indicators can also be used to<br />
compare the regression effect between OLS, SLM and<br />
SEM.<br />
B. Econometric Model<br />
In this ( article, 6)<br />
we proposes the main factors that<br />
affecting industrial agglomeration from four perspectives:<br />
comparative advantage, new economic geography,<br />
knowledge spillovers and the role <strong>of</strong> government.<br />
Considering measurability <strong>of</strong> indicators and availability<br />
<strong>of</strong> data, we proposes the following twelve indicators in<br />
table 2.
1828 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
Table 2: Explanatory variables and their settings<br />
Explanatory variables Symbol Meaning<br />
Agriculture endowment agrgift Added value <strong>of</strong> agriculture/GDP<br />
Endowment <strong>of</strong> natural resources natgift Output <strong>of</strong> extractive industries/GDP<br />
Capital endowment capgift Rate <strong>of</strong> capital formation<br />
Endowment <strong>of</strong> human resources humgift Number <strong>of</strong> students in higher school/Population<br />
Level <strong>of</strong> urbanization city Urban employment/Total employment<br />
Market demand demand Provincial GDP per capita/National GDP per capita<br />
Transportation road Classified road density<br />
Industrial foundation industry Added value <strong>of</strong> industry /GDP<br />
Openness open Foreign direct investment/GDP<br />
Supporting services service Added value <strong>of</strong> tertiary industry /GDP<br />
Patent approvals patent Provincial patent approvals/National patent approvals<br />
Financial income fisc-income Financial income/GDP<br />
In table 2, Agriculture endowment, Endowment <strong>of</strong><br />
natural resources, Capital endowment and Endowment <strong>of</strong><br />
human resources are the indicators <strong>of</strong> comparative<br />
advantage; Level <strong>of</strong> urbanization, Market demand,<br />
Transportation and Industrial foundation are the<br />
indicators <strong>of</strong> new economic geography; Openness,<br />
Supporting services and Patent approvals are the<br />
indicators <strong>of</strong> knowledge spillovers; Financial income is<br />
the indicator that reflect the role <strong>of</strong> government. On this<br />
basis, We recommend the following double logarithmic<br />
model:<br />
ln LQ = β0<br />
+ β1<br />
ln agrgift + β2<br />
ln natgift + β3<br />
ln capgift<br />
+ β6<br />
ln demand + β7<br />
ln road + β8<br />
lnindustry<br />
+ β9<br />
+ β11<br />
ln patent + β12<br />
ln fisc − income + ε<br />
β<br />
In the equation above, i are regression coefficient,<br />
i = 1, 2,<br />
⋅⋅<br />
⋅30,<br />
ε is a random error term. In the<br />
following empirical analysis, adjustment may be made to<br />
the model based on the actual situation.<br />
The sample in this paper includes all the provinces,<br />
autonomous regions and municipalities in China except<br />
Hong Kong, Macao and Taiwan (Chongqing is taken into<br />
Sichuan for convenience). All data can be found in<br />
"China Statistical Yearbook", "China Industrial Economy<br />
Statistical Yearbook" from 2000-2009 and the website <strong>of</strong><br />
The People's Bank <strong>of</strong> China.<br />
IV. EMPIRICAL ESTIMATION AND RESULTS<br />
A. Econometric Analysis<br />
Considering the formation <strong>of</strong> industrial agglomeration<br />
is a process, it will take some time for the effect<br />
becoming apparent. In this article, we set the Location<br />
Quotient <strong>of</strong> provincial Manufacturing <strong>of</strong> 2008 as<br />
+ dependent β4<br />
ln humgift variable + β5<br />
ln and city the twelve indicators <strong>of</strong> 2006 as<br />
lnexplanatory<br />
open + β variables.<br />
10 ln service<br />
( 7)<br />
First <strong>of</strong> all, we make a OLS estimation including all<br />
the factors. As can be seen from estimation 1 <strong>of</strong> table 3,<br />
the t value <strong>of</strong> most variables is not significant and severe<br />
multicollinearity exists in the model. So some adjustment<br />
need to be made until t value <strong>of</strong> most variables become<br />
significant and model’s multicollinearity is weaken.<br />
Results after adjustment is shown in estimation 2 <strong>of</strong> table<br />
3. Obviously, the model <strong>of</strong> estimation 2 is better than the<br />
model <strong>of</strong> estimation 1.<br />
Table 3: OLS estimation <strong>of</strong> the factors <strong>of</strong> manufacturing agglomeration (2006-2008)<br />
estimation 1 estimation 2<br />
Variables<br />
Coefficient<br />
Standard<br />
deviation<br />
t value p value Coefficient<br />
Standard<br />
deviation<br />
t value p value<br />
C 1.0524 2.1146 0.4976 0.6250 1.5389 0.5981 2.5728 0.0166<br />
lnagrgift -0.2014 0.2413 -0.8345 0.4155<br />
lnnatgift -0.0147 0.0641 -0.2294 0.8212<br />
lncapgift 0.3917 0.3991 0.9814 0.3401<br />
lnhumgift -0.0791 0.2610 -0.3031 0.7654<br />
lncity 0.2485 0.3863 0.6432 0.5286<br />
lndemand 0.2833 0.4734 0.5984 0.5574 0.8064 0.1133 7.1131 0.0000<br />
lnroad 0.1051 0.1081 0.9720 0.3446 0.0898 0.0522 1.7205 0.0982<br />
lnindustry 1.2566 0.4098 3.0661 0.0069 1.3435 0.2508 5.3550 0.0000<br />
lnopen 0.0744 0.0769 0.9678 0.3466<br />
lnservice 1.1140 0.8528 1.3062 0.2088 1.3270 0.6093 2.1779 0.0394<br />
lnpatent 0.0852 0.0849 1.0029 0.3299<br />
lnfisc-income -0.5922 0.3889 -1.5228 0.1461 -0.3536 0.2448 -1.4441 0.1616<br />
R 2 0.8936 0.9142<br />
LogL 6.6697 4.7362<br />
AIC 12.6606 2.5274<br />
SC 30.8761 10.9347<br />
F 21.3008 0.0000 62.8634 0.0000<br />
© 2011 ACADEMY PUBLISHER
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1829<br />
As can be seen from the above analysis, the main<br />
factors that influence China’s provincial manufacturing<br />
agglomeration include Market demand, Transportation,<br />
Industrial foundation, Supporting services and Financial<br />
income. Industrial foundation and Supporting services<br />
have positive impact on manufacturing agglomeration.<br />
Their elasticities are 1.3435 and 1.3270. While Financial<br />
income has a negative impact on manufacturing<br />
agglomeration with its elasticity -0.3536. Due to the fault<br />
<strong>of</strong> OLS method dealing with spatial autocorrelation, we<br />
intend to use spatial econometric models to analyze the<br />
influencing factors <strong>of</strong> manufacturing agglomeration. The<br />
new model is the one in estimation 2 <strong>of</strong> table 3:<br />
ln LQ = β0<br />
+ β1<br />
ln demand + β2<br />
ln road + β3<br />
ln<br />
+ β4<br />
ln service + β5<br />
ln fisc − income + ε<br />
We need to verify the existence <strong>of</strong> spatial<br />
autocorrelation before peforming spatial econometric<br />
analysis. The result <strong>of</strong> spatial dependence test is shown in<br />
table 4.<br />
Table 4: Spatial dependence test<br />
Spatial dependence MI/DF Statistics p value<br />
Moran’s I (Error) 0.1047 1.9679 0.0489<br />
LMLAG 1 0.2102 0.6465<br />
R-LMLAG 1 0.4540 0.5003<br />
LMERR 1 0.0931 0.7602<br />
R-LMERR 1 0.3370 0.5615<br />
LM-SARMA 2 0.5472 0.7606<br />
As is shown in table 4, LMERR, LMLAG, R-LMERR<br />
and R-LMLAG do not pass the test at significance level<br />
<strong>of</strong> 5%. And SEM is better comparing the value <strong>of</strong> LogL,<br />
industry AIC, SC and LR <strong>of</strong> SLM and SEM. The result <strong>of</strong><br />
estimation <strong>of</strong> SLM and SEM ( 8)<br />
is displayed in table 5.<br />
Table 5: SLM and SEM estimation <strong>of</strong> the factors <strong>of</strong> manufacturing agglomeration (2006-2008)<br />
SLM SEM<br />
Variables<br />
β<br />
Standard<br />
deviation<br />
t value p value β<br />
Standard<br />
deviation<br />
t value p value<br />
C 1.6128 0.5600 2.8797 0.0039 1.5984 0.5300 3.0154 0.0025<br />
lndemand 0.7757 0.1192 6.5054 0.0000 0.8018 0.0975 8.2223 0.0000<br />
lnroad 0.0742 0.0573 1.2952 0.1952 0.0851 0.0437 1.9457 0.0516<br />
lnindustry 1.3642 0.2306 5.9145 0.0000 1.3741 0.2224 6.1785 0.0000<br />
lnservice 1.3462 0.5464 2.4635 0.0137 1.3585 0.5442 2.4960 0.0125<br />
lnfisc-income 0.3370 0.2200 1.5319 0.1255 0.3516 0.2199 1.5985 0.1099<br />
ρ/λ 0.0621 0.1208 0.5146 0.6067 0.3324 0.1275 2.6064 0.0074<br />
Statistical test DF Statistics p value DF Statistics p value<br />
R 2 0.9296 0.9495<br />
LogL 4.8505 4.8034<br />
AIC 4.2989 2.3930<br />
SC 14.1073 10.8002<br />
LR 1 0.2285 0.6325 1 0.1344 0.7138<br />
As is shown in table 5, the coefficients <strong>of</strong> all variables<br />
are positive. It indicate that this estimation is more<br />
consistent with theoretical analysis. Industrial foundation<br />
and Supporting services have positive impact on<br />
manufacturing agglomeration. Their elasticities are<br />
1.3642 and 1.3462 in SLM, 1.3741 and 1.3585 in SEM.<br />
B. Prediction <strong>of</strong> Location Quotient <strong>of</strong> China’s Provincial<br />
Manufacturing<br />
BP neural network is error back propagation neural<br />
network and feedforward network. It is widely used in<br />
function approximation, pattern recognition and data<br />
compression. In this article, we will predict the 2009’s<br />
Location Quotient <strong>of</strong> China’s Provincial Manufacturing<br />
using Neural Network Toolbox in Matlab (2006, 2007<br />
and 2008’s location quotient are references for<br />
comparison). The Matlab program <strong>of</strong> prediction is as<br />
follows (take Beijing for example):<br />
clc<br />
close all<br />
clear all<br />
p0=[2.4289 2.1880 2.0585 1.6110 1.3667 1.4216<br />
1.4169 1.3470 1.2994 1.3316];<br />
day=1999:2008;<br />
plot(day, p0,'b+')<br />
© 2011 ACADEMY PUBLISHER<br />
hold on<br />
plot(day, p0, 'r-.')<br />
p1=(p0-min(p0))./(max(p0)-min(p0));<br />
for i=1:5;<br />
p(:,i)=p1(i:i+2);<br />
t(:,i)=p1(i+3);<br />
end<br />
p;<br />
t;<br />
for i=1:5;<br />
testp(:,i)=p1(i+1:i+3);<br />
testt(:,i)=p1(i+4);<br />
end<br />
net=newff(minmax(p),[20,1],{'logsig','purelin'},'trainl<br />
m');<br />
net.trainParam.lr=0.8;<br />
net.trainParam.epochs = 500;<br />
net.trainParam.goal = 0.001;<br />
net=train(net,p,t);<br />
y=sim(net,testp);<br />
E=testt-y;
1830 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
p01=p1(5:7)';<br />
y01=sim(net,p01);<br />
y08=min(p0)+y01*(max(p0)-min(p0));<br />
E01=(p0(8)-y08)/p0(8);<br />
p02=p1(6:8)';<br />
y02=sim(net,p02);<br />
y09=min(p0)+y02*(max(p0)-min(p0));<br />
E02=(p0(9)-y09)/p0(9);<br />
p03=p1(7:9)';<br />
y03=sim(net,p03);<br />
y10=min(p0)+y03*(max(p0)-min(p0));<br />
E03=(p0(10)-y10)/p0(10);<br />
y=[y08 E01;y09 E02;y10 E03]<br />
p04=p1(8:10)';<br />
y04=sim(net,p04);<br />
y11=min(p0)+y04*(max(p0)-min(p0))<br />
p2=[p0 y11];<br />
day1=[day 2009];<br />
figure<br />
plot(day, p0,'g-.')<br />
hold on<br />
plot(day, p0,'b+')<br />
hold on<br />
plot(2006,y08,'ro')<br />
plot(2007,y09,'ro')<br />
plot(2008,y10,'ro')<br />
plot(2009,y11,'ro')<br />
plot(day1,p2, 'k-.')<br />
Table 6 shows the details about the prediction <strong>of</strong> the<br />
location quotient <strong>of</strong> China's provincial manufacturing.<br />
table 6: Result <strong>of</strong> BP neural network prediction (2006-2009)<br />
2006 2007 2008 2009<br />
Provinces Actual Predicted Relative Actual Predicted Relative Actual Predicted Relative Predicted<br />
value value error value value error value value error value<br />
Beijing 1.3471 1.3471 0.0000 1.2995 1.3513 -0.0399 1.3317 1.3417 -0.0076 1.3234<br />
Tianjin 3.0678 3.0680 -0.0001 3.3412 3.2362 0.0314 3.1160 3.1768 -0.0195 3.3121<br />
Hebei 0.9264 0.9268 -0.0004 0.9620 0.9702 -0.0085 0.9860 0.9975 -0.0117 0.9678<br />
Shanxi 1.1078 1.1077 0.0001 1.1452 1.1093 0.0314 1.1828 1.1927 -0.0084 1.2705<br />
Inner Mongolia 0.8289 0.8293 -0.0004 0.8529 0.8756 -0.0266 0.8782 0.8837 -0.0063 0.9291<br />
Liaoning 1.3766 1.3741 0.0019 1.4282 1.3442 0.0589 1.4928 1.5446 -0.0347 1.7410<br />
Jilin 0.9328 0.9325 0.0004 0.9951 1.0093 -0.0142 1.0031 1.0477 -0.0445 1.1994<br />
Heilongjiang 0.5891 0.5921 -0.0052 0.6210 0.5937 0.0440 0.6533 0.6673 -0.0215 0.7849<br />
Shanghai 3.5631 3.5114 0.0145 3.8370 3.8636 -0.0069 4.0874 4.0309 0.0138 4.0215<br />
Jiangsu 2.3332 2.3264 0.0029 2.4230 2.4105 0.0052 2.5108 2.4587 0.0207 2.4824<br />
Zhejiang 2.1312 2.1315 -0.0002 2.1339 2.2604 -0.0593 2.2352 2.2784 -0.0193 2.2766<br />
Anhui 0.4583 0.4587 -0.0008 0.4807 0.4570 0.0492 0.5093 0.4927 0.0326 0.4968<br />
Fujian 1.2987 1.2985 0.0002 1.3475 1.3451 0.0018 1.3897 1.3361 0.0386 1.3027<br />
Jiangxi 0.6510 0.6511 -0.0001 0.6724 0.6739 -0.0022 0.7028 0.6804 0.0318 0.6711<br />
Shandong 1.4882 1.4889 -0.0005 1.5514 1.5563 -0.0031 1.6114 1.5556 0.0346 1.5575<br />
Henan 0.5783 0.5780 0.0005 0.6038 0.5920 0.0196 0.6261 0.6312 -0.0081 0.6476<br />
Hubei 0.8619 0.8687 -0.0079 0.8907 0.9224 -0.0356 0.9010 0.9064 -0.0060 0.8571<br />
Hunan 0.6870 0.6870 0.0000 0.7222 0.7075 0.0205 0.7530 0.7612 -0.0110 0.7743<br />
Guangdong 2.6261 2.6265 -0.0001 2.7000 2.7044 -0.0016 2.8909 2.9614 -0.0244 2.9368<br />
Guangxi 0.3574 0.3576 -0.0006 0.3706 0.3692 0.0036 0.3812 0.3970 -0.0413 0.4505<br />
Hainan 0.2636 0.2641 -0.0019 0.2611 0.2734 -0.0472 0.2768 0.2730 0.0139 0.2851<br />
Sichuan 0.6065 0.6068 -0.0005 0.6266 0.6306 -0.0064 0.6456 0.6691 -0.0364 0.6766<br />
Guizhou 0.3089 0.3089 0.0002 0.3182 0.3157 0.0078 0.3310 0.3386 -0.0227 0.3437<br />
Yunnan 0.3094 0.3111 -0.0055 0.3119 0.2997 0.0389 0.3156 0.3227 -0.0224 0.3359<br />
Xizang 0.1206 0.1207 -0.0004 0.1162 0.1125 0.0318 0.1152 0.1125 0.0234 0.1125<br />
Shaanxi 0.6657 0.6649 0.0011 0.7090 0.6814 0.0389 0.7447 0.7702 -0.0342 0.6921<br />
Gansu 0.5786 0.5789 -0.0005 0.6076 0.5766 0.0511 0.6346 0.6518 -0.0272 0.6594<br />
Qinghai 0.4761 0.4763 -0.0004 0.4896 0.4783 0.0231 0.5105 0.5374 -0.0527 0.5895<br />
Ningxia 0.7761 0.7792 -0.0039 0.8054 0.8101 -0.0058 0.8593 0.8469 0.0145 0.8456<br />
Xinjiang 0.4726 0.4714 0.0026 0.4837 0.4655 0.0376 0.4990 0.4868 0.0244 0.4834<br />
We have high precision in the prediction using BP<br />
neural network (Relative errors are within 6%). This also<br />
shows that BP neural network has a strong function in<br />
learning, association, fault-tolerant and highly nonlinear<br />
function mapping with a good ability <strong>of</strong> generalization.<br />
© 2011 ACADEMY PUBLISHER<br />
ACKNOWLEDGMENT<br />
The authors are grateful to Pr<strong>of</strong>essor Huanming Zhang<br />
and Pr<strong>of</strong>essor Erpo Lu <strong>of</strong> Anhui University <strong>of</strong> Financial<br />
and Economics (AUFE) for their valuable comments. The<br />
authors are also grateful to Zhongsheng Xu, a graduate<br />
student <strong>of</strong> department <strong>of</strong> statistics <strong>of</strong> AUFE for our earlier<br />
joint work. However, any remaining errors are the<br />
author’s responsibility.<br />
This work was supported in part by a grant from XYZ.
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1831<br />
REFERENCES<br />
[1] Anselin L. Spatial Econometrics: Methods and Models.<br />
Dordrecht: Kluwer Academic, 1988.<br />
[2] Anselin L. “Space and applied econometrics”. Special<br />
Issue, Regional Science and Urban Economics, Vol. 22,<br />
No. 3, pp.509-536, 1992.<br />
[3] Anselin L, Florax R. “Small sample properties <strong>of</strong> tests for<br />
spatial dependence in regression models: some further<br />
results”. In New Directions in Spatial Econometrics,<br />
Edited by Anselin and Florax. Berlin: Springer-Verlag,<br />
1995, pp.21–74.<br />
[4] Anselin L. GeoDa 0.9.3 User's Guide. Center for Spatially<br />
Integrated Social Science, 2003.<br />
[5] Mohamad H. Hassoun. Fundamentals <strong>of</strong> Artificial Neural<br />
Networks. The MIT Press, 1995.<br />
[6] Saeed Moshiri and Norman Cameron. “Neural Network<br />
Versus Econometric Models in Forecasting Inflation”.<br />
<strong>Journal</strong> <strong>of</strong> Forecasting, No. 19, pp.201-217, 2000.<br />
[7] Paul Krugman. Geography and Trade. The MIT Press,<br />
1992.<br />
[8] G. Ellison and E. Glaeser. “Geographic concentration in<br />
US manufacturing industries: A dartboard approach”.<br />
<strong>Journal</strong> <strong>of</strong> Political Economy, Vol. 105, No. 5,<br />
pp.889–927, 1997.<br />
© 2011 ACADEMY PUBLISHER<br />
[9] Marius Brulhart. “Economic Geography, Industry location<br />
and trade: the evidence”. The World Economy, Vol. 21,<br />
No. 6, pp.775-801, 1998.<br />
[10] Antje. “Determinations <strong>of</strong> Geographical concentration<br />
patterns in central and eastern European countries”,<br />
unpublished.<br />
[11] Stuart S. Rosenthal. “The determinants <strong>of</strong><br />
agglomenration”. <strong>Journal</strong> <strong>of</strong> Urban, Vol. 50, No. 2,<br />
pp.191-229, 2001.<br />
[12] Kim, S. “Expansion <strong>of</strong> Markets and the Geographic<br />
Distribution <strong>of</strong> Economic Activities: The Trends in U.S.<br />
Regional Manufacturing Structure, 1860-1987”. Quarterly<br />
<strong>Journal</strong> <strong>of</strong> Economics, Vol. 110, pp.881-908, 1995.<br />
Huayin Yu (1962-), pr<strong>of</strong>essor in Department <strong>of</strong> Statistics <strong>of</strong><br />
Anhui University <strong>of</strong> Financial and Economics, master tutor.<br />
Pr<strong>of</strong>essor Huayin Yu majors in numerical calculation and data<br />
analysis. He is the corresponding author <strong>of</strong> this article.<br />
Weiping Gu (1985-), graduate student in Department <strong>of</strong><br />
Statistics <strong>of</strong> Anhui University <strong>of</strong> Financial and Economics. He<br />
majors in numerical calculation and data analysis.
1832 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
Application <strong>of</strong> Computer Technology in<br />
Efficiency Analysis <strong>of</strong> China Life Insurance<br />
Company<br />
Hongling Wu<br />
School <strong>of</strong> Economics/Anhui University <strong>of</strong> Technology, Maanshan, China<br />
Email: wuhongling76@163.com<br />
XiaoFei Zeng<br />
School <strong>of</strong> Economics/Anhui University <strong>of</strong> Technology, Maanshan, China<br />
Abstract—During the recent 100 years, the third<br />
technological revolution has promoted the development <strong>of</strong><br />
computer technology dramatically, which thus has brought<br />
a great change in the economic society <strong>of</strong> human beings such<br />
as economics structure, employment direction, the form <strong>of</strong><br />
international economic and the form <strong>of</strong> business. Besides,<br />
new concepts and ideas have been brought into the mode <strong>of</strong><br />
production and life style <strong>of</strong> human beings. By the<br />
instrumentality <strong>of</strong> LINDO s<strong>of</strong>t ware and SAS system, this<br />
research was conducted to evaluate the super efficiency,<br />
technical efficiency, pure technical efficiency and scale<br />
efficiency <strong>of</strong> life insurance companies <strong>of</strong> China in recent<br />
years by using the method <strong>of</strong> DEA and to analyze and find<br />
out the main and secondary factors that influenced the<br />
operational efficiency <strong>of</strong> insurance companies by using the<br />
measurement method. On this basis, it was concluded that<br />
efficiency <strong>of</strong> life insurance companies in our country could<br />
be enhanced by increasing underwriting quality,<br />
strengthening service awareness and optimizing business<br />
structure, etc.<br />
Index Terms—DEA model; efficiency; Life Insurance<br />
Company; insurance market; s<strong>of</strong>tware LINDO; SAS system<br />
I. INTRODUCTION<br />
Application <strong>of</strong> computer technology and computer<br />
programs pervades every field <strong>of</strong> human life and<br />
production and also alters the development mode <strong>of</strong><br />
human economic society. For example, in recent years,<br />
Lindo and Lingo are widely used in the fields <strong>of</strong><br />
economic management and empirical analysis. S<strong>of</strong>tware<br />
<strong>of</strong> Lindo and Lingo which were developed by American<br />
Lindo System Company are computer programs to solve<br />
the problem <strong>of</strong> optimization. The basic function <strong>of</strong> Lindo<br />
is to solve problems <strong>of</strong> linear programming and quadratic<br />
programming. Furthermore, Lingo not only has all the<br />
functions <strong>of</strong> Lindo but also can solve the problem <strong>of</strong><br />
nonlinear programming and Lingo can be used in the<br />
solution <strong>of</strong> linear and nonlinear equations. In the practical<br />
process <strong>of</strong> application, we find that the most significant<br />
Foundation item: Project <strong>of</strong> Non Fiction <strong>of</strong> Department <strong>of</strong> Education<br />
Anhui Province (Grant No.2010sk181).<br />
© 2011 ACADEMY PUBLISHER<br />
doi:10.4304/jcp.6.9.1832-1841<br />
feature <strong>of</strong> Lindo and Lingo is that an integer acting as a<br />
decision variable is available (integer programming) and<br />
the execution speed <strong>of</strong> these two kinds <strong>of</strong> s<strong>of</strong>tware is<br />
much faster.<br />
In fact, Lingo is a modeling language <strong>of</strong> the problem <strong>of</strong><br />
optimization, which includes many common functions <strong>of</strong><br />
mathematics, economics and management and it is<br />
available for users’ fitting the optimization model and it<br />
can supply interfaces <strong>of</strong> other data files such as text files,<br />
excel, database files and so on and it is very convenient,<br />
fast and simple for inputting, solving and analyzing<br />
spacious problems <strong>of</strong> optimization.<br />
Maybe thanks to these characteristics, Lindo and<br />
Lingo’s solving programs <strong>of</strong> linear, nonlinear and integer<br />
programming are used to analyze maximizing pr<strong>of</strong>its and<br />
minimizing costs by the broad masses <strong>of</strong> theory<br />
researchers and practical managers and the programs can<br />
be used in various fields and have been proved to be<br />
playing a significant role in commercial, industry,<br />
research and government including affairs <strong>of</strong> production<br />
distribution, ingredient mixing, arrangement between<br />
production and personal affairs, inventory management<br />
etc and especially the field <strong>of</strong> finance and insurance.<br />
SAS is a large-scale integrated computer s<strong>of</strong>tware<br />
system in which a set <strong>of</strong> computer programs worked<br />
together. The SAS users can make reasonable choices<br />
according to their demands. Since SAS is a kind <strong>of</strong><br />
integrated system, it has complete functions <strong>of</strong> data<br />
access, data management, data analysis, data report and<br />
so on. This computer system was promoted by American<br />
SAS S<strong>of</strong>tware Research Institution in 1976, and now has<br />
been adopted by 120 countries and 30,000 departments in<br />
the world. SAS when running under WINDOWS<br />
environment can fully utilize the eminent graphical<br />
interface <strong>of</strong> WINDOWS operating system and good<br />
connectivity with other system and data, which brings a<br />
lot <strong>of</strong> convenience on program editing and data<br />
manipulation and management. The operation <strong>of</strong> SAS<br />
system is flexible and functional; furthermore, its<br />
language is a powerful programming designing language<br />
and it integrates a variety <strong>of</strong> high-level language features<br />
and flexible format. It is an integration <strong>of</strong> data
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1833<br />
progressing and statistical analysis and also has a strong<br />
scalability. Therefore, the system is widely used by lots<br />
<strong>of</strong> general theory researchers and practical managers.<br />
Since 1980 when the domestic business was restored,<br />
the insurance industry <strong>of</strong> China has developed rapidly.<br />
The total volume <strong>of</strong> premium income was 1,600,000,000<br />
yuan in 1980, and an increase to 139,322,000,000 yuan<br />
(life insurance premium income was 87, 210,000,000<br />
yuan), and an increase to 978,400,000,000 yuan (life<br />
insurance premium income was 744, 738,000,000 yuan).<br />
The total volume <strong>of</strong> premium income in 2008 was<br />
increased by 6 times that in 1998 and life insurance<br />
premium income was increased 7.5 times. The capital <strong>of</strong><br />
insurance increased from 260,409,000,000 yuan in 1999<br />
to 3,341,844,000,000 yuan in 2008, an increase <strong>of</strong> 11.8<br />
times. The total investment <strong>of</strong> insurance increased from<br />
89,142,000,000 yuan in 1999 to 2,246,522,000,000 yuan<br />
in 2008, an increase <strong>of</strong> 24.2 times. The insurance density<br />
increased from 110 yuan per person in 1999 to 736.74<br />
yuan per person in 2008, and the insurance density was<br />
increased. As a result, insurance is playing a more and<br />
more important role in the development <strong>of</strong> economics <strong>of</strong><br />
the society. On aspect <strong>of</strong> attracting foreign investment:<br />
there were 18 overseas-funded enterprises <strong>of</strong> all insurance<br />
companies in China in 2000, and this number increased to<br />
89 in 2009. The increase <strong>of</strong> insurance companies<br />
especially on overseas-funded enterprises will lead to an<br />
increased competition in the insurance market. The<br />
financial strength, product development technology,<br />
development <strong>of</strong> industry approach and the business<br />
management level <strong>of</strong> foreign insurance companies are<br />
obviously better than those <strong>of</strong> domestic insurance<br />
companies. And because <strong>of</strong> the better salary, higher<br />
strategy on investment and management, a large number<br />
<strong>of</strong> excellent talents <strong>of</strong> management will be attracted by<br />
foreign insurance companies and this is a huge pressure<br />
for domestic insurance companies. Insurance companies<br />
<strong>of</strong> China always pay attention to underwriting income<br />
and scale <strong>of</strong> growth, and ignore claims service, efficiency<br />
and investing management and emphasize the premium<br />
income, thus under the macroeconomic environment that<br />
large numbers <strong>of</strong> foreign insurance companies flush into<br />
the market <strong>of</strong> insurance <strong>of</strong> China, the efficiency <strong>of</strong><br />
insurance companies is becoming a focus in this field.<br />
With the linear, nonlinear and quadratic solution<br />
programs <strong>of</strong> Lindo and through the method <strong>of</strong> Date<br />
Envelopment Analysis (DAE), this research is conducted<br />
to evaluate values <strong>of</strong> super efficiency, comprehensive<br />
efficiency, pure efficiency and scale efficiency, and<br />
analyze changes <strong>of</strong> efficiency <strong>of</strong> different insurance<br />
companies, and establish relevant econometric model to<br />
analyze the key factors affecting the efficiency <strong>of</strong><br />
insurance companies, and make appropriate comments<br />
and suggestions on enhancing the efficiency <strong>of</strong> insurance<br />
companies. Related researches in China only measured<br />
the efficiency <strong>of</strong> a certain value; however, this paper<br />
especially estimates the super efficiency <strong>of</strong> insurance<br />
companies to compare the pros and cons between<br />
insurance companies <strong>of</strong> which technology are effective<br />
© 2011 ACADEMY PUBLISHER<br />
and at last it analyzes the influencing factors on<br />
corresponding values <strong>of</strong> efficiency.<br />
II. INTRODUCTION EVALUTAION OF THE EFFICIENCY OF<br />
INSURANCE COMPANIES OF CHINA<br />
A. Sample Selection<br />
According to the principle <strong>of</strong> availability and<br />
comparability on data, 22 life insurance companies in<br />
2003-2008 were selected as the research samples. While<br />
newly established insurance companies that have been<br />
operated for 10 months and the premium income <strong>of</strong><br />
which was in the forefront <strong>of</strong> all newly established<br />
insurance companies were also selected, and at last, 22<br />
companies in 2003, 25 companies in 2004, 29 companies<br />
in 2005, 33companies in 2006, 35 companies in 2007,<br />
and 39companies in 2008 were chosen. In the data <strong>of</strong><br />
sampling companies, because PICC (People’s Life<br />
Insurance Company <strong>of</strong> China) was establish on 6.30.<br />
2003, and inherited relevant insurance business <strong>of</strong> CICL<br />
(China Life Insurance Company Limited), the increase <strong>of</strong><br />
reserve fund <strong>of</strong> PICC in 2003 is presented as the product<br />
<strong>of</strong> total increase reserve fund <strong>of</strong> CICL*( premium<br />
income <strong>of</strong> PICC / premium income <strong>of</strong> CICL), and<br />
compensation duty and pr<strong>of</strong>it margin are instead<br />
according CICL, others are instead according to data <strong>of</strong><br />
PICC. Relevant data come from the “China Insurance<br />
Yearbook” and relevant documents.<br />
B. Variable Selection and Comparison<br />
Efficiency is a reflection <strong>of</strong> result on microscopic<br />
behavior <strong>of</strong> enterprises, and it specially presents<br />
relationship between input and output or costs and<br />
benefits <strong>of</strong> insurance companies. According to the<br />
definitions given by Charnes and Coopers, the most<br />
important characteristics <strong>of</strong> input and output are that the<br />
increase <strong>of</strong> output and the decrease <strong>of</strong> input are the<br />
fundamental approach <strong>of</strong> pursuing aim and improving<br />
efficiency level <strong>of</strong> a production decision-making unit.<br />
There are three main methods to define input and output<br />
<strong>of</strong> a financial institution, namely intermediate approach,<br />
cost approach and added value. Financial institutions are<br />
generally calculated as a pure financial intermediary<br />
financial institution in the intermediate approach, that is,<br />
financial institution only earns the differences <strong>of</strong> interests<br />
through borrowing funds and transforming funds into<br />
assets. Obviously, this method is not proper for insurance<br />
companies. It is determined by the contribution to the<br />
income <strong>of</strong> financial institutions that whether a financial<br />
product can be acted as an input or output. If the proceed<br />
<strong>of</strong> the asset is greater than the opportunity cost <strong>of</strong> assets ,<br />
or liabilities <strong>of</strong> the financing cost is less than the<br />
opportunity cost, then the product can be considered to be<br />
financial outputs; otherwise that is input. This approach is<br />
theoretically feasible; however, it is not practically<br />
available because it needs accurate data <strong>of</strong> product<br />
benefits and opportunity costs which are difficult to<br />
estimate. Berger& Humphery(1997) considered that<br />
added value was an appropriate method to measure the<br />
output in researches on the efficiency <strong>of</strong> insurance
1834 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
companies, and this method can bring significant added<br />
value factors as output variables and the value reduction<br />
factors as input variables. There is basic agreement on the<br />
selection <strong>of</strong> input variables in China and the variables are<br />
mainly labor investment, capital investment (including<br />
forms <strong>of</strong> physical assets, paid-in capital, total capital and<br />
etc.) and operating costs (including forms <strong>of</strong> claim<br />
amounts and operating expenses etc.); however, there is a<br />
big difference on output variables. Zhao Xu (2003)<br />
adopted pr<strong>of</strong>its as the variable, while Hui min and Li Xin<br />
dan (2003) the asset pr<strong>of</strong>it margin and business income,<br />
Hou jing and Zhu lei (2004) the actual expected loss and<br />
investment income, Yao Shu jie et. (2005) premium<br />
income and investment income, Sun lin and Li Guang jin<br />
(2005) per capita pr<strong>of</strong>it and asset margin and He jing and<br />
Li Cunpu (2005) premium income. Insurance companies<br />
are different from general companies. It is clearly<br />
inappropriate to measure the operating efficiency by<br />
using a particular index <strong>of</strong> pr<strong>of</strong>its, pr<strong>of</strong>it margin or the<br />
premium income, such as premium income, and per<br />
capita premium is a quantitative measurement <strong>of</strong><br />
operating results, and it is difficult to evaluate the income<br />
and risk status objectively only by considering the<br />
premium income. The reserve fund is a indicator <strong>of</strong><br />
measuring business risk <strong>of</strong> insurance companies. The<br />
more adequate reserve fund the stronger ability <strong>of</strong><br />
insurance companies resisting risks, and pr<strong>of</strong>its and pr<strong>of</strong>it<br />
margins are pr<strong>of</strong>itability indicators <strong>of</strong> insurance<br />
companies, and the higher pr<strong>of</strong>it margin, the greater<br />
development potential <strong>of</strong> companies. The amount <strong>of</strong><br />
investment income presents the management and<br />
investment competence <strong>of</strong> companies. Modern insurance<br />
companies should not only pursue pr<strong>of</strong>its and investment<br />
income but also carry out their social duties, thus it is a<br />
key to measure the performance <strong>of</strong> insurance companies<br />
that considers premium income, pr<strong>of</strong>its, changes in the<br />
insurance reserve and investment income<br />
comprehensively.<br />
In summary, we adopt added value approach on input<br />
and output. Total fixed assets (equal to half <strong>of</strong> total fixed<br />
assets in early and the fixed assets at the end), total cost<br />
(including fees, commission costs and operating<br />
expenses), net amount <strong>of</strong> compensation payout (including<br />
direct insurance and reinsurance claims net <strong>of</strong><br />
compensation ), and total number <strong>of</strong> employees are<br />
selected as the input indicators. Premium income (equal<br />
to the direct insurance and reinsurance premium income),<br />
total pr<strong>of</strong>its, the amount <strong>of</strong> reserve growth (the amount <strong>of</strong><br />
preparation for the end <strong>of</strong> the year - the early mount <strong>of</strong><br />
preparation) and the amount <strong>of</strong> investment income are<br />
selected as output indicators.<br />
C. Selection <strong>of</strong> Model<br />
Estimating efficiency <strong>of</strong> insurance companies using<br />
DEA linear model includes: 1.Measure the technical<br />
efficiency value by using C2R model, thus in order to<br />
compare the comprehensive efficiency <strong>of</strong> insurance<br />
companies; 2. Measure the pure efficiency value by using<br />
BC2 model and compare efficiency <strong>of</strong> insurance<br />
companies after removing the scale factor; 3. Measure the<br />
super efficiency value by using super efficiency model<br />
© 2011 ACADEMY PUBLISHER<br />
and thus in order to compare and distinguish the<br />
achievements and failures among insurance companies; 4.<br />
Measure the returns to scale changes by NIRS model,<br />
when the technical efficiency value in NIRS model is not<br />
equal to that in BC2 model that TENIRS ≠ TE BC2 ,it<br />
means the unit being evaluated is in the increase region <strong>of</strong><br />
returns to scale, and the scale invalid is due to the small<br />
size and that means companies can increase efficiency<br />
through the expansion <strong>of</strong> scale. When TENIRS = TE<br />
BC2¸it means unit being evaluated is in the decrease<br />
region <strong>of</strong> returns to scale, and the scale invalid is due to<br />
the overlarge size <strong>of</strong> decision-making unit, and that<br />
means companies can increase efficiency through<br />
narrowing the scale. In this paper, origin and evolution <strong>of</strong><br />
the models are omitted, and the returns to scale status <strong>of</strong><br />
companies are not listed in this paper.<br />
(a) C 2 R Model<br />
T − T +<br />
ρ = θ − ε ( l s + l s )]<br />
min[ 1 2<br />
n<br />
∑ λi<br />
xi<br />
i=<br />
1<br />
−<br />
+ s =<br />
n<br />
∑ λi<br />
yi<br />
i=<br />
1<br />
+<br />
+ s =<br />
s .t.<br />
θx<br />
;<br />
(b) BC2 Model<br />
T − T<br />
ρ = θ − ε ( l s + l s<br />
min[ 1 2<br />
n<br />
∑ λi<br />
xi<br />
i=<br />
1<br />
−<br />
+ s =<br />
n<br />
∑ λi<br />
yi<br />
i=<br />
1<br />
+<br />
− s =<br />
n<br />
∑ λi<br />
= 1<br />
i=<br />
1<br />
s .t.<br />
k<br />
θy<br />
;<br />
+<br />
k<br />
k<br />
)]<br />
θx<br />
;<br />
(c) “Super Efficiency” Model<br />
T − T<br />
ρ = θ − ε ( l s + l s<br />
min[ 1 2<br />
n<br />
∑λ i xi<br />
i=<br />
1,<br />
j≠<br />
k<br />
−<br />
+ s =<br />
n<br />
∑λ i yi<br />
i=<br />
1,<br />
j≠<br />
k<br />
+<br />
− s =<br />
s .t.<br />
(d) NIRS Model<br />
T − T<br />
ρ = θ − ε ( l s + l s<br />
min[ 1 2<br />
n<br />
∑ λi<br />
xi<br />
i=<br />
1<br />
−<br />
+ s =<br />
n<br />
∑ λi<br />
yi<br />
i=<br />
1<br />
+<br />
− s =<br />
n<br />
∑ λi<br />
≤ 1<br />
i=<br />
1<br />
+ −<br />
i S S ≥<br />
s .t.<br />
y<br />
k<br />
+<br />
k<br />
;<br />
)]<br />
θx<br />
;<br />
y<br />
k<br />
+<br />
k<br />
;<br />
)]<br />
θx<br />
;<br />
λ ,<br />
In the model,<br />
, 0; ε<br />
are non-archimedean<br />
infinitesimal, and i λ<br />
is the weight <strong>of</strong> DMU decision-<br />
y<br />
k<br />
;
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1835<br />
x = x , x , ,<br />
x<br />
( 2<br />
making unit, and i 1i i mi is the input<br />
variable <strong>of</strong> DMU decision-making unit, and<br />
y = y , y , ,<br />
y<br />
( 1i 2i<br />
si<br />
)<br />
i<br />
is the output variable <strong>of</strong> f DMU<br />
+ −<br />
decision-making unit, and<br />
S , S<br />
is the slack variable,<br />
and S −<br />
is the m-dimensional column vector variable, and<br />
S +<br />
the s-dimensional column vector variable, and ρ is<br />
the ration <strong>of</strong> narrowing input. If<br />
DMU i decision-making unit is DAE effective, and if<br />
ρ = 1<br />
DMU<br />
and there is non-zero value in S+, S- , i<br />
)<br />
ρ = 1<br />
, = = 0<br />
− +<br />
S S<br />
decision-making unit is DAE weekly effective, and<br />
ρ ≺ 1 DMU<br />
if , i decision-making unit is DAE invalid.<br />
ρ is the index <strong>of</strong> relative efficiency, and in<br />
T T<br />
l1 = (1,1,...,1) 1× m, l2<br />
= (1,1,...,1) 1×<br />
s , s is the output<br />
variable and m is the input variable.<br />
D. Selection and Application <strong>of</strong> S<strong>of</strong>tware<br />
(a) Characteristics and Application <strong>of</strong> LINDO<br />
S<strong>of</strong>tware.<br />
LINDO was developed by the Linnus Schrage and is a<br />
kind <strong>of</strong> s<strong>of</strong>tware package that specially used to solve the<br />
mathematical programming problem. The s<strong>of</strong>tware<br />
package contained a complete series since its inception<br />
including LINDO, GINO, LINGO and LINGO NL. As<br />
mentioned above, LINDO is mainly used to solve linear<br />
programming, integer programming and quadratic<br />
programming problems, and GINO can be used to solve<br />
nonlinear programming problem, and to solve linear and<br />
nonlinear equations, inequalities and the roots <strong>of</strong><br />
algebraic equations, besides, GINO includes certain<br />
finance, probability and trigonometric functions and a<br />
variety <strong>of</strong> common mathematical functions which is<br />
available for user to invoke when creating the problem<br />
model, and LINGO can be used in solving linear and<br />
integer programming problem, and LINGO NL can be<br />
used for solving linear, nonlinear and integer<br />
programming problems.<br />
Because LINDO’s high speed on implementation and<br />
the convenience on inputting, solving and analyzing<br />
mathematical programming problems, LINDO is widely<br />
used in the fields <strong>of</strong> mathematics, scientific research and<br />
industry and LINDO has been developed several<br />
versions. Current versions <strong>of</strong> LINDO are powerful and<br />
are mainly used in solving linear, quadratic and integer<br />
programming problems. Interactive environment is<br />
available for beginners to set up and solve the<br />
optimization problem easily. On the other hand, it can<br />
also be used to solve some complex quadratic integer<br />
programming problems practically. Like on the largescale<br />
machine, it can be used to solve large-scale<br />
complex problems with more than 50,000 constraints and<br />
2,000,000,000 variables. Using LINDO s<strong>of</strong>tware, this<br />
paper gets the value <strong>of</strong> DEA value <strong>of</strong> several insurance<br />
companies through selecting input and output variables.<br />
© 2011 ACADEMY PUBLISHER<br />
Entering the following procedure in LINDO6.1<br />
window, the technical efficiency value <strong>of</strong> the insurance<br />
company Pacific-Antai Life Insurance Company Limited<br />
(PALIC) is obtained, and the procedures <strong>of</strong> other<br />
efficiency values <strong>of</strong> insurance companies are similar, and<br />
they are omitted. Here just presents the following<br />
procedure in LINDO6.1 window:<br />
MINX26<br />
ST<br />
2)149987.00X1+54876.91X2+34618.16X3+18820.71X4<br />
+17674.05X5+6607.45X6+1245.22X7+4758.05X8+597.<br />
6X9+617.36X10+188.26X11+174.88X12+50.32X13+65<br />
0.98X14+75.69X15+333.9X16+213.04X17+116.74X18+<br />
1206.1X19+120.13X20+83.48X21+3.66X22+53.59X23+<br />
292.22X24+21.56X25>21.56<br />
3)3157X1+2384.7X2-1602.29X3-360.57X4+110.68X5-<br />
570.52X6-126.35X7-37.93X8-1.5X9-72.95X10-<br />
40.63X11-54.35X12-10.39X13-131.33X14-22.85X15-<br />
60.73X16-80.38X17-54.61X18-126.54X19-40.52X20-<br />
56.45X21-26.43X22-72.15X23+25.23X24-19.48X25>-<br />
19.48<br />
4)93294X1+34412.44X2+25341.38X3+14419.62X4+126<br />
50.46X5+5774.42X6+1213.97X7+3348.81X8+321.27X9<br />
+360.1X10+131.36X11+78.43X12+34.56X13+282.75X1<br />
4+43.62X15+281.293X16+113.75X17+82.08X18+1035.<br />
6X19+132.88X20+64.5X21+2.2X22+42.76X23+291.73<br />
X24+7.54X25>7.54<br />
5)3669X1+2848.06X2+1288.79X3+453.59X4+710.95X5<br />
+164.01X6+18.73X7+369.49X8+28.93X9+16.74X10+11<br />
.41X11+5.04X12+1.65X13-<br />
5.88X14+3.28X15+5.14X16+6.33X17+5.21X18+48.82X<br />
19+3.59X20+2.56X21+1.96X22-<br />
0.6X23+14.88X24+3.51X25>3.51<br />
6)12773.5X1+4331.735X2+1418.055X3+929.485X4+57<br />
5.3X5+240.775X6+54.58X7+127.4X8+25.27X9+21.57X<br />
10+10.79X11+10.43X12+0.6X13+13.5X14+5.955X15+7<br />
.622X16+22.89X17+14.695X18+36.97X19+17.99X20+1<br />
4.945X21+5.095X22+12.16X23+10.81X24+2.845X25-<br />
2.845X26
1836 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
by COROLINA STATE UNIVERSICIY in 1966. The<br />
SAS INSTITUTE INC was established in 1976 and since<br />
then began the work <strong>of</strong> maintenance, development,<br />
marketing and training <strong>of</strong> SAS system. During the period,<br />
SAS had gone through many versions, and after several<br />
years’ improvement and development, SAS system has<br />
been valued as the international standard statistical<br />
analysis s<strong>of</strong>tware and is widely used in various fields.<br />
SAS is a modular, integrated large-scale application<br />
s<strong>of</strong>tware system. It consists <strong>of</strong> dozens <strong>of</strong> specialized<br />
modules and its functions include data access, data<br />
storage and management, application development,<br />
graphics processing, data analysis, report preparation,<br />
operation research approach, econometrics and<br />
forecasting etc.<br />
On one hand, SAS has characteristics <strong>of</strong> powerful<br />
functions and its statistical methods are abundant and<br />
new. SAS provides not only basic statistics calculation<br />
but also variance analysis, correlation and regression<br />
analysis and multivariate analysis <strong>of</strong> various statistical<br />
analysis processes <strong>of</strong> various experimental designs, and<br />
its technology <strong>of</strong> analysis is advanced and reliable. The<br />
analysis method is realized through the process call.<br />
Many processes also provide a variety <strong>of</strong> algorithms and<br />
options. For example, in the analysis <strong>of</strong> variance <strong>of</strong><br />
multiple comparisons, more than 10 kinds <strong>of</strong> methods<br />
including LSD, DUNCAN, and TUKEY are provided. A<br />
choice <strong>of</strong> 9 various methods (such as STEPWISE,<br />
BACKWARD, FORWARD, RSQUARE etc) is provided<br />
in regression analysis. In the regression model, users can<br />
choose whether to include the intercept and can also predesignate<br />
some independent variable word groups<br />
(SUBSET) in the model. For the intermediate results,<br />
those can be all output, not output or selecting output and<br />
can also be stored to a file for further analysis procedure<br />
call. On the other hand, SAS is easy to use and flexible to<br />
operate. It yields data sets through a common data<br />
(DATA) and later complete various data analysis through<br />
different procedure calls. Its programming statements are<br />
concise and short, and generally a number <strong>of</strong> complex<br />
operations with satisfactory results can be completed by a<br />
only a few statements. Results are presented by concise<br />
English prompt, and statistical terminology is standard<br />
and easily understand, and it is available for preliminary<br />
Company Number<br />
TABLE1. EFFICIENCY OF VARIOUS DIFFERENT INSURANCE COMPANIES 1<br />
Super<br />
efficiency<br />
English and statistical basis. Users just tell SAS what to<br />
do without telling how to do. Design <strong>of</strong> SAS make users<br />
do not have to tell SAS something that can be “guessed”<br />
by SAS ( that is without setting), and SAS also can<br />
correct some minor errors automatically. Besides, SAS<br />
can give reasons and correction method <strong>of</strong> running-time<br />
errors. As a result, SAS organically combines the<br />
scientific, precise and accurate <strong>of</strong> statistics and the feature<br />
<strong>of</strong> easily use together, which greatly facilitates the users.<br />
In SAS9.0 window, entering the following procedure,<br />
main factors influencing technical efficiency value <strong>of</strong><br />
insurance companies can be obtained. Procedure as<br />
followed: data A2;set A1;run;proc reg; model Y1=X1-<br />
X8/selection=stepwise sls=0.05 sle=0.2 r;run; A1: data<br />
files imported in SAS s<strong>of</strong>tware, including the 8<br />
assumptive influencing factors and specific values <strong>of</strong><br />
various efficiency, and other efficiency values are<br />
regression similar, and they are omitted in this paper.<br />
E. Efficiency Value and Evaluations <strong>of</strong> Company<br />
Table 1 shows us the super efficiency technical<br />
efficiency, pure technical efficiency and scale efficiency<br />
<strong>of</strong> different insurance companies. The top five insurance<br />
companies <strong>of</strong> super efficiency are Zhaoshangxinruo<br />
company, Ruitai life insurance company,<br />
Zhongbaokanglian v, PICC, Yangguang life insurance<br />
company, and the last five companies are Changcheng<br />
life insurance company, Hezhong life insurance<br />
company, Haier –NewYork company, Haikang life<br />
insurance company and Guangdianrisheng company,<br />
short term and long term companies are both include, and<br />
at the forefront only PICC is large scale company, while<br />
others are all small companies. The top five insurance<br />
companies <strong>of</strong> technical companies are Zhongbaokanglian<br />
company, Yangguang life insurance company,<br />
Yingdataihe company, Xingfu life insurance company<br />
and PICC, and the last five companies are Yingzhong life<br />
insurance company, Hezhong life insurance company,<br />
Haier-NewYork company, and Guangdianrensheng<br />
company. Pure technical efficiency and scale efficiency<br />
are similar, and this is because total value <strong>of</strong> technical<br />
efficiency is determined by pure technical efficiency and<br />
scale efficiency.<br />
Technical<br />
efficiency<br />
Average value<br />
Pure<br />
technical<br />
efficiency<br />
Scale<br />
efficiency<br />
Operating<br />
time:<br />
year<br />
China Life Insurance 1 1.8363 0.9771 1.0000 0.9771 2003-2008<br />
Ping An Life Insurance 2<br />
1.2635 0.9468 1.0000 0.9468<br />
2003-2008<br />
Pacific Life Insurance 3 1.0985 0.9571 1.0000 0.9571 2003-2008<br />
Xinhua Life Insurance 4 1.3150 0.9307 1.0000 0.9307 2003-2008<br />
Taikang Life Insurance 5 1.1445 0.9576 0.9933 0.9628 2003-2008<br />
Tai Ping Life Insurance 6 0.9604 0.7162 0.7979 0.8889 2003-2008<br />
Sino-Life Insurance 7 0.9095 0.6306 0.7971 0.7546 2003-2008<br />
AIA 8 1.0822 0.7964 0.9755 0.8200 2003-2008<br />
© 2011 ACADEMY PUBLISHER
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1837<br />
Zhonghong Life Insurance 9 1.4103 0.7883 0.8849 0.8417 2003-2008<br />
Pacific-Aetna Life Insurance 10 0.7050 0.5844 0.6307 0.8633 2003-2008<br />
Allianz Dazhong Life Insurance 11 0.5566 0.5566 0.8155 0.6497 2003-2008<br />
AXA-Minmetals Assurance 12 0.5001 0.4769 0.5607 0.7727 2003-2008<br />
China Life CMG 13 1.9438 1.0000 1.0000 1.0000 2003-2008<br />
Prudential Life Insurance 14 0.6829 0.5793 0.6557 0.7623 2003-2008<br />
John Hancock Tianan Life<br />
Insurance<br />
15<br />
0.5597 0.4378 0.8528 0.5012<br />
2003-2008<br />
Generali China Life Insurance 16 1.3654 0.9032 0.9162 0.9738 2003-2008<br />
Sun Life Everbright 17 0.5754 0.5163 0.6188 0.8224 2003-2008<br />
Haier New York Life 18 0.2565 0.2565 0.4327 0.5969 2003-2008<br />
Minsheng Life Insurance 19 0.6306 0.5326 0.6680 0.8020 2003-2008<br />
ING Insurance Company 20 0.8471 0.4018 0.5179 0.7364 2003-2008<br />
Sino-British Life Insurance 21 0.3992 0.3702 0.6227 0.5745 2003-2008<br />
Nissay-SVA Life Insurance<br />
Company<br />
22<br />
0.1805 0.1805 0.9527 0.1915<br />
2003-2008<br />
AEGON-CNOOC Insurance 23 0.2391 0.2391 0.4533 0.5286 2004-2008<br />
Heng An Standard Life<br />
Insurance<br />
24<br />
0.5092 0.5092 0.7096 0.7587<br />
2004-2008<br />
CIGNA and CMC Life Insurance 25 3.4377 0.6899 0.8611 0.7781 2004-2008<br />
China. MetLife 26 0.4016 0.3857 0.4340 0.6718 2005-2008<br />
Greatwall Life Insurance 27 0.3815 0.3815 0.6035 0.6295 2005-2008<br />
Cathay Life Insurance 28 0.4200 0.4200 0.6235 0.6692 2005-2008<br />
Winterthur Life 29 3.3269 0.7500 1.0000 0.7500 2006-2008<br />
United Metlife Insurance 30 0.6844 0.6492 0.7818 0.7918 2006-2008<br />
Union Life Insurance 31 0.2996 0.2996 0.4014 0.7529 2006-2008<br />
Huatai Life Insurance 32 0.6042 0.4977 0.7045 0.6348 2006-2008<br />
Jiahe Life Insurance 33 1.1850 0.6497 0.9100 0.6976 2007-2008<br />
Dragon Life Insurance 34 1.3203 0.6618 0.8649 0.8818 2007-2008<br />
Huaxia Life Insurance 35 0.6755 0.6755 0.8041 0.7096 2007-2008<br />
Sinatay Life Insurance 36 0.5879 0.5879 0.7880 0.7461 2008<br />
YingDaTaiHe Life Insurance 37 1.5471 1.0000 1.0000 1.0000 2008<br />
Happy Insurance 38 1.0442 1.0000 1.0000 1.0000 2008<br />
Sunshine Life Iunsurance 39 1.5893 1.0000 1.0000 1.0000 2008<br />
1 Data from the “China insurance Yearbook” from 2003 to 2008 and other relevant documents<br />
III. ANALYSIS OF FACORSIY OF INFLUENCING EFFICIENCY<br />
OF INSURANCE COMPANY<br />
A. Theory Analysis <strong>of</strong> Factors Influencing Efficiency <strong>of</strong><br />
Insurance Company<br />
According to current domestic research on the<br />
efficiency <strong>of</strong> insurance companies, it is considered that<br />
size, ownership structure, human capital, proprietorship<br />
structure, operating time and business scope <strong>of</strong> insurance<br />
companies will affect the efficiency <strong>of</strong> insurance<br />
companies. Based on the domestic researches’<br />
conclusions, this paper assumed the following factors<br />
influencing efficiency <strong>of</strong> insurance companies and did<br />
tests accordingly.<br />
a) X1Factor <strong>of</strong> the Capacity <strong>of</strong> Insurance Services:<br />
A key function <strong>of</strong> contracted business <strong>of</strong> insurance<br />
company is risking-sharing, and when insurers suffer<br />
© 2011 ACADEMY PUBLISHER<br />
losses, timely payments by insurance companies is one <strong>of</strong><br />
the keys that insurance companies can get businesses.<br />
Therefore, the loss ratio is measured as an index <strong>of</strong><br />
service competency <strong>of</strong> insurance companies. Low loss<br />
ratio will not only improve operational efficiency, but<br />
rather be in a disadvantage situation because <strong>of</strong> lack <strong>of</strong><br />
appeal in the fierce competition. Insurance companies<br />
always pay attention to premium income but ignore the<br />
claims, and when the overall loss ratio is low, it is<br />
assumed that the higher the loss ratio is, the more<br />
premium and the better operating efficiency the insurance<br />
companies get. Loss ratio = current total amount <strong>of</strong><br />
claim/current total amount <strong>of</strong> premium.<br />
b) X 2 Factor <strong>of</strong> Asset Scale:<br />
Insurance companies are enterprises operating risk<br />
business, and larger-scale insurance companies have<br />
higher ability <strong>of</strong> acceptance <strong>of</strong> risk, and small-scale<br />
insurance companies are disadvantage on both credibility
1838 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
and popularity, and insurance industry in China is in the<br />
stage that development and improvement <strong>of</strong><br />
competitiveness are relying on the growth <strong>of</strong> scale, so<br />
natural logarithm <strong>of</strong> asset amount <strong>of</strong> an insurance<br />
company is selected to present the scale <strong>of</strong> an insurance<br />
company.<br />
c) X 3Factor <strong>of</strong> Human Capital:<br />
Human resources play an important role in the<br />
development and competition <strong>of</strong> modern enterprise. An<br />
insurance company is composed <strong>of</strong> stuffs with different<br />
levels. It is generally believed that higher educated<br />
employees with stronger pr<strong>of</strong>essional knowledge are<br />
good for the development <strong>of</strong> an insurance company.<br />
Therefore the ratio <strong>of</strong> well-educated stuff (number <strong>of</strong><br />
employees who are undergraduate and over / total<br />
number <strong>of</strong> employees) is adopted as the human capital.<br />
d) X4 Factor <strong>of</strong> Productivity Per Labor Unit:<br />
Premium income per person is to measure the<br />
operating efficiency <strong>of</strong> an insurance company through the<br />
production efficiency per labor unit. Higher production<br />
efficiency per labor unit can produce better benefits and<br />
efficiency in an insurance company. Higher premium<br />
income per capita induces higher operating efficiency in a<br />
insurance company and vice versa. Premium income per<br />
capita = total premium income / total number <strong>of</strong><br />
employees.<br />
e) X5 Factor <strong>of</strong> Operating Time <strong>of</strong> a Company:<br />
It takes a long time to manifest the operating<br />
performance <strong>of</strong> an insurance company, and companies <strong>of</strong><br />
short operating time are at a disadvantage on business<br />
network, reputation and scale while companies <strong>of</strong> long<br />
operating time are at an advantage on business network,<br />
reputation and scale. Therefore, it is assumed that longer<br />
operating time, better operating efficiency.<br />
f) X6 Factor <strong>of</strong> Insurance Type:<br />
Business <strong>of</strong> life insurance companies can be divided<br />
into group insurance and individual insurance. Group<br />
insurance is better than individual insurance at terms <strong>of</strong><br />
size and quality, so a higher proportion <strong>of</strong> individual<br />
insurance in the premium income <strong>of</strong> an insurance means<br />
lower operating efficiency and vice versa. The type <strong>of</strong> an<br />
insurance business is account to the individual<br />
proportion, and an individual proportion = an individual<br />
premium/ total amount <strong>of</strong> premium income.<br />
g) X7 Factor <strong>of</strong> Underwriting Quality:<br />
For an insurance company, more surrender will bring<br />
negative effects on normal operation and development.<br />
High surrender ratio stands for low efficiency in an<br />
insurance company. Surrender ratio = amount <strong>of</strong><br />
surrender/ total amount <strong>of</strong> premium income.<br />
h) X8 Factor <strong>of</strong> Competence <strong>of</strong> Investment and<br />
Management:<br />
When an insurance company develops to a certain<br />
stage, the underwriting pr<strong>of</strong>it generally is low because<br />
competition increases. The insurance company enhances<br />
its competitiveness and developing ability mainly relying<br />
on high investment rate and good risk management. The<br />
insurance market is developing gradually, and the<br />
investment scope is expanding, and the investment risk is<br />
also expanding, and the rate <strong>of</strong> return on investment<br />
© 2011 ACADEMY PUBLISHER<br />
(ROI) influences greatly on the operating efficiency in an<br />
insurance company. Therefore, in this paper it is assumed<br />
that the higher rate <strong>of</strong> ORI the higher efficiency <strong>of</strong> an<br />
insurance company. ROI = net investment income/ total<br />
amount asset <strong>of</strong> the insurance company.<br />
B. Establishment <strong>of</strong> model<br />
The macroeconomic environment that influences the<br />
super efficiency, technical efficiency, pure technical<br />
efficiency, scale efficiency value <strong>of</strong> an insurance<br />
company includes insurance regulatory policy,<br />
macroeconomic conditions, as well as the operation<br />
situation <strong>of</strong> the enterprise itself such as ROI and<br />
premiums per capita. The enterprise can not alter the<br />
external factors like macroeconomic policy; however, the<br />
only changes that can be done by enterprise are to<br />
strengthen their management, improve operational<br />
efficiency. Coelli et al. (1998) proposed the famous “twostage”<br />
method, and its main thought is first calculates the<br />
efficiency value by DEA model, and then selecting<br />
appropriate environment variables to do regression<br />
analysis and then make sure <strong>of</strong> the factors influencing<br />
efficiency. In China, least squares regression and Tobit<br />
models are always used to estimate the influencing<br />
factors, and because the efficiency value <strong>of</strong> former<br />
model has a restrict range between 0-1, parameter<br />
estimation is biased and non-consistent. The technical<br />
efficiency value determined by DEA method can not<br />
distinguish the advantages and disadvantages among<br />
companies, and efficiency values <strong>of</strong> effective companies<br />
are all 1, and the restrict range <strong>of</strong> efficiency value which<br />
between 0-1 make the Tobit model no longer available in<br />
this situation. According to Hardwicketal’s method<br />
(2003), this paper did regression analysis on super<br />
efficiency and when did regression on technical<br />
efficiency, pure technical efficiency and scale efficiency<br />
value, convertible regression <strong>of</strong> efficiency value is<br />
adopted. Transform form is as below:<br />
Yi = Ln(<br />
TEi<br />
/ 1−<br />
TEi<br />
)<br />
TE i is technical efficiency, pure technical efficiency<br />
and scale efficiency value calculated in the DEA model,<br />
and the rang is 0-1, and Ln is natural logarithm, and in<br />
order to convert conveniently, all the efficiency value<br />
minus 0.0005, and regression model is established:<br />
Yi = β 0 + β1X<br />
1 + β2<br />
X 2 + +<br />
βn<br />
X n + εi<br />
After that, we can adopt the least squares method to do<br />
stepwise regression on dependent variable Y at different<br />
influencing factors and investigate the factors that have<br />
significant effect on the efficiency <strong>of</strong> insurance industry.<br />
C. Empirical Results Analysis<br />
From the regression results, we can find that the<br />
factors affecting the efficiency <strong>of</strong> an insurance company<br />
are loss ratio, human capital, premium income per capita<br />
and proportion <strong>of</strong> individual insurance and operating<br />
time. Loss ratio and proportion <strong>of</strong> individual insurance<br />
show significant difference at the level <strong>of</strong> 5% in t-test and<br />
others show significant difference at the level <strong>of</strong> 1% in ttest.<br />
Loss ratio, surrender rate, premium income per
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1839<br />
capita and company scale plays a significant effect on<br />
influencing efficiency <strong>of</strong> an insurance company. Of these<br />
four factors, surrender rate shows significant difference at<br />
the level <strong>of</strong> 5% in t-test and the other three show<br />
significant difference at the level <strong>of</strong> 1%. Details are<br />
shown in Table 2.<br />
TABLE 2. REGRESSION RESULTS OF FACTORS INFLUENCING EFFICIENCY OF LIFE INSURANCE COMPANIES OF CHINA<br />
Variable Model 1 Model 2 Model 3<br />
Constant -2.0209(4.18)**<br />
Loss ratio 7.4912(4.33)** 7.3445(5.86)**<br />
Human capital 5.8667(6.89)***<br />
Surrender ratio 6.2194(8.67)***<br />
Company scale 0.7632(41.7)*** 0.2844(44.82)***<br />
Premium per capita 0.2209(11.39)*** 0.1706(8.12)***<br />
Individual insurance 1<br />
-0.3144(4.46)**<br />
Operating time 0.1923(18.34)***<br />
Adjusted R 0.4795 0.1873 0.6231<br />
F statistics 32.79*** 41.70*** 73.99***<br />
Number <strong>of</strong> observation 183 183 183<br />
***、**means difference at level <strong>of</strong> 1 and 5%,T statistics are in brackets.<br />
1<br />
Propotion <strong>of</strong> individual insurance<br />
Models are obtained according the results <strong>of</strong> regression these will be improved. There is no doubt that premium<br />
Y<br />
( 1 , Y2,<br />
Y3<br />
are technical efficiency, pure technical<br />
efficiency, scale efficiency value separately)<br />
Y1<br />
= 7.<br />
4912X<br />
1 + 5.<br />
8667X<br />
3 +<br />
0.<br />
2209X<br />
Model1<br />
4 + 0.<br />
1923X<br />
5 − 0.<br />
3144X<br />
6<br />
per capita is the index <strong>of</strong> output per unit, and low output<br />
unit will never bring high operational efficiency, thus as<br />
premium per capita increases 1%, technical efficiency<br />
increases 0.2209%. Compared with other indexes, the<br />
impact <strong>of</strong> premium per capita is less, and it is relative to<br />
with that insurance marketing <strong>of</strong> China only seeks the<br />
Model 2<br />
expanding in scale a few years ago, and it means that it is<br />
Y 2 = −2.<br />
0209 + 0.<br />
7632X<br />
2<br />
Model3<br />
Y3<br />
= 7.<br />
3445X<br />
1 + 0.<br />
2844X<br />
2<br />
not feasible that insurance companies <strong>of</strong> China increases<br />
efficiency by expanding scale. Proportion <strong>of</strong> individual<br />
insurance and technical efficiency are negatively<br />
correlated, and 1% increase in proportion <strong>of</strong> individual<br />
+ 0.<br />
1706X<br />
4 + 6.<br />
2194X<br />
7<br />
In model 1, the loss ratio, human capital, premium<br />
income per capita, company operating time is positively<br />
correlated with technical efficiency, and as loss ratio<br />
increased by 1 percentage, technical efficiency increases<br />
7.4912 percents, and this shows that insurance companies<br />
<strong>of</strong> China should strengthen management on claims and<br />
improve the function on claims and security. As human<br />
capital increases 1%, technical efficiency increases<br />
5.8667%. Human capital generally does not show<br />
significant difference in the previous researches and thus<br />
it is <strong>of</strong>ten removed. However, this paper shows that<br />
human capital begins to play an important role on<br />
efficiency <strong>of</strong> insurance companies according to the last 6<br />
years data, and it is inseparable from the practical<br />
environment that Chinese insurance industry has fully<br />
opened to foreign countries and insurance companies has<br />
enhanced competition and talents have began to play a<br />
great role on the development <strong>of</strong> insurance companies<br />
since 2003. Premiums per capita and established time<br />
have a positive impact on the technical efficiency <strong>of</strong><br />
insurance companies, and this is consistent with previous<br />
analysis, and it means that the established time <strong>of</strong> the<br />
insurance company has a certain impact on efficiency <strong>of</strong><br />
the insurance company, and this is mainly related to with<br />
the marketing channels <strong>of</strong> the insurance company. Newly<br />
established insurance companies are poor at brand<br />
influence and marketing networks, and as time goes on,<br />
insurance and 0.3144% decrease in technical efficiency<br />
and this is consistent with the current situation in China.<br />
China’s insurance market is relatively underdevelopment,<br />
and compared with individual insurance, group insurance<br />
has superiority on scale and quality. If the individual<br />
insurance proportion is high in the business <strong>of</strong> insurance<br />
companies, operating efficiency will be relatively low.<br />
Therefore, it is a better choice to increase the proportion<br />
<strong>of</strong> group insurance in the business <strong>of</strong> insurance<br />
companies.<br />
In model 2, scale <strong>of</strong> insurance companies affects the<br />
pure technical efficiency <strong>of</strong> insurance companies. If the<br />
scale <strong>of</strong> the company increases 1%, pure technical<br />
efficiency increases 0.7632%. Overall, compares with<br />
other factors, scale <strong>of</strong> companies has a lower influence on<br />
efficiency <strong>of</strong> insurance companies.<br />
In model 3, loss ratio, surrender rate, premium income<br />
per capita and company scale affect the scale efficiency<br />
<strong>of</strong> insurance companies. These four factors are positive<br />
correlation. Loss ratio increases 1%, and the scale<br />
efficiency increases 7.3445%. This is basically consistent<br />
with model 1. This means that loss ratio influences<br />
technical efficiency <strong>of</strong> the company through influencing<br />
scale efficiency. Increase <strong>of</strong> loss ratio contributes to the<br />
increase <strong>of</strong> scale, thus increase scale efficiency, and the<br />
increasing scale efficiency can help improving the<br />
technical efficiency. 1% decrease <strong>of</strong> surrender rate<br />
contributes 6.2194% increase <strong>of</strong> scale efficiency and<br />
© 2011 ACADEMY PUBLISHER
1840 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
which may be related to China's current surrender terms<br />
and that insured is in a weak position compares with<br />
insurance company. In other words, the insurance<br />
company constrains the insured through a relatively harsh<br />
terms <strong>of</strong> surrender, so the insured once surrenders, the<br />
insurance companies benefit. A 1 % increase <strong>of</strong> company<br />
scale and the premium per capita followed a crease <strong>of</strong><br />
0.2844%, 0.1706% scale efficiency separately. As can be<br />
seen, it is very weak that insurance companies enhance<br />
technical efficiency through increasing scale efficiency.<br />
Overall, factors influencing technical efficiency <strong>of</strong><br />
insurance companies include loss ratio, surrender rate,<br />
premiums per capita, company scale, company operating<br />
time and proportion <strong>of</strong> personal insurance in companies’<br />
business. On the aspect <strong>of</strong> impact that ROI affects the<br />
operating efficiency, because significant issues in the<br />
regression analysis have been removed, and this is the<br />
disadvantage <strong>of</strong> the paper and it may be due to that when<br />
determining the DEA efficiency, return on investment is<br />
taken as the output item and thus make DEA efficiency<br />
highly related to ROI.<br />
IV. CONCLUSION AND SUGGESTION<br />
Statistical economics is playing a more and more<br />
important role in the modern society <strong>of</strong> economics and<br />
life. In order to grasp the pulse <strong>of</strong> the economy,<br />
government and enterprise collect and release large<br />
amounts <strong>of</strong> digital information every year, and in order to<br />
constitute the developing plan <strong>of</strong> society economy,<br />
several order differential equations, hundreds <strong>of</strong><br />
simultaneous linear equations and solving large-scale<br />
matrix are processed, and it is inconceivable without the<br />
help <strong>of</strong> computer. The actual shapes <strong>of</strong> various curves in<br />
economics mainly come from the analysis <strong>of</strong> statistical<br />
data and knowledge <strong>of</strong> database, procedures, and systems<br />
etc. are needed in computer application science. It is not<br />
only because that economics gives us inspires to<br />
understand the complex economical society, but also that<br />
it makes the market economy go through smoothly and<br />
get better control. To accurately grasp the subtleties <strong>of</strong><br />
economic and social development, we consider that we<br />
should first combine mathematics, economics and<br />
computer organically. Since we are familiar with our<br />
research object: the efficiency <strong>of</strong> life insurance<br />
companies, we should have a mathematical basis<br />
meanwhile we should also have to grasp the application<br />
<strong>of</strong> computer programs in some extent: LINDO s<strong>of</strong>tware<br />
and SAS s<strong>of</strong>tware, so we combine the two and carry out<br />
the thought <strong>of</strong> this article.<br />
In this paper, by the means <strong>of</strong> the s<strong>of</strong>tware <strong>of</strong> LINDO<br />
and SAS, efficiency values <strong>of</strong> different insurance<br />
companies are fast calculated and it supplies a scientific<br />
tool for comparing business efficiency <strong>of</strong> each insurance<br />
company and it promotes the application <strong>of</strong> LINDO and<br />
SAS in the field <strong>of</strong> economic management. The<br />
appearance <strong>of</strong> LINDO and SAS well solve the<br />
shortcoming <strong>of</strong> lack <strong>of</strong> tools in the field <strong>of</strong> economic<br />
analysis. Especially in the problem <strong>of</strong> linear<br />
programming, LINDO s<strong>of</strong>tware is simple, fast, and<br />
© 2011 ACADEMY PUBLISHER<br />
convenient to operate, and it is suitable for the users in<br />
the field <strong>of</strong> economic analysis.<br />
A. Improve the Underwriting Quality <strong>of</strong> Insurance<br />
Companies<br />
To the newly established insurance companies, it is<br />
pivotal that how to improve their popularity and get<br />
customers’ recognition, and only with the appropriate<br />
market scale, newly established companies can compete<br />
with the old famous state insurance companies;<br />
otherwise, everything is impossible. For the large scale<br />
insurance companies like PICC, it is an important issue to<br />
handle the relationship between scale and quality<br />
correctly. At the time <strong>of</strong> expanding scale, underwriting<br />
quality should be improved, and avoid ignoring<br />
underwrite quality because <strong>of</strong> the expansion <strong>of</strong> scale.<br />
B. Greatly Improve Loss Ratio and Service Quality <strong>of</strong><br />
Insurance<br />
The appropriate loss ratio is a key indicator to attract<br />
interests <strong>of</strong> insured and is a basic function <strong>of</strong> insurance<br />
companies. Some insurance companies attempt to<br />
improve pr<strong>of</strong>its by deliberately suppressing loss ratio, set<br />
barriers <strong>of</strong> coverage and claims; however, these measures<br />
will not enhance the pr<strong>of</strong>itability, operating efficiency<br />
and competitiveness <strong>of</strong> insurance companies but will<br />
make themselves in a vicious circle and make themselves<br />
in a disadvantage situation in the competition <strong>of</strong><br />
insurance. Appropriately increasing loss ratio will<br />
stimulate the enthusiasm <strong>of</strong> insured and also is<br />
conductive to the growth and maturity <strong>of</strong> the insurance<br />
market, and insurance companies also can deconcentrate<br />
the risk through the advantage <strong>of</strong> scale.<br />
C. Optimize Insurance Structure and Speed Up the<br />
Development <strong>of</strong> Insurance Products.<br />
In the current insurance market <strong>of</strong> China, group<br />
insurance is better than individual insurance on scale and<br />
quality; therefore, insurance companies should pay<br />
attention on promoting group insurance, increase the<br />
proportion <strong>of</strong> group insurance, and insurance companies<br />
should speed up the development <strong>of</strong> insurance products<br />
depending on the development <strong>of</strong> group insurance and<br />
especially focus on the market <strong>of</strong> individual insurance<br />
and promote various kinds <strong>of</strong> products to meet different<br />
individual need. As the development <strong>of</strong> economy, the<br />
individual insurance market will gradually develop and<br />
mature, and insurance companies will have a great space<br />
in developing individual insurance market.<br />
D. Improve the Investment and Management Capabilities<br />
<strong>of</strong> Insurance Companies and ROI, and Provide Supports<br />
For the Development <strong>of</strong> Insurance Companies.<br />
Modern insurance companies obtain a large number <strong>of</strong><br />
insurance funds relying on the insurance market and then<br />
get high returns depending on the excellent operating on<br />
investment and management and thus it will support the<br />
development <strong>of</strong> insurance business. The traditional<br />
situation <strong>of</strong> getting pr<strong>of</strong>its relying on the underwriting are<br />
not existed anymore, and the fierce competition make the<br />
underwriting pr<strong>of</strong>its <strong>of</strong> insurance companies become
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1841<br />
smaller and smaller, or even loss, so it is the inevitable<br />
choice for the insurance companies to improve the ROI.<br />
Facing the situation that foreign large insurance<br />
companies have joined the Chinese insurance market and<br />
the competition is increasing day by day, it is greatly<br />
effective to improve the insurance companies’ efficiency<br />
through strengthening the claim service, increasing the<br />
loss ratio, improving insurance structure, especially<br />
raising the proportion <strong>of</strong> group insurance.<br />
REFERENCES<br />
[1] Wei Quanling. DEA Data Packet Analysis[M].Beijing:<br />
Science Press.2004<br />
[2] Hou Jin, Zhu Lie.Non-life insurance empirical analysis <strong>of</strong><br />
operating efficiency <strong>of</strong> China Insurance Company<br />
[J].Nakai Economic Studies.2004(4),108_112<br />
[3] Li Kecheng. Empirical analysis <strong>of</strong> operating efficiency <strong>of</strong><br />
China Life Insurance Company[J].Insurance<br />
Studies.2005(2),37_41<br />
[4] Yao Shujie, Feng Genfu, Han Zhongwei. The Empirical<br />
Analysis <strong>of</strong> Efficiency <strong>of</strong> China's Insurance Industry<br />
[J].Economic Research Joural.2005(7),56_65<br />
© 2011 ACADEMY PUBLISHER<br />
[5] Hu Ying, Ye Yugang.An Empirical Study on the Factors<br />
Influencing the Efficiency <strong>of</strong> Insurers in China[J].<strong>Journal</strong><br />
<strong>of</strong> Jinan University.2008(4),28_34<br />
[6] Yue Chaolong. SAS system and economic statistical<br />
analysis[M].Hefei: University <strong>of</strong> Science and Technology<br />
<strong>of</strong> China Press.2003<br />
Hongling WU, female, was born on December 31th, 1976 in<br />
Dingyuan Anhui Province. Education background: Master <strong>of</strong><br />
Economics. Gain the Master <strong>of</strong> Quantitative Economics in 2008<br />
at Anhui University <strong>of</strong> Technology. Now, she applies herself to<br />
finance and transnational corporations. Work Experience: In<br />
1999-2000, Business College <strong>of</strong> East China University <strong>of</strong><br />
Metallurgy. Teaching Assistant; In 2001-2010, School <strong>of</strong><br />
Economics <strong>of</strong> Anhui University <strong>of</strong> Technology. Lecturer<br />
Xia<strong>of</strong>ei Zeng, man, was born on August in 1979 in Yinbin<br />
Sichuan Province.<br />
Education background: Master <strong>of</strong> Economics<br />
Major: Quantitative Economics
1842 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
A Bayesian Belief Net Model for Evaluating<br />
Organizational Safety Risks<br />
Li Song<br />
School <strong>of</strong> Economy and Management, Anhui University <strong>of</strong> Science and Technology, Anhui Huainan, China<br />
Email: lilysong@ustc.edu<br />
Li Yang<br />
School <strong>of</strong> Economy and Management, Anhui University <strong>of</strong> Science and Technology, Anhui Huainan, China<br />
E-mail: yangli081003@163.com<br />
Jing Han<br />
Huainan Vocational & Technical College, Huainan, Anhui, China<br />
E-mail: hanjing623@163.com<br />
Abstract—A Bayesian Belief Network (BBN) is a valuable<br />
tool to represent the causal relationships that exist in a given<br />
set <strong>of</strong> variables. This paper presents a methodology for<br />
organizational risk analysis for safety management.<br />
Learning a BBN from data is a difficult and<br />
resource-consuming task, we presents the implementation <strong>of</strong><br />
a greedy algorithm that automatically constructs a BBN<br />
from a dataset <strong>of</strong> cases obtained. The resulting BBN reflect<br />
installation specific factors respect to organizational factors<br />
and show the dependencies that exist among key variables<br />
that are associated to the trip generation process.<br />
Index Terms—Bayesian Belief Network; organizational risk<br />
factors; reliability analysis<br />
I. INTRODUCTION<br />
Complex socio-technical systems are comprised <strong>of</strong><br />
physical system and human system. The performance <strong>of</strong> a<br />
complex socio-technical system is dependent on the<br />
interaction <strong>of</strong> technical, human, social, organizational,<br />
managerial and environmental factors [1]. Safety<br />
performances <strong>of</strong>ten depend on complex and distributed<br />
interactions between human operators and technical<br />
systems. In the present dynamic society, a very fast pace<br />
<strong>of</strong> change <strong>of</strong> technology is found at the operative level <strong>of</strong><br />
society within many domains, and the rapid development<br />
<strong>of</strong> information and communication technology have<br />
leaded to a high degree <strong>of</strong> integration and coupling <strong>of</strong><br />
systems and the effects <strong>of</strong> a single decision can have<br />
dramatic effects that propagate rapidly and widely<br />
through the global society. Living in a very aggressive<br />
and competitive environment, companies today would<br />
focus the incentives <strong>of</strong> decision makers on short term<br />
financial and survival criteria rather than long term<br />
criteria concerning safety. It is widely recognized that<br />
© 2011 ACADEMY PUBLISHER<br />
doi:10.4304/jcp.6.9.1842-1846<br />
Jinkai Li<br />
Guanghua Management School,Peking University;<br />
E-mail: lijinkai@sina.com<br />
accidents in which ‘human error’ plays a part are <strong>of</strong>ten<br />
not solely attributable to errors made by an operator but<br />
have deeper causes, arising from the behavior <strong>of</strong> many<br />
others within the organizational context <strong>of</strong> a system[2].<br />
Investigations <strong>of</strong> accidents in complex systems have<br />
shown that events attributed to human error and blamed<br />
on an operator have systemic causes, such as procedural<br />
or organizational weaknesses. Many such failures and<br />
accidents do not have a simple explanation, particularly<br />
those that have significant contributions from human and<br />
organizational behaviors. Increasing interest over the past<br />
two decades in causal modeling <strong>of</strong> organizational safety<br />
behavior is in part motivated by the desire to understand<br />
the deeper more fundamental causes <strong>of</strong> accidents and<br />
incidents. Reason [8] describes the gradual relaxation <strong>of</strong><br />
safety alertness following a period <strong>of</strong> safe operation,<br />
followed by increased alertness after an accident as<br />
‘currents in the safety space’. Rusmussen [3] stresses<br />
environmental pressure will cause the operation <strong>of</strong> a<br />
system to migrate towards the boundary <strong>of</strong> safety. To<br />
analyze the risk <strong>of</strong> accidents and to improve safety,<br />
organizational risk factors need to be understood and<br />
evaluated. Several frameworks for analyzing the<br />
organizational context <strong>of</strong> accidents have been proposed,<br />
but without the capability to assess risks numerically.<br />
Event trees are usually used to model accident process,<br />
while organizational weaknesses have only an indirect<br />
effect on the accident and are therefore not readily<br />
represented as events. We outline an alternative method,<br />
using Bayesian Networks to model accidents, with<br />
explicit representation <strong>of</strong> both events and root causes.<br />
II. APPLICATIONS OF BBN IN ORGANIZATIONAL RISK<br />
ANALYSIS
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1843<br />
A. Bayesian Belief Network<br />
Bayesian probability has existed for many years.<br />
Implementation algorithms and s<strong>of</strong>tware tools have<br />
become available in recent years. Bayesian Networks are<br />
a network-based framework for representing and<br />
analyzing models involving uncertainty. They handle<br />
uncertainty in a mathematically rigorous yet efficient and<br />
simple way compared with other knowledge-based<br />
systems. Due to its power to deal with the s<strong>of</strong>t data in<br />
reliability, it has stimulated a strong interest [4]. A<br />
Bayesian Belief Network consists <strong>of</strong> a set <strong>of</strong> variables<br />
(causes and effects) and a set <strong>of</strong> directed edges between<br />
variables (paths <strong>of</strong> influence). Each variable has a finite<br />
set <strong>of</strong> mutually exclusive states. The variables together<br />
with the directed edges form a directed acyclic graph<br />
(DAG). Conditional probabilities carry the strength <strong>of</strong> the<br />
links between the causes and their potential effects. For<br />
example for a given state <strong>of</strong> a variable A with parents B1,<br />
…, Bn, we have the conditional probability <strong>of</strong> the state<br />
(A) occurring given the state <strong>of</strong> the contributing parent<br />
nodes: P(A|B1, …, Bn). Bayes' theorem in the subjective<br />
theory <strong>of</strong> probability is at the core <strong>of</strong> the inference engine<br />
<strong>of</strong> BBNs. In the definition <strong>of</strong> Bayesian Belief Networks,<br />
the DAG restriction is critical. Feedback cycles are<br />
difficult to model quantitatively and no calculus has been<br />
developed for causal networks that can cope with<br />
feedback cycles in a reasonably general way.<br />
Figure 1. Simple BBN<br />
B. Organizational risks factors<br />
The empirical studies <strong>of</strong> organizational safety<br />
performance have revealed a number <strong>of</strong> organizational<br />
factors in developing a predictive causal model <strong>of</strong><br />
organizational risks. In Bella’s view, large organizations<br />
are complex systems, which adaptively change and<br />
self-organize, the global patterns <strong>of</strong> organizational<br />
behavior that tend to reduce the safety <strong>of</strong> systems are<br />
common to all systems. Biondi (1998) [5] states that the<br />
organization system can have an affect on the reliability<br />
through numerous interrelated ways, such as work<br />
overload, time pressure and systemic distortion <strong>of</strong><br />
information. Many recent disasters happened not because<br />
<strong>of</strong> the way that safety was managed through the formal<br />
controls and procedures, but because <strong>of</strong> the safety culture<br />
in which safety management approaches were<br />
implemented [6]. Safety culture is a sub-facet <strong>of</strong><br />
organizational culture and is defined as common safety<br />
value in organization [7]. Certain works on the<br />
organizational factors have been devoted mainly to the<br />
© 2011 ACADEMY PUBLISHER<br />
classification <strong>of</strong> such factors. Embrey(1992) [8]analyzes<br />
railway accidents in the United Kingdom, and points that<br />
organizational risks factors have three levels: Level 1<br />
includes latent, active, and recovery errors; Level 2<br />
includes error-inducing factors such as training,<br />
procedures, time pressure, responsibilities, etc.; and Level<br />
3 includes policy deficiencies such as project<br />
management, safety culture, training policy, etc.<br />
Davoudian (1994) [9] proves that organizational factors<br />
should include overall culture (communication, decision<br />
making, etc.) and certain attributes <strong>of</strong> decision making,<br />
communication, etc. Leveson (2004) [10] views safety as<br />
a control problem and managed by a control structure<br />
developed for a socio-technical system.<br />
III. BUILD BBN MODEL OF ORGANIZATIONAL SAFETY<br />
RISKS<br />
For risk management purpose, it is necessary to have a<br />
technique that is capable <strong>of</strong> assessing the impacts <strong>of</strong><br />
potential changes. BBNs can be applied for predicting the<br />
effects <strong>of</strong> changes[11]. As a “probabilistic” technique<br />
rooted in Artificial Intelligence, BBN has the capability<br />
<strong>of</strong> utilizing subjective expert opinions. Adapting this<br />
technique makes the quantification <strong>of</strong> the organizational<br />
accident causation theory possible, even with a lack <strong>of</strong><br />
actual data. Figure 2 shows a schematic process model<br />
representing hierarchical structure <strong>of</strong> the process system<br />
<strong>of</strong> an organization, different activities at different layers<br />
construct organizational safety activities. Two activities<br />
are either sequential or hierarchical. For example, A2 and<br />
A 3<br />
A<br />
are related sequentially, and 22<br />
and 2 A are<br />
related hierarchically (<br />
A 22 is a sub-activity <strong>of</strong><br />
A 2 ). But<br />
in realty, it is possible that a safety output performance <strong>of</strong><br />
an organization is the result <strong>of</strong> two parallel activities,<br />
which are neither sequentially nor hierarchically related.<br />
The total safety performance 0 S<br />
can be broken down<br />
S that are the outputs <strong>of</strong> parallel<br />
into output 1 S to k<br />
activities<br />
A 11 to<br />
A1 K . Each <strong>of</strong> these activities have their<br />
own Resource ( R ), Input ( I ) and Control/Criteria ( C ).<br />
The second layer <strong>of</strong> activities, comprise those that<br />
have R , I , and C <strong>of</strong> the layer one as their outputs. For<br />
example,<br />
R 12 is the resource for activity 12<br />
A22 ; I12 is the input <strong>of</strong> activity 12 A<br />
output <strong>of</strong> activity R<br />
and the output <strong>of</strong> activity<br />
A and the<br />
A22 I C<br />
;and 12 is the control for<br />
A<br />
activity 12<br />
C<br />
and the output <strong>of</strong> activity<br />
A 22 . The same<br />
logic will hold for the activities from layer 1 to layer N<br />
(the layer where the modeler stops decomposition).<br />
The above schematic process model can be converted<br />
to BBN as shown in Figure 3. In Figure 3 the quality <strong>of</strong><br />
safety output would be a function <strong>of</strong> the quality <strong>of</strong><br />
A to K A . Knowing the state <strong>of</strong> 1 A to K A<br />
activities 1<br />
S<br />
given 1 A<br />
as well as the conditional probabilities for 0<br />
to A K , we can reach the probability <strong>of</strong> safety output with<br />
specific state. For example, considering a binary state for<br />
the factor (success and failure), by knowing the
1844 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
probability <strong>of</strong> success and failure for activities in layer N,<br />
and also the conditional probabilities, one can find the<br />
According to scope <strong>of</strong> organizational behavior, we<br />
usually divide organizational behavior into three layers,<br />
from individual, team to organization. Shown in table 1,<br />
we select training and workload & time pressure to reflect<br />
individual level, communication, safety administration<br />
and safety decision to reflect team level, salary policy and<br />
safety culture reflect organizational level. Each node and<br />
its all possible conditions should be definite and give a<br />
conditional probability in order to analyze effect <strong>of</strong> each<br />
organizational factor on system reliability using BBN.<br />
First <strong>of</strong> all, we need consider all nodes that construct<br />
BBN comprehensively. We classify all nodes into two<br />
© 2011 ACADEMY PUBLISHER<br />
probability <strong>of</strong> total output being in the success state.<br />
Fig.2 Schematic process model <strong>of</strong> organizational process system<br />
Fig. 3. Bayesian belief network for the organizational process system<br />
categories: human error nodes and organizational factors<br />
nodes which result in human errors. Then calculate<br />
conditional probability <strong>of</strong> each node based on the data<br />
sample. Human error and severe loss are selected as<br />
accident nodes. Secondly, we separate organizational<br />
factors in human error accidents database, that is to say,<br />
each organizational factor is divided into several states,<br />
and each state corresponding to a discrete value. Table 1<br />
shows category and characteristics <strong>of</strong> human error data<br />
discretization. Table 2 shows samples <strong>of</strong> organizational<br />
factors data.
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1845<br />
Table1 category and characteristics <strong>of</strong> organizational factors<br />
layer code Organizational factors characteristics<br />
Layer1 X 1<br />
safety culture<br />
Layer2<br />
X 2<br />
X 3<br />
X 4<br />
X 5<br />
Salary policy<br />
Communication<br />
Safety administration<br />
Safety decision<br />
good (1)<br />
average (2)<br />
Poor(3)<br />
Layer3 X 6<br />
training<br />
X 7<br />
Workload & time pressure<br />
Accident X 8<br />
X 9<br />
human errors<br />
severe loss<br />
Yes(1) No(2) Poor(3)<br />
Table 2 samples <strong>of</strong> organizational factors data<br />
X1 X2 X3 X4 X5 X6 X7 X8 X9<br />
1 1 2 1 2 1 2 2 2<br />
2 2 2 2 1 2 1 2 2<br />
1 1 3 1 2 2 2 2 2<br />
3 2 1 1 3 1 3 1 2<br />
3 2 2 2 3 2 2 1 2<br />
2 3 3 3 2 2 3 1 1<br />
Traditionally, BBNs were constructed from knowledge<br />
<strong>of</strong> human experts. However, during the last decade<br />
several methods had been developed to build them<br />
directly from databases. In order to ensure configuration<br />
<strong>of</strong> BBN, all variables 1 X , X 2 ,…, X 9 need to be ordered<br />
according to topology order . Father node set <strong>of</strong> each<br />
variable should be determined and partial probability <strong>of</strong><br />
each state need to be assigned. For the purposes <strong>of</strong> this<br />
paper the K 2 algorithm was applied to database [12].<br />
K 2 finds the optimal structure through a greedy search <strong>of</strong><br />
a reduced space <strong>of</strong> possible networks. The greedy<br />
criterion is based on a scoring function that represents the<br />
probability <strong>of</strong> a structure given data. In K 2 algorithm,<br />
Let Z be a set <strong>of</strong> n discrete variables (nodes) , where a<br />
X i variable in Z has i r<br />
possible value assignments.<br />
Let D be a database <strong>of</strong> m cases, where each case<br />
contains a value assignment for each variable in Z . Let<br />
Bs denote a belief network structure containing just the<br />
variables in Z , and p B<br />
the conditional probabilities.<br />
X i B s<br />
Each variable in has a set <strong>of</strong> parents, which are<br />
W ij<br />
represented with a list <strong>of</strong> variables. Let denote the<br />
jth<br />
unique instantiation <strong>of</strong> i relative to D . Suppose there<br />
are i q<br />
such unique instantiations <strong>of</strong> i N ijk<br />
. Define to be<br />
the number <strong>of</strong> cases in D in which variable i X<br />
has the<br />
r ik value and i W ij<br />
is instantiated as . According to the<br />
following theorem, the K 2 algorithm determines the<br />
optimal structure through a greedy procedure that<br />
identifies if a node can increase the network probability<br />
by adding a new parent to it, the structure <strong>of</strong> BBNs can be<br />
calculated.<br />
The basic structure <strong>of</strong> the K 2 algorithm is described<br />
with the following pseudocode:<br />
1. for i = 1 to n<br />
© 2011 ACADEMY PUBLISHER<br />
π i =<br />
2.<br />
{ }<br />
3.<br />
gn [ i]<br />
= g(<br />
i,<br />
null)<br />
4. finish=false<br />
π i<br />
< u<br />
5. while not finish and<br />
gnnew<br />
= −∞<br />
6.<br />
j = 1<br />
7. for to pred<br />
g( pred<br />
8. if<br />
[ j]<br />
, π i )<br />
> [ ] i gn<br />
then<br />
gnnew = g(<br />
pred<br />
9.<br />
[ j]<br />
, π i )<br />
z = j<br />
10.<br />
gnnew<br />
11. if > [ ] i gn<br />
then<br />
gn<br />
12. [ i]<br />
gnnew<br />
=<br />
π i = π i ∪ { z}<br />
13.<br />
14. else finish=true<br />
Where n is number <strong>of</strong> nodes in the network, i π<br />
is<br />
array <strong>of</strong> parents <strong>of</strong> node i , u is maximum allowable<br />
parents that any node can have, gn is array that stores the<br />
maximum values <strong>of</strong> ( ) g<br />
associated each node, pred is<br />
array <strong>of</strong> predecessor nodes to each node i , z is<br />
prospective parent with the highest probability.<br />
Figure 3 presents structure <strong>of</strong> BBNs applied above<br />
organizational factors database. The model indicates the<br />
value <strong>of</strong> safety-minded companies creating a safety<br />
culture that enhances communicating, decision and<br />
monitoring procedures, thereby reducing human error and<br />
severe loss. This direction will probably necessitate both<br />
restructuring the net in order to account for influences<br />
that were neglected at this stage and introducing other<br />
influences on organizational behaviour, such as training<br />
and salary policy, etc.
1846 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
Fig. 4 structure <strong>of</strong> BBN <strong>of</strong> organizational factors <strong>of</strong> the sample database<br />
IV. CONCLUSION<br />
Bayesian belief networks provide a robust probabilistic<br />
method <strong>of</strong> reasoning with uncertainty and are is more<br />
suitable to represent complex dependencies among<br />
components and to include uncertainty in modeling. In<br />
this paper we have demonstrated in principle that BBNs<br />
can be used for evaluating accident probability <strong>of</strong><br />
organizational factors. We have also shown how such a<br />
model can be used for practical applications. As stated in<br />
one <strong>of</strong> the assumptions <strong>of</strong> the K2 algorithm, an ordering<br />
<strong>of</strong> the nodes has to be established to define the structure<br />
<strong>of</strong> a BBN. The model succeeds in building a quantitative<br />
tier on top <strong>of</strong> the qualitative explanations <strong>of</strong><br />
organizational risks. We do believe that such models can<br />
become a reliable tool for predicting influence <strong>of</strong><br />
organizational risks changes and even in orienting safety<br />
investments at this level.<br />
ACKNOWLEDGMENT<br />
This work is supported by the National Natural<br />
Science Foundation <strong>of</strong> China(71071003), the MOE<br />
Project <strong>of</strong> Youth Foundation <strong>of</strong> Humanities and Social<br />
Science (09YJC630004), and the Natural Science<br />
Foundation key project <strong>of</strong> Anhui University<br />
(KJ2009A59).<br />
REFERENCES<br />
[1] Gordon , R.P.E. The contribution <strong>of</strong> Human Factors to<br />
accidents in the Offshore Oil Industry[J]. Reliability<br />
Engineering and System Safety, 1998, 61: 95-108.<br />
[2] Reason J. A systems approach to organizational error<br />
[J].Ergonomics, 1995, 38(8): 1708-1721.<br />
© 2011 ACADEMY PUBLISHER<br />
[3] Rasmussen, J. Risk Management in a Dynamic Society: a<br />
modeling problem[J]. Safety Science, 1997,27: 183-213<br />
[4] Oien K. A framework for the establishment <strong>of</strong><br />
organiza-tional risk indicators[J]. Reliability Engineering<br />
and System Safety, 2001, 7: 147-167<br />
[5] Zahra Mohaghegh, RezaKazemi,AliMosleh. Incorporating<br />
organizational factors into Probabilistic Risk Assessment<br />
(PRA) <strong>of</strong> complex socio-technical systems: A hybrid<br />
technique formalization[J]. Reliability Engineering and<br />
System Safety, 2009, 94:1000~1018<br />
[6] JEAN-CLAUDE ANDRE. Complexity and occupational<br />
safety and health prevention research. Theoretical Issues in<br />
Ergonomics Science, 2005, 6 (6): 483~507<br />
[7] Cooper G, Herskovits E.A Bayesian Method for the<br />
Induction <strong>of</strong> Probabilistic Networks from Data[J].Machine<br />
Learning, 1992,9: 309.<br />
[8] Embrey, D.E.. Incorporating management and<br />
organizational factors into probabilistic safety<br />
assessment[J]. Reliability Engineering and Systems Safety,<br />
1992(38): 199-208<br />
[9] Davoudian, K., et.al.. Incorporating Organizational Factors<br />
into, Risk Assessment through the analysis <strong>of</strong> Work<br />
Processes[J]. Reliability Engineering and System Safety,<br />
1994, 45: 85-91<br />
[10] Leveson, N. A new accident model for engineering safer<br />
systems[J]. Safety Science, 2004, 42(4): 237-270<br />
[11] M. Jaeger.Complex probabilistic modeling with recursive<br />
relational Bayesian networks[J]. Annals <strong>of</strong> Mathematics<br />
and Artificial Intelligence, 2001,32:179–220<br />
[12] Xiao qingkunl. theory and application <strong>of</strong> dynamic<br />
Bayesian networks learning [M]. National defence industry<br />
press, 2007<br />
Mrs. Li Song is an associate pr<strong>of</strong>essor in School <strong>of</strong> Economics<br />
and Management, Anhui University <strong>of</strong> Science & Technology,<br />
Huainan, Anhui, China. Her major field <strong>of</strong> study includes safety<br />
evaluation, organizational behavior and risk management<br />
(E-mail: lilysong@ustc.edu)
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1847<br />
Research and Application <strong>of</strong> J2EE and AJAX<br />
Technologies in Industry Report<br />
Min Hu<br />
School <strong>of</strong> Computer & Information /Hefei University <strong>of</strong> Technology, Hefei, China<br />
Email: uhnim@163.com<br />
Ding-ding Pan and Pei-en Zhou<br />
School <strong>of</strong> Computer & Information /Hefei University <strong>of</strong> Technology, Hefei, China<br />
Email: panding1986@sina.com, albertzpe@163.com<br />
Abstract—The traditional system <strong>of</strong> industry report is highly<br />
influenced by the speed <strong>of</strong> Internet and has low efficiency on<br />
report. In order to solve these problems, this paper studies<br />
J2EE and AJAX technologies, combine them and propose<br />
an industry report system which based on J2EE and AJAX<br />
technologies. The system which makes full advantages <strong>of</strong><br />
both technologies, has solved the problems such as easily<br />
impacted by the bandwidth, reported in low efficiency, also<br />
increased the server’s load capacity. It obtains a good result<br />
in the practical application.<br />
Index Terms—Industry Report, J2EE, AJAX<br />
I. INTRODUCTION<br />
With the deepening <strong>of</strong> China’s economic reform,<br />
various economic types and operational forms <strong>of</strong><br />
companies are emerging, number and size <strong>of</strong> enterprises<br />
are constantly expanding, and the traditional way <strong>of</strong><br />
industry report encountered a series <strong>of</strong> problems and<br />
faced a serious challenge in practice.<br />
In this case, it is imperative to establish an online<br />
industry report system using computer and network<br />
technology. Enterprises could connect to data networks <strong>of</strong><br />
management institutions and submit the industry reports<br />
directly through the Internet. The realization <strong>of</strong> online<br />
industry report system which has changed industrial data<br />
acquisition, is an inevitable reform trend <strong>of</strong> report<br />
method, also speeds up the construction <strong>of</strong> statistical<br />
information and achieves paperless report. The<br />
introduction <strong>of</strong> network-based work, brings a huge<br />
impact in industry report, and has greatly improved the<br />
capacity <strong>of</strong> data collection, analysis and aggregate while<br />
upgrading data quality and work efficiency.<br />
At present, despite that the online report system has<br />
got a certain application, most <strong>of</strong> them have many<br />
problems, such as slow access, system instability, report<br />
on low efficiency, poor server load capacity and so on. In<br />
view <strong>of</strong> these problems, after making detailed studies in<br />
J2EE and AJAX technologies, we apply them to the<br />
development <strong>of</strong> industry report system and achieve good<br />
results; both <strong>of</strong> them fully play respective advantages. For<br />
Project supported by the National Natural Science Foundation <strong>of</strong><br />
China (No. 60773043).<br />
© 2011 ACADEMY PUBLISHER<br />
doi:10.4304/jcp.6.9.1847-1851<br />
example, J2EE technology holds high scalability and<br />
steady availability; and AJAX technology owns strong<br />
response capability between the client and server.<br />
II. THEORIES OF J2EE AND AJAX<br />
A. The Theory <strong>of</strong> J2EE<br />
J2EE [1] is a system structure which uses Java 2<br />
platform standard edition as the core to simplify the<br />
development <strong>of</strong> enterprise solutions and deploy and<br />
manage some complex issues. It not only consolidates the<br />
advantages <strong>of</strong> the standard such as “write once, run<br />
anywhere”, to facilitate database access JDBC API,<br />
CORBA technology, security model <strong>of</strong> protecting data in<br />
Internet applications and so on, but also provides full<br />
support <strong>of</strong> EJB, Java Servlets API, JSP and XML<br />
technology. J2EE has many technical advantages, as<br />
follows:<br />
a) Supporting heterogeneous environment: J2EE<br />
could develop transplantable program which deployed in<br />
heterogeneous environment, and the program that<br />
developed once can be deployed to a variety <strong>of</strong> platforms.<br />
b) Scalability: Applications based on J2EE platform<br />
could be deployed to a variety <strong>of</strong> operating system. The<br />
provider <strong>of</strong> J2EE field <strong>of</strong>fers a wide range <strong>of</strong> load<br />
balancing strategies, allows integrated deployment <strong>of</strong><br />
multiple servers, and achieves highly scalable system.<br />
c) Steady availability: A server-side platform must<br />
be able to run uninterrupted. J2EE supports long-term<br />
availability while being deployed to a reliable operating<br />
system.<br />
J2EE uses multi-tier distributed application model.<br />
Application logic is divided into components in light <strong>of</strong><br />
the function. And all application components locate in<br />
different machines according to their location in different<br />
tiers. Now J2EE multi-tier enterprise applications divide<br />
different levels <strong>of</strong> two-tier model into many tiers. A<br />
multi-tiered application can provide an independent tier<br />
for each different service. The following is a typical<br />
four-tier structure <strong>of</strong> J2EE [2] (Shown in Fig. 1):<br />
a) Client tier components running on the client<br />
machine.<br />
b) Web tier components running on J2EE server.
1848 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
c) Business logic tier components running on J2EE<br />
server.<br />
d) Enterprise information system tier components<br />
running on EIS server.<br />
B. The Theory <strong>of</strong> AJAX<br />
Figure 1. Four-layer Structure Diagram.<br />
AJAX [3] works is to use the XMLHttpRequest object<br />
to transfer requests and responses asynchronously<br />
between the client and server. Fig. 2 shows the process<br />
flow <strong>of</strong> communication between client and server.<br />
XMLHttpRequest object is the core <strong>of</strong> AJAX and has<br />
become the actual standard <strong>of</strong> asynchronous transfer for<br />
XML data via HTTP. Asynchronous interaction means<br />
that the browser could continue processing the events<br />
page while sending request. Data is transferred in the<br />
background, and automatically loaded to the page without<br />
refreshing. Using AJAX technology has the following<br />
advantages:<br />
a) No page refreshing, communicating with server<br />
within the page, and providing a good user experience.<br />
b) Communicating with server using asynchronous<br />
mode without interrupting the user’s operation, and<br />
holding a more rapid response capability.<br />
c) Passing some <strong>of</strong> the burden work from server to<br />
the client, using the client’s ability to deal with, reducing<br />
the burden on server and bandwidth, saving space and<br />
bandwidth rental costs. And AJAX reduces the burden on<br />
the principles <strong>of</strong> “on-demand access <strong>of</strong> data”, shows the<br />
greatest degree <strong>of</strong> reduction <strong>of</strong> redundant request and<br />
response on the server.<br />
d) Based on a standardized and widely supported<br />
technology, no need to download plug-ins or applets.<br />
Figure 2. The Process Flow Diagram <strong>of</strong> Communication between<br />
Client and Server<br />
© 2011 ACADEMY PUBLISHER<br />
C. The Relation <strong>of</strong> J2EE and AJAX<br />
J2EE and AJAX are two technologies <strong>of</strong> Java, or two<br />
frameworks. They can not communicate with each other<br />
and both have their own advantages. On J2EE, all tasks<br />
are on the server, because, on the one hand, the client is<br />
relatively simple and does not need to do complex logic;<br />
on the other hand, data processing on the server is securer<br />
than the client. However, if the server capacity is limited,<br />
to improve data bandwidth and processing capabilities are<br />
also limited, many customers can not bear the burden. On<br />
AJAX, almost all services are placed on the client and<br />
processing speed is fast, but the client would be so<br />
complex that leading to poor compatibility. In general, all<br />
operations are focusing on the client and server. In<br />
response, we combine the two technologies and give full<br />
play to both <strong>of</strong> superiority. The core part <strong>of</strong> the<br />
implementation will use J2EE on server, while a<br />
relatively minor operation will be implemented with<br />
AJAX on the client.<br />
III. SYSTEM DESIGN<br />
A. The Analysis <strong>of</strong> System<br />
The principle system design is to ensure stability, high<br />
reliability, security and scalability <strong>of</strong> the data, implement<br />
unified interface for data interchange, exchange standards<br />
and authentication. Industry report system accomplishes<br />
the task <strong>of</strong> data collection. In order to facilitate on-line<br />
industry data report and ensure the access to the Internet<br />
effectively, we need to create a unified plan <strong>of</strong> report<br />
platform and data centre, specify normative<br />
organizational structure and data exchange standards<br />
while providing data interface to other systems.<br />
Since the work <strong>of</strong> special report, industry report system<br />
uses three-tier B/S architecture. This structure fully<br />
accounts the special report, not only provides users with a<br />
simple operating environment, but also ensures a quick<br />
and easy transfer report effectively. Enterprises connect<br />
to higher authorities via the Internet; they are linked into<br />
a seamless system by WEB technology and database.<br />
During industry report system B/S structure, enterprise<br />
users and the authorities have always been at the client.<br />
Enterprise users could transmit the data to authorities<br />
through the IE browser. The authorities can also carry out<br />
audit, management, statistics and summary on real-time<br />
data reported and print out reports (Shown in Fig. 3). The<br />
client is responsible for user authentication, input, report<br />
and audit <strong>of</strong> the data; and server is responsible for data<br />
reception and management.<br />
B. The Architecture <strong>of</strong> System<br />
System architecture is three-layer structure achieved<br />
by Struts [4][5] framework (shown in Fig. 4). It includes<br />
three parts: Model layer, View layer and Control layer.<br />
Compared with four-layer structure <strong>of</strong> J2EE, the view<br />
layer corresponds to the client tier components and the<br />
WEB tier components; model layer corresponds to<br />
business logic tier.<br />
a) Model Layer
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1849<br />
Figure 3. Flow Chart <strong>of</strong> Industry Report<br />
Model Layer is the main part <strong>of</strong> the system<br />
architecture. In the Struts framework, model layer is<br />
composed <strong>of</strong> ActionForm and JavaBean. ActionForm<br />
will encapsulate the user’s parameters ActionForm<br />
object, the object is forwarded to the Action by<br />
ActionServlet, and Action process client requests<br />
according to the request parameter in the ActionForm.<br />
JavaBean then encapsulates the underlying business logic<br />
like database access etc. The system uses DAO to access<br />
operations on database and protect the security <strong>of</strong><br />
database.<br />
b) View Layer<br />
View layer which composed by the JSP page is the<br />
interactive interface and achieves development and<br />
design <strong>of</strong> the main page <strong>of</strong> each functional module. It<br />
could check operational status to model, synchronize and<br />
update the user interface. Struts framework provides a<br />
rich tag library which could reduce the use <strong>of</strong> scripts;<br />
custom tag library can achieve an effective interaction<br />
with the model layer.<br />
c) Control Layer<br />
Control layer is the core <strong>of</strong> system architecture. Struts<br />
uses built-in Servlet—ActionServlet—as a controller,<br />
which receives a request from the client, enables event<br />
scheduling mechanism, selects the model <strong>of</strong> the<br />
corresponding business logic layer upon request, and then<br />
sends the results <strong>of</strong> the response to the client. While there<br />
are more concurrent operations in the client, use data<br />
scheduling to reduce pressure on the client by load<br />
balancing access technology.<br />
Figure 4. Framework Diagram <strong>of</strong> Struts MVC.<br />
© 2011 ACADEMY PUBLISHER<br />
C. The Design <strong>of</strong> Module<br />
On the base <strong>of</strong> the demand, we finalize the system<br />
modules, namely, enterprise information, enterprise data<br />
reporting, enterprise data auditing, statistical summary,<br />
query analysis and system management, a total <strong>of</strong> six<br />
modules. In accordance with the functional requirements,<br />
each module contains several corresponding sub-module<br />
(Shown in Fig. 5).<br />
Figure 5. Diagram <strong>of</strong> System Module<br />
D. The Design <strong>of</strong> Database<br />
In the database design [6], using Power Designer [7]<br />
for the design <strong>of</strong> logical model and physical model, and<br />
automatically generate the SQL script <strong>of</strong> SQL Server<br />
2008. According to the system design and function<br />
modules, design database tables as follows:<br />
a) Enterprise basic information table: used to store<br />
basic information <strong>of</strong> enterprises.<br />
b) User table and permissions group table: used to<br />
store user names and distribution <strong>of</strong> user rights.<br />
c) Production table and benefit table: used to store<br />
the data <strong>of</strong> production and benefit which reported by<br />
enterprise.<br />
d) Summary table: used to store the summary results<br />
<strong>of</strong> the data <strong>of</strong> production and benefit.<br />
IV. SYSTEM IMPLEMENTATION<br />
A. Development Environment<br />
Based on J2EE platform, B/S structure is applied to<br />
achieve system’s cross-platform deployment and<br />
operation. And MS SQL Server 2008 is chosen as<br />
background database, Tomcat 5.5 as publishing tools, and<br />
Eclipse 3.1 as programming tools.<br />
B. Implementation <strong>of</strong> Persistence with DAO<br />
During the development <strong>of</strong> J2EE-based system, in<br />
order to take full advantage <strong>of</strong> object-oriented features <strong>of</strong><br />
Java, developers <strong>of</strong>ten design the required Java classes to<br />
manipulate business data. Now databases used commonly<br />
are relational database rather than object database.<br />
Therefore, the data to add, delete, change and other<br />
operations in the Java class can not be directly persistent<br />
to the relationship table <strong>of</strong> database. In this paper we<br />
propose JDBC and DAO pattern to solve the problem<br />
using JDBC to establish a connection to the database and
1850 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
DAO [8] to abstract and encapsulate all access on data<br />
sources. And DAO is also responsible for connection<br />
management and data sources to obtain and store the<br />
data. Figure 6 shows the principle diagram <strong>of</strong> DAO.<br />
Figure 6. Principle Diagram <strong>of</strong> DAO<br />
C. Implementation <strong>of</strong> J2EE and AJAX Technology<br />
Struts framework can achieve MVC model <strong>of</strong> J2EE<br />
easily, a clear division <strong>of</strong> application makes the<br />
application logic and display logic independent <strong>of</strong> each<br />
other. However, users need to interact with server while<br />
the server is running. In this process users need to fill a<br />
large number <strong>of</strong> forms, and these operations directly<br />
affect the response speed <strong>of</strong> user interface. To solve the<br />
problem, DWR [9] which is a kind <strong>of</strong> AJAX technology<br />
is introduced in this paper. Its biggest advantage is<br />
safeguarding data without updating the page, combination<br />
<strong>of</strong> asynchronous nature <strong>of</strong> AJAX and synchronous nature<br />
<strong>of</strong> normal Java method calls. In asynchronous mode, the<br />
resulting data could be accessed asynchronously after the<br />
call has been executed for a long time.<br />
DWR which is an open source library contains two<br />
main parts: First, JavaScript could get data from a Servlet<br />
which is in the WEB server and follows the principle <strong>of</strong><br />
AJAX; second, a JavaScript library could help WEB<br />
developers to use the obtained data and change the<br />
content <strong>of</strong> page dynamically. In addition, DWR has<br />
adopted a new method which is similar to AJAX to<br />
dynamically generate JavaScript code which based on<br />
Java class, so the WEB developers could use the Java<br />
code in JavaScript. However, the Java code runs on<br />
server and is free to visit the WEB server resources.<br />
Finally, considering to the security, WEB developers<br />
must properly configure the Java class which can be used<br />
outside safely.<br />
V. EXAMPLES OF J2EE AND AJAX<br />
In this paper, we introduce the application <strong>of</strong> J2EE and<br />
AJAX technologies with the example: monthly report <strong>of</strong><br />
economic benefits.<br />
A. The Implementation <strong>of</strong> J2EE Technology<br />
Struts framework is mainly used to implement the<br />
MVC pattern <strong>of</strong> J2EE technology. It provides the<br />
controller which is inherited HttpServlet class and<br />
intercepts all HTTP requests, then calls the model layer to<br />
complete the request upon the HTTP requests and passes<br />
the final result to the client. To achieve the functionality,<br />
we need configure the struts-config.xml file as follows:<br />
<br />
<br />
© 2011 ACADEMY PUBLISHER<br />
<br />
<br />
name=”success”<br />
<br />
<br />
<br />
……<br />
<br />
<br />
<br />
name=”failure”<br />
And then implement the subclass which inherited the<br />
corresponding Action class.<br />
B. The Implementation <strong>of</strong> AJAX Technology<br />
The response <strong>of</strong> user interface is too slow while system<br />
is running. In order to solve the problem, this paper<br />
successfully introduced DWR which is a kind <strong>of</strong> AJAX<br />
technology. DWR allows passing a callback function<br />
which used to process the Java function call<br />
asynchronously.<br />
a) Configure dwr.xml file as follows:<br />
<br />
<br />
<br />
creator=”new”<br />
<br />
<br />
name=”class”<br />
<br />
……<br />
<br />
<br />
b) Javascript Calls<br />
<br />
<br />
<br />
<br />
Function dosubmit(){<br />
……<br />
JjzbybbManager.submitJjzbybb();<br />
……<br />
}<br />
<br />
Then it enables the client to call the function in class<br />
JjzbybbDao to complete the operations <strong>of</strong> economic<br />
reports.<br />
VI. CONCLUSIONS<br />
Based on B/S architecture and relational database, in<br />
this paper we design and implement industry report<br />
system based on J2EE and AJAX technologies. During<br />
the implementation, we achieve the MVC pattern <strong>of</strong> J2EE<br />
with Struts framework, design the logical and physical<br />
models with Power Designer tool, and implement the<br />
development <strong>of</strong> data persistence with DAO technology.<br />
Also, with the combination <strong>of</strong> J2EE and AJAX, the
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1851<br />
system has good scalability and security, fully embodies<br />
the design advantages <strong>of</strong> MVC pattern, and receives the<br />
desired results in practical application.<br />
ACKNOWLEDGMENT<br />
The authors would like to thank the guest editors<br />
reviewer for their valuable comments and insightful<br />
suggestions. This research was supported by the National<br />
Natural Science Foundation <strong>of</strong> China (No. 60773043).<br />
REFERENCES<br />
[1] W. Grawford and J. Kaplan. J2EE Design Patterns,<br />
Beijing: China Electric Power Press, 2005.<br />
[2] Liu Yang, Gao Liansheng and Wang Bin. “Study and<br />
implement <strong>of</strong> distribution system based on J2EE and MVC<br />
design pattern”. Computer Engineering and Design, 2007,<br />
7:1655-1658.<br />
[3] D. Johnson, A. White and A. Charland. Enterprise AJAX<br />
strategies for building high performance applications,<br />
Beijing: People’s Posts and Telecommunications<br />
Publishing House, 2008.<br />
[4] Yang Shaobo. Struts framework technology, Beijing:<br />
Tsinghua University Press, 2008.<br />
[5] J. Carnell, R. Harrop and K. Mittal. Pro apache struts with<br />
ajax, Beijing: People’s Posts and Telecommunications<br />
Publishing House, 2008.<br />
[6] Wang Shan and Sa Shixuan. Introduction to Database<br />
System, 4 th ed., Beijing: Higher Education Press, 2006.<br />
[7] Jiang Xueying. Web Database Design and Development,<br />
Beijing: Tsinghua University Press, 2007.<br />
[8] D. Alur, J. Crupi and D. Malks. J2EE Core Model, Beijing:<br />
China Machine Press, 2002.<br />
[9] Frank W.Zammetti. Practical DWR 2 projects, Beijing:<br />
People’s Posts and Telecommunications Publishing House,<br />
2009.<br />
© 2011 ACADEMY PUBLISHER<br />
Min Hu received her bachelor's degree<br />
in Automation Engineering from<br />
Nanjing University <strong>of</strong> Technology,<br />
Nanjing, China (1988), master's degree<br />
in Automation Engineering from Hefei<br />
University <strong>of</strong> Technology, Hefei, China<br />
(1994) and Ph.D. in Computer<br />
Application Technology from Hefei<br />
University <strong>of</strong> Technology, Hefei, China (2004). Since 1994, she<br />
is teaching at the School <strong>of</strong> Computer and Information <strong>of</strong> Hefei<br />
University <strong>of</strong> Technology, where currently she is a Pr<strong>of</strong>essor<br />
and actively participates in a number <strong>of</strong> national natural science<br />
foundation research projects concerning application and<br />
research <strong>of</strong> the theory and methods <strong>of</strong> multivariate rational<br />
interpolatory approximation in graphics, and research on the<br />
application <strong>of</strong> digital image processing based on continued<br />
fraction methods. Her current teaching and research interests<br />
include computer application, digital image processing and<br />
digital watermarking. She has authored more than 30 papers in<br />
international conferences and journals.<br />
Ding-ding Pan received his bachelor's degree in Computer<br />
Science and Technology from Anhui University <strong>of</strong> Architecture,<br />
Hefei, China (2008) and master's degree in Computer<br />
Application Technology from Hefei University <strong>of</strong> Technology,<br />
Hefei, China. His research interests include computer<br />
application, s<strong>of</strong>tware engineering.<br />
Pei-en Zhou received his bachelor's degree in Computer<br />
Science and Technology from AnQing Teachers College,<br />
Anqing, China (2008) and master's degree in Computer<br />
Application Technology from Hefei University <strong>of</strong> Technology,<br />
Hefei, China. His research interests include computer aided<br />
design.
1852 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
The Analysis <strong>of</strong> China New Energy Vehicle<br />
Industry Alliance Status based on UCINET<br />
S<strong>of</strong>tware<br />
Xiongfei Guo *<br />
School <strong>of</strong> Economics and Management/Beijing Jiaotong University, Beijing, China<br />
Yingqi Liu<br />
School <strong>of</strong> Economics and Management/Beijing Jiaotong University, Beijing, China<br />
Abstract—New energy vehicle industry acquires highly<br />
complex techniques. The new energy vehicle industry<br />
alliance is one <strong>of</strong> the most effective origination form and has<br />
been developed fast in China. This paper mainly use the<br />
s<strong>of</strong>tware UCINET to draw up the picture <strong>of</strong> China’s new<br />
energy vehicle industry alliance network, and study the<br />
cooperation relationships within the alliances through<br />
analyzing their elements. The results find that the key factor<br />
that effects the development <strong>of</strong> China’s new energy vehicle<br />
industry alliance is automobile companies. The key point in<br />
future management and research <strong>of</strong> China’s new energy<br />
vehicle industry alliance are cooperation and management<br />
between automobile companies and other members in the<br />
alliance.<br />
Index Terms—new energy vehicle; industry alliance;<br />
element; automobile company<br />
I. INTRODUCTION<br />
The development <strong>of</strong> new energy vehicle industry not<br />
only can help solving problems like energy security,<br />
carbon dioxide emissions, but also increase companies’<br />
ability <strong>of</strong> innovation, motivate industrial upgrade. In the<br />
past two decades, because <strong>of</strong> the complexity <strong>of</strong> new<br />
energy vehicle’s technologies and its research involves<br />
too many fields, there’re few companies that can master<br />
all technologies. Foreign new energy vehicle industries<br />
mainly adopted the organization <strong>of</strong> industry alliance, one<br />
<strong>of</strong> the most important organization forms to process the<br />
technical innovations. For example, in the year <strong>of</strong> 2010,<br />
Toyota and other Japanese automobile companies<br />
announced an alliance called CHAdeMo. Its group<br />
members include Toyota, Nissan, Mitsubishi, Fuji and<br />
Tokyo elec ∗ tric vehicle company. There’re 160<br />
companies joined the alliance in total, including foreign<br />
capital and government agencies (Japan Electric Vehicle<br />
Alliance, 2010).<br />
New energy vehicle industry was treated as a national<br />
strategic emerging industry, thus to gain fast<br />
development. From 2009, when China’s first self-owned<br />
brand new energy vehicle, Changan Jiexun, entered the<br />
market, to the end <strong>of</strong> 2010, many companies had carried<br />
∗ corresponding author <strong>of</strong> this article.<br />
© 2011 ACADEMY PUBLISHER<br />
doi:10.4304/jcp.6.9.1852-1856<br />
out their own new energy vehicle products. Inspired by<br />
the development <strong>of</strong> foreign new energy vehicle<br />
industries, China’s new energy vehicle industry alliance<br />
developed fast. Nearly 30 different types <strong>of</strong> new energy<br />
vehicle industry alliances had been established in China<br />
by the end <strong>of</strong> 2010 1 .<br />
This article mainly starts with the status <strong>of</strong> China’s<br />
new energy vehicle industry alliances, analyzing the<br />
elements within each alliance and seeks for related<br />
conclusions.<br />
II. ALLIANCE AND PARTNER TYPE<br />
Alliances are voluntary, cooperative agreements<br />
between two or more firms designed to achieve an<br />
advantage for the partners (Das and Teng, 2000). More<br />
generally, according to Gulati (1998), we define alliances<br />
are voluntary agreements between independent firms to<br />
develop and commercialize new products, technologies or<br />
services. Portfolio companies in the form <strong>of</strong> the formation<br />
<strong>of</strong> industry alliances and new product development to<br />
overcome the inherent risk associated with the control<br />
process <strong>of</strong> innovation and better results have reached a<br />
consensus (Jarillo, 1988; Gulati, Nohria, and Zaheer,<br />
2000). Hence, in the past two decades, Industry Alliance<br />
as an important form <strong>of</strong> industrial organization abounds<br />
in the automobile industry (Garcia-Pont and Nohria,<br />
2002).<br />
An entrepreneurial venture using R&D alliances within<br />
the new product development process has three distinct<br />
choices <strong>of</strong> partners differentiated by their position along<br />
the industry value chain (Baum et al., 2000). According<br />
to Rothaermel and Deeds (2006), partner can be divided<br />
into three different levels: upstream partners, horizontal<br />
partners and downstream partners. However, the division<br />
will be different to each other due to different types <strong>of</strong><br />
alliances and industrial chains.<br />
In dimensions <strong>of</strong> new energy vehicle industry and<br />
division <strong>of</strong> work, industries’ upstream partners mainly<br />
involve in research developments, and horizontal partners<br />
mainly involve in products fabrication, while downstream<br />
1 Gathered by authors.
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1853<br />
partner mainly involve in sell or services. Theoretically,<br />
upstream partners are research institutions, they basically<br />
do research developments from basis study to tapered<br />
technologies, and they all belong to institutional units.<br />
Horizontal partners should consist <strong>of</strong> automobile<br />
companies and automobile parts manufacturers. Their<br />
duty is to provide productivity. Although both automobile<br />
companies and automobile parts manufacturers are in<br />
same level, in fact, comparing to automobile assemble<br />
companies, automobile parts manufacturers are still a<br />
kind <strong>of</strong> upstream units. Downstream partners mainly<br />
consist <strong>of</strong> transportation companies and dealers etc., they<br />
mainly involve in sells or service providing. Both<br />
horizontal partners and downstream partners belong to<br />
enterprise units.<br />
However, not all alliance involve all these three levels,<br />
some alliance may consist <strong>of</strong> the companies come from<br />
the same level, which is called horizontal alliance. Others<br />
may be more integrated, its member may involve the<br />
most part <strong>of</strong> the industrial chain (Like SOE electric<br />
vehicle alliance or alliance in Beijing and Chongqing),<br />
which named vertical alliance. Factors that affect the<br />
formation <strong>of</strong> the alliance could be various.<br />
III. THE STATUS OF CHINA NEW ENERGY<br />
VEHICEL INDUSTRY ALLIANCE<br />
According to New Energy Automobile Manufacturing<br />
Companies Product Standards and Managing Rules<br />
Published by China Ministry <strong>of</strong> Industry, "New energy<br />
vehicles refers to the vehicles using unconventional<br />
vehicle fuels as a power source (or use the conventional<br />
vehicle fuel power plant using the new device), integrated<br />
power control and vehicle's advanced driving technology,<br />
adopting Advanced technological principles, and with<br />
new technology and structure." Include hybrid vehicles,<br />
electric vehicles (including solar power vehicles), fuel<br />
cell electric vehicles, hydrogen vehicles. And other new<br />
energy sources (such as high energy storage devices,<br />
DME) vehicles and products. In short, the new energy<br />
vehicles refers in the fuel or power systems differ from<br />
traditional internal combustion engine vehicle motor<br />
vehicles (Ministry <strong>of</strong> Industry, 2009).<br />
Currently, among China new energy vehicle industry<br />
alliances, in the geographical point <strong>of</strong> view, are divided<br />
into three categories: the first is established in their<br />
respective regions alliance; second is the alliance<br />
established at the national level; third category is<br />
international Union. The early established alliance in<br />
China is the regional alliance. Beijing new energy vehicle<br />
industry alliance as the first new energy vehicles industry<br />
alliance in China began its operation <strong>of</strong>ficially on March<br />
13, 2009. With Beijing's new energy vehicle industry<br />
alliance formed up, Chongqing, Hubei, Shanghai,<br />
Tianjin, Jilin, Zhejiang, Guangzhou, Anhui and Chengdu<br />
and other places to set up their own new energy vehicle<br />
industry alliance. The number <strong>of</strong> new energy vehicle<br />
industry is raised to 30. These alliances are categorized<br />
into national level, regional level and so-called<br />
international level. There’re both full <strong>of</strong> new energy<br />
vehicles alliances that covered entire industrial chain and<br />
© 2011 ACADEMY PUBLISHER<br />
alliances that only covered the chain <strong>of</strong> production and<br />
market services. The scale <strong>of</strong> these alliances ranges from<br />
60 to 2. There are both alliance with foreign enterprises<br />
and those involve only domestic firms. Meanwhile, the<br />
constitution <strong>of</strong> the various alliances has great differences<br />
among these alliances. Some requested the federal<br />
procurement, and some are not mandatory.<br />
Most <strong>of</strong> the new energy vehicle industry alliances have<br />
strong research and development abilities. According to<br />
the statistics data from 2009, the top 15 patent applicants<br />
in the main manufacturers are: Chery, BYD, FAW,<br />
Changan, and research institutions: Tsinghua University,<br />
Chongqing University, Shanghai Jiaotong University,<br />
Zhejiang University and Jilin University. Meanwhile, the<br />
core members <strong>of</strong> the industry alliances also invested<br />
strong R&D funding. For example, Chongqing Changan<br />
invested 2.52 billion yuan for R&D, FAW Group’s<br />
investment in new energy research and development in<br />
the Eleventh Five-Year period up to 3 million, receiving<br />
65 patents, 38 patents by the U.S. Patent Office (SAC,<br />
2010).<br />
IV. ELEMENTS ANALISIS IN INDUSTRY<br />
ALLIANCE<br />
To show the status <strong>of</strong> China’s new energy industry<br />
alliance better. We used UCINET (Borgatti, Everett, &<br />
Freeman, 2002) to build the network <strong>of</strong> China’s new<br />
energy vehicle industry alliance (Figure 1). We used the<br />
data to build an organization-organization binary matrix<br />
to define the relationship between one independent<br />
company and another.<br />
As Figure 1 shows, most Chinese new energy vehicle<br />
industry alliances are connected to each other except few<br />
ones. Members in these alliances, such as Tsinghua<br />
University, Beijing Institute <strong>of</strong> Technology, Wuhan<br />
Institute <strong>of</strong> Technology etc. are research institutions.<br />
There also have been relationships between automobile<br />
companies and its division companies, such as Dongfeng<br />
Automobile and Dongfeng Yunnan, FAW and Tianjin<br />
FAW etc.. There’re relationships between automobile<br />
parts manufacturers and division companies, such as<br />
Chunlan Electric and Chunlan clean-energy research<br />
institution etc.. Noted some companies have joined more<br />
than one alliance. Among our sample alliances, there’re<br />
only 3 alliances are not connected to others. Hence, these<br />
three alliances are lack <strong>of</strong> connection among the entire<br />
industry alliance.<br />
Our sample alliances consist <strong>of</strong> 183 companies or<br />
institutions. According to results, we divided them into 6<br />
groups: 1) Automobile parts manufacturer; 2)<br />
Automobile company; 3) Research institution; 4)<br />
Transportation company; 5) Government agency; 6)<br />
Financial Institution. Among this companies and<br />
institutions, automobile parts manufacturer has the largest<br />
number, 63; automobile company is second on the list,<br />
50. The third one is research institution, 48. Followed<br />
with, transportation company (9), government agency (9)<br />
and financial institution (5). (Table I) Some <strong>of</strong> the<br />
companies involved more than one business, like<br />
companies that both doing research and product
1854 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
manufacturing, we distinguish them according its major<br />
role in the alliance. Other institutions looked like<br />
enterprise on surface, which mainly involve in<br />
coordination and management will be treated as<br />
government agencies. In fact there’s no place for<br />
government agencies and financial institutions in the<br />
industry chain, so we simply can’t put them into industry<br />
chain.<br />
Considering automobile parts manufacturer,<br />
automobile company, transportation company and<br />
26%<br />
research institution, according to the previews theory,<br />
there’re 48 upstream partners (Research Institution),<br />
26.2% in total; 113 horizontal partners (automobile parts<br />
manufacturer and automobile company), 61.7% in total; 9<br />
downstream partners (transportation company), 4.9% in<br />
total. Consequently, there’s more than half <strong>of</strong> the<br />
members in new energy vehicle industry alliance is in the<br />
position <strong>of</strong> horizontal level.<br />
Figure 1. China’s new energy vehicle industry alliance network<br />
Source: all the data gathered by authors<br />
5%<br />
5%<br />
3%<br />
27%<br />
34%<br />
Automobile Parts Manufacturer<br />
Automobile Company<br />
Research Institution<br />
Transportation Company<br />
Government Agency<br />
Financial Institution<br />
Figure 2. Members in China’s new energy vehicle industry alliance network<br />
Source: all the data get by authors<br />
From the point <strong>of</strong> view <strong>of</strong> cooperation, the total<br />
number <strong>of</strong> cooperative relationships involved in all<br />
alliances is 496. Among these cooperative relationships,<br />
partnerships between automobile companies and<br />
automobile parts manufacturers reached 126 (25.4%);<br />
partnerships between automobile companies reached 110<br />
(22.2%); partnerships between automobile companies and<br />
research institutions reached 96 (19.4%); partnerships<br />
between automobile companies and government agencies<br />
reached 39 (7.9%); partnerships between research<br />
© 2011 ACADEMY PUBLISHER<br />
institutions and automobile parts manufacturers reached<br />
23 (4.6%); partnerships between automobile parts<br />
manufacturers reached 21 (4.2%); partnerships between<br />
automobile parts manufacturers and government agencies<br />
reached 17 (3.4%); partnerships between research<br />
institutions reached 14 (2.8%); partnerships between<br />
research institution and government agencies reached 11<br />
(2.2%); partnerships between automobile companies and<br />
transportation companies reached 5 (1.0%); partnerships<br />
between government agencies and transportation
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1855<br />
companies reached 2 (0.4%); partnerships between<br />
government agencies and financial institutions reached 2<br />
(0.4%); partnerships between government agencies<br />
reached 1 (0.2%). In general, the major types <strong>of</strong><br />
partnerships are automobile company-automobile parts<br />
manufacturers, automobile company- automobile<br />
company and automobile company-research institution<br />
(66.9% in total).<br />
TABLE I. CHINA’S NEW ENERGY VEHICLE INDUSTRY ALLIANCE PARTNERSHIP<br />
Automobile Parts Manufacturer Partnership: Research Institution Partnership:<br />
Automobile Company 126 65.6% Automobile Company 96 64.0%<br />
Research Institution 23 12.0% Automobile Parts Manufacturer 23 15.3%<br />
Automobile Parts Manufacturer 21 10.9% Research Institution 14 9.3%<br />
Government Agency 17 8.9% Government Agency 11 7.3%<br />
Financial Institution 5 2.6% Financial Institution 6 4.0%<br />
Automobile Company Partnership: Government Agency Partnership:<br />
Automobile Parts Manufacturer 126 32.1% Automobile Company 39 54.2%<br />
Automobile Company 110 28.1% Automobile Parts Manufacturer 17 23.6%<br />
Research Institution 96 24.5% Research Institution 11 15.3%<br />
Government Agency 39 9.9% Financial Institution 2 2.8%<br />
Financial Institution 14 3.6% Transportation Company 2 2.8%<br />
Transportation Company 7 1.8% Government Agency 1 1.4%<br />
Transportation Company Partnership: Financial Institution Partnership:<br />
Automobile Company 7 77.8% Automobile Company 14 66.7%<br />
Government Agency 2 22.2% Automobile Parts Manufacturer 5 23.8%<br />
*. Source: all the data gathered by authors<br />
As TABLE I shows, we can generally find different<br />
traits <strong>of</strong> different types <strong>of</strong> partners. The major<br />
partnerships <strong>of</strong> automobile parts manufacturers are those<br />
with automobile companies; the major partnerships <strong>of</strong><br />
automobile companies are those with automobile parts<br />
manufacturers, other automobile companies and research<br />
institutions; the major partnerships <strong>of</strong> Transportation<br />
companies are those with automobile companies; the<br />
major partnerships <strong>of</strong> research institutions are those with<br />
automobile companies; the major partnerships <strong>of</strong><br />
government agencies are those with automobile<br />
companies; the major partnerships <strong>of</strong> financial<br />
institutions are those with automobile companies. In<br />
results, automobile company is the focal role <strong>of</strong> the entire<br />
new energy vehicle industry alliance.<br />
Generally, there’re four types <strong>of</strong> alliance formation we<br />
can find in Figure 1:<br />
(a) Dyadic alliance that automobile company acts as<br />
focal firm: This kind <strong>of</strong> alliance is formed in tree shape in<br />
Figure 2. Basically one company or more established<br />
partnership with focal firm, and no partnership existed<br />
among non-focal firm.<br />
(b) Dyadic alliance that government agency acts as<br />
focal firm: This kind <strong>of</strong> alliance is much similar to the<br />
first one, the only difference is that the focal firm is<br />
government agency. Meanwhile, no partnership existed<br />
among non-focal firm.<br />
(c) Alliance network: Partnerships existed between<br />
multiple companies and institutions, each alliance<br />
© 2011 ACADEMY PUBLISHER<br />
Government Agency 2 9.5%<br />
member established partnership to every other member,<br />
thus to form up a network.<br />
(d) Compound Alliance: Compound alliance refers to<br />
an alliance network with dyadic relationships, which<br />
means not all companies established partnerships to each<br />
other.<br />
We find Dyadic alliance that automobile company acts<br />
as focal firm is the main stream <strong>of</strong> alliance formation in<br />
Chinese new energy vehicle industry alliance. Others like<br />
Dyadic alliance that government agency acts as focal<br />
firm; Alliance network and Compound Alliance are lesser<br />
when compared with the first alliance formation.<br />
V. CONCLUSION<br />
After our analysis elements within China’s new energy<br />
vehicle industry alliance, our findings are summarized as<br />
follow:<br />
(a) Chinese new energy vehicle industry alliance<br />
formation can generally be divided into four types:<br />
Dyadic alliance that automobile company acts as focal<br />
firm, Dyadic alliance that government agency acts as<br />
focal firm, Alliance network and Compound Alliance.<br />
Dyadic alliance that automobile company acts as focal<br />
firm is the main stream <strong>of</strong> alliance formation in Chinese<br />
new energy vehicle industry alliance, while others are<br />
lesser when compared with the first alliance formation.<br />
(b) There’s more than half <strong>of</strong> the members in new<br />
energy vehicle industry alliance is in the position <strong>of</strong><br />
horizontal level. In general, the major types <strong>of</strong><br />
partnerships are automobile company-automobile parts
1856 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
manufacturers, automobile company- automobile<br />
company and automobile company-research institution.<br />
(c) Automobile company is the focal role <strong>of</strong> the entire<br />
new energy vehicle industry alliance. The major<br />
partnerships <strong>of</strong> automobile parts manufacturers are those<br />
with automobile companies; the major partnerships <strong>of</strong><br />
automobile companies are those with automobile parts<br />
manufacturers, other automobile companies and research<br />
institutions; the major partnerships <strong>of</strong> Transportation<br />
companies are those with automobile companies; the<br />
major partnerships <strong>of</strong> research institutions are those with<br />
automobile companies; the major partnerships <strong>of</strong><br />
government agencies are those with automobile<br />
companies; the major partnerships <strong>of</strong> financial<br />
institutions are those with automobile companies.<br />
In summarize, we conclude that automobile company<br />
is the key factor <strong>of</strong> the entire China’s new energy vehicle<br />
industry alliance. Thus, improving and strengthening the<br />
cooperation between automobile companies and other<br />
members in the alliance is the essential to the<br />
development <strong>of</strong> alliance. This paper provides the basis for<br />
quantification study in the future. Further research will<br />
involve in theoretical and empirical studies on innovation<br />
model.<br />
REFERENCES<br />
[1] Baum, Joel A. C., Calabrese Tony. Silverman. Brian<br />
S.(2000). Don't go it alone: alliance network composition<br />
and startups' performance in Canadian biotechnology.<br />
© 2011 ACADEMY PUBLISHER<br />
Strategic Management <strong>Journal</strong>. Special Issue: Strategic<br />
Networks, 21(3): 267–294.<br />
[2] Das, T.K., Teng, B.-S. (2000). A resource-based theory <strong>of</strong><br />
strategic alliances. <strong>Journal</strong> <strong>of</strong> Management, 26 (1):31–60.<br />
[3] Garcia-Pont, C. and N. Nohria (2002) “Local versus<br />
Global Mimetism: The Dynamics <strong>of</strong> Alliance Formation in<br />
the Automobile Industry”, Strategic Management <strong>Journal</strong>,<br />
23, 307-21.<br />
[4] Gulati Ranjay(1998).l Alliances and networks. Strategic<br />
Management <strong>Journal</strong>.Special Issue: Editor's Choice. 19(4):<br />
293-317.<br />
[5] Gulati, R., Nohria, N., Zaheer, A. (2000). Strategic<br />
Networks, Strategic Management <strong>Journal</strong>, 21: pp. 203-215.<br />
[6] Japan Electric Vehicle Alliance.<br />
www.CHAdeMOchademo.com.<br />
[7] Jarillo, J. Carlos(1998). On strategic networks. Strategic<br />
Management <strong>Journal</strong>. 9(1): 31–41.<br />
[8] Ministry <strong>of</strong> Industry . New Energy Automobile<br />
Manufacturing Companies Product Standards and<br />
Managing Rules [EB /<br />
OL].http://www.miit.gov.cn/n11293472/n11293832/nl129<br />
3907/n11368223112425871.html.2009-06-25.<br />
[9] Rothaermel, Frank T. & Deeds, David L., 2006. "Alliance<br />
type, alliance experience and alliance management<br />
capability in high-technology ventures," <strong>Journal</strong> <strong>of</strong><br />
Business Venturing, Elsevier, vol. 21(4):429-460<br />
[10] State-owned Assets Supervision and Administration<br />
Commission. www.sasac.gov.cn/<br />
[11] Stephen P Borgatti, M G Everett, Linton C Freeman(2002).<br />
Ucinet for Windows: S<strong>of</strong>tware for Social Network<br />
Analysis. Harvard Analytic Technologies, 2002.
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1857<br />
Efficiency Evaluation Information System Based<br />
on Data Envelopment Analysis<br />
Jing Han<br />
Department <strong>of</strong> Economics and Management, Huainan Vocational & Technical College, Huainan, Anhui, China<br />
E-mail: hanjing623@163.com<br />
Malin Song<br />
School <strong>of</strong> Statistics and Applied Mathematics, Anhui University <strong>of</strong> Finance and Economics, Anhui Bengbu, China<br />
Email: songmartin@163.com<br />
Abstract—As data envelopment analysis (DEA) has been<br />
developed both in theory and application, the calculation <strong>of</strong><br />
models become more and more important. Although many<br />
DEA s<strong>of</strong>tware tools have been built for the calculation <strong>of</strong> the<br />
DEA models, there are some deficiencies in embedding them<br />
into enterprise management information system (MIS). As<br />
an extension <strong>of</strong> this work, an idea was generated in this<br />
paper, which could both calculate the DEA and further<br />
support the decision making for decision making units<br />
(DMUs), i.e., the organizations, in the information<br />
environment. This is an attempt to bridge between DEA and<br />
MIS. And we could demonstrate this approach for building<br />
efficiency evaluation information system. Furthermore, an<br />
efficiency evaluation information system <strong>of</strong> company A,<br />
which was built by ourselves, was shown to illustrate our<br />
purpose.<br />
Index Terms—Data envelopment analysis; Management<br />
information system; Efficiency evaluation information<br />
system; Decision support<br />
I. INTRODUCTION<br />
The efficiency evaluation is becoming more and more<br />
considerable in companies' daily management operating.<br />
By looking at the efficiency evaluation, the enterprises<br />
can be aware <strong>of</strong> their specific position, and find out the<br />
gap between them and their competitors, so as to<br />
determine how they could improve the quality <strong>of</strong><br />
products on practical and scientific aspects.<br />
Data envelopment analysis (DEA), as a non-parametric<br />
programming technique, has becoming more and more<br />
popular in evaluating the performance efficiency <strong>of</strong> a set<br />
<strong>of</strong> homogenous decision making units (DMUs). It was<br />
first proposed by Charnes, Cooper and Rhodes in 1978<br />
[1] and extensively applied in multiple inputs and<br />
multiple outputs complex systems. Since the CCR model,<br />
there has been an impressive growth both in theoretical<br />
developments and applications <strong>of</strong> DEA. DEA researchers<br />
have developed a number <strong>of</strong> updated models, such as<br />
BCC model [2], additive model [3], multilevel models [4,<br />
5], super efficiency models [6] and so on. At the same<br />
time, DEA has also been extensively applied in<br />
performance evaluation and benchmarking <strong>of</strong> hospitals,<br />
universities, cities, courts, business firms, and others,<br />
including the performance <strong>of</strong> regions, countries etc [7].<br />
© 2011 ACADEMY PUBLISHER<br />
doi:10.4304/jcp.6.9.1857-1861<br />
However, the applications <strong>of</strong> DEA in the enterprise<br />
management information system are few.<br />
There have been several DEA s<strong>of</strong>tware tools in market.<br />
They can be divided into two groups: one group is<br />
pr<strong>of</strong>essional, such as DEA Solver pro, DEAP, Efficiency<br />
Measurement System (EMS), DEA excel solver and so<br />
on [8]. For these DEA s<strong>of</strong>tware tools, we can get the<br />
results just by inputting the DMUs’ data and choosing the<br />
appropriate model. The other is universal, such as Matlab,<br />
Lingo, Lindo and so on. Using these universal s<strong>of</strong>tware<br />
tools, we must program the procedure by ourselves.<br />
However, all DEA s<strong>of</strong>tware tools mentioned above can’t<br />
be embedded into management information system (MIS)<br />
perfectly. This limits their application strongly. Based on<br />
the theory <strong>of</strong> DEA, this paper tries to set up an evaluation<br />
information system for company A upon the platform <strong>of</strong><br />
DEA in order to supply some useful management<br />
information for it. Company A’s MIS may contain a lot<br />
<strong>of</strong> sub-systems, such as staff information management,<br />
salary management, performance management and so on.<br />
We focused on the evaluation system and its relationship<br />
with others. For simplifying illustration, we construct the<br />
efficiency evaluation information system just based on<br />
CCR and BCC for real company A.<br />
This paper is aimed at evaluating DMUs and<br />
benchmarking by using efficiency evaluation information<br />
system. This approach has some applied advantages,<br />
especially in the information management. Section 2<br />
briefly reviews the traditional DEA models <strong>of</strong> CCR and<br />
BCC. Section 3 introduces the efficiency evaluation<br />
sub-system. In Section 4, we apply the idea to build MIS<br />
<strong>of</strong> a real company A which contains the efficiency<br />
evaluation sub-system. Finally, Conclusions are given in<br />
Section 5.<br />
II. DEA MODELS<br />
We assume that there are n DMUs to be evaluated,<br />
where each DMU contains s different outputs and m<br />
different inputs. We denote the ith input and rth output<br />
for DMUj(<br />
j = 1, 2,..., n<br />
) as ij x ( i = 1, 2,..., m ) and<br />
yrj<br />
( r = 1, 2,..., m ) respectively. We assume that<br />
x ≥ 0 ij ,<br />
yrj ≥ 0<br />
and each DMU must has at least one
1858 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
positive input and one positive output value.<br />
The CCR model, proposed by Charnes et al in 1978<br />
[1], for measuring the technical efficiency <strong>of</strong> the 0 jth<br />
DMU<br />
DMU ( 0 ) was first stated as follows.<br />
s m<br />
CCR<br />
e = max u y / vx<br />
s m<br />
∑ ∑<br />
∑ ∑<br />
r rj0 i ij0<br />
r= 1 i=<br />
1<br />
s.t. uy − vx ≤ 0, j= 1,2,..., n<br />
r rj i ij<br />
r= 1 i=<br />
1<br />
ur, vi ≥ 0, r = 1,2,..., s, i = 1,2,..., m.<br />
(1)<br />
Through the Charnes and Cooper transformation [9]<br />
for linear fractional programming yielded the equivalent<br />
programming problem as follows.<br />
s<br />
CCR<br />
e = max∑µ ryrj0 r=<br />
1<br />
s m<br />
∑ ∑<br />
s.t. µ y − ω x ≤ 0, j = 1, 2,..., n<br />
r rj i ij<br />
r= 1<br />
m<br />
i=<br />
1<br />
∑<br />
i=<br />
1<br />
ω x<br />
i ij0<br />
= 1<br />
µ r, ωi<br />
≥ 0, r = 1, 2,..., s, i = 1, 2,..., m.<br />
(2)<br />
for which the LP dual problem is<br />
CCR<br />
e = min θ<br />
n<br />
∑<br />
s.t. λ x + s = θx<br />
, i = 1,2,..., m<br />
j=<br />
1<br />
n<br />
∑<br />
j=<br />
1<br />
j ij<br />
−<br />
i i0<br />
λ y − s = y , r = 1,2,..., s<br />
j rj<br />
+<br />
r r0<br />
λ j ≥ 0, j = 1,2,..., n.<br />
(3)<br />
Model (3) is sometimes referred to as the “Farrell<br />
model” because it is the one used in Farrell [10]. In the<br />
economics portion <strong>of</strong> the DEA literature, it is said to<br />
conform to the assumption <strong>of</strong> “strong disposal”, because<br />
it ignores the presence <strong>of</strong> non-zero slacks. Besides, it is<br />
also under the assumption <strong>of</strong> constant returns to scale<br />
(CRS).<br />
Then, based on CCR model, Banker, Charnes and<br />
Cooper built the BCC model as follows [2].<br />
BCC<br />
e = min θ<br />
n<br />
∑<br />
s.t. λ x + s = θx<br />
, i = 1,2,..., m<br />
j=<br />
1<br />
n<br />
∑<br />
j=<br />
1<br />
n<br />
∑<br />
j=<br />
1<br />
j ij<br />
−<br />
i i0<br />
λ y − s = y , r = 1,2,..., s<br />
j rj<br />
+<br />
r r0<br />
λ = 1<br />
j<br />
λ ≥ 0, j = 1,2,..., n.<br />
j<br />
© 2011 ACADEMY PUBLISHER<br />
(4)<br />
In the economics portion <strong>of</strong> the DEA literature, the<br />
BCC model, that is (4), are under the assumption <strong>of</strong><br />
variable returns to scale (VRS). BCC model could be<br />
used to determine the returns to scale, including<br />
decreasing, constant and increasing.<br />
DMU 0 is efficient if and only<br />
For model (3) and (4),<br />
− * + *<br />
*<br />
if θ = 1 s = s = 0<br />
DMU<br />
and i r for all i and r. 0 is<br />
*<br />
*<br />
weakly efficient if θ = 1 s 0<br />
and i<br />
− ≠<br />
and (or)<br />
*<br />
s 0 r<br />
+ =<br />
for some i and r in some alternate optima.<br />
DMU *<br />
0 is inefficient if θ < 1 [11]. Assume the<br />
CCR and BCC scores <strong>of</strong> a DMU are CCR<br />
e and BCC<br />
e<br />
respectively. The scale efficiency is defined by<br />
scale CCR* BCC*<br />
e = e / e [12].<br />
The following conditions identify the situation for<br />
returns to scale (RTS) for the CCR model given in (3).<br />
xˆ ˆ<br />
(i) Increasing RTS prevail at ( 0 : y 0)<br />
if and only if<br />
n<br />
∑<br />
*<br />
λ < 1 j<br />
for all optimal solutions.<br />
xˆ ˆ<br />
(ii) Decreasing RTS prevail at ( 0 : y 0)<br />
if and only if<br />
j=<br />
1<br />
n<br />
∑<br />
*<br />
λ > 1 j<br />
for all optimal solutions.<br />
xˆ ˆ<br />
(iii) Constant RTS prevail at ( 0 : y 0)<br />
if and only if<br />
j=<br />
1<br />
n<br />
∑<br />
j=<br />
1<br />
*<br />
λ = 1 j<br />
for at least one optimal solution [13].<br />
III. BASIC FUNCTION OF EFFICIENCY EVALUATION<br />
INFORMATION SYSTEM<br />
We thought that the efficiency evaluation sub-system<br />
should contain at least three parts. One is evaluation<br />
among it and its homogeneous DMUs, another is<br />
evaluation among its performance in different time, and<br />
the last is benchmarking.<br />
The first evaluation aims to determine the efficiency <strong>of</strong><br />
DMU 0 when it compares with other DMUs from<br />
cross-sectional data. The<br />
DMU 0 can be aware <strong>of</strong> its<br />
location exactly among the same kind <strong>of</strong> DMUs by the<br />
results above. It is very useful for the DMU to understand<br />
the gap between itself and other DMUs.<br />
The second evaluation aims to determine the efficiency<br />
<strong>of</strong><br />
DMU 0 in series times. Through the results, we can<br />
know whether the DMU’s performance has been<br />
improved. Here, we assume the DMU is a company.<br />
Through specifically understanding the developments and<br />
trends <strong>of</strong> the company, we can do better preparation for<br />
its future development, in order to avoid irreparable loss<br />
caused by the company management’s delaying.
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1859<br />
An important part <strong>of</strong> organizational planning and<br />
control is the selection <strong>of</strong> proper performance<br />
benchmarks [14]. Benchmarking is a means to evaluate<br />
their own businesses and study other organizations. It<br />
takes the internal or external best practices in business<br />
enterprises as its own internal development goals, and<br />
then applies the goal to their business practice. Through<br />
the results, we can determine the practical and scientific<br />
path for efficiency improvement.<br />
Now, we began to introduce the efficiency evaluation<br />
sub-system in details. This sub-system contains three<br />
parts. The first part was evaluation among Company A<br />
and its homogeneous companies, the second part was<br />
evaluation among its performance in different time, and<br />
IV. REAL COMPANY A’S EFFICIENCY EVALUATION<br />
INFORMATION SYSTEM<br />
Before introducing the efficiency evaluation<br />
sub-system, we should introduce the management<br />
information system briefly which developed by ourselves.<br />
It contained five parts: system management, basic<br />
information management, sale data analysis, efficiency<br />
evaluation, data inquiry. The interface was designed as<br />
Fig.1.<br />
FIG.1 COMPANY A MANAGEMENT INFORMATION SYSTEM INTERFACE<br />
the last was benchmarking. The system’s interfaces<br />
were shown as Fig.2, 3, 4 respectively when they were<br />
running based on CCR model.<br />
FIGURE.2 EVALUATION OF OPERATIONAL EFFICIENCY THROUGH CROSS-SECTIONAL DATA<br />
From Figure.2, we could gain the companies’<br />
efficiency value and their sizes intuitionally.<br />
© 2011 ACADEMY PUBLISHER<br />
From Figure.3, we could gain the company’s<br />
efficiency value and their trend in time series<br />
intuitionally.
1860 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
FIGURE.3 EVALUATION OF OPERATIONAL EFFICIENCY OF COMPANY A OVER THE YEARS<br />
From Figure.3, we could gain the company’s<br />
efficiency improvement proposals for efficient and its<br />
returns to scale intuitionally.<br />
Part <strong>of</strong> the data used for DMUs’ evaluation was<br />
generated automatically by the information system. The<br />
others were inputted by an interface from outside.<br />
The calculations <strong>of</strong> the DEA models were operated at<br />
computer background. The results were stored in<br />
database. When we need the related data, we could call<br />
them by programming directly. This approach could<br />
reduce the running time apparently. For example, if we<br />
should evaluation the Company A’s efficiency among all<br />
DMUs, we just called the evaluation results in database<br />
by using SQL language instead <strong>of</strong> calculating the DEA<br />
models, that was, linear programming. If operations are<br />
frequent, the former’s advantages, which had low time<br />
and space complexity, would show out.<br />
V. CONCLUSION<br />
DEA has been used in many fields popularly.<br />
Nowadays, DEA has been used widely in many fields.<br />
There are several s<strong>of</strong>tware tools for dealing with DEA<br />
© 2011 ACADEMY PUBLISHER<br />
FIGURE.4 BENCHMARKING BASED ON CCR<br />
model, including pr<strong>of</strong>essional and universal tools. All<br />
these tools can deal with some DEA models. However,<br />
they can not become part <strong>of</strong> enterprise management<br />
information system perfectly, which is popularly used for<br />
management in our firms now. To build efficiency<br />
evaluation information system is very useful. This paper<br />
briefly introduces the DEA model and the parts which the<br />
system should contain. At last, for illustrating our idea,<br />
we take an efficiency evaluation information system <strong>of</strong> a<br />
real company A as an example.<br />
As one <strong>of</strong> the solutions, our proposed approach is only<br />
one way to integrate DEA into management information<br />
system. This will help managers grasp the state <strong>of</strong> their<br />
company among the same kind companies better. It is<br />
also useful to gain the company development trend<br />
during the time series. Last but by no means least, the<br />
system can make benchmarking and propose some useful<br />
suggestions for company too. However, our system is<br />
based on personal platform. This may limit its usage in<br />
some degree. Therefore some extensions can be studied<br />
in the future. The next work we will do is to build an<br />
efficiency evaluation information system based on Web.
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1861<br />
REFERENCE<br />
[1] Charnes, A., Cooper, W.W., and Rhodes, E., “Measuring<br />
the efficiency <strong>of</strong> decision making units”, European <strong>Journal</strong><br />
<strong>of</strong> Operational Research, vol. 2, No. 6, 1978, pp. 429–444.<br />
[2] Banker, R.D., Charnes, A., Cooper, W.W., “Some models<br />
for estimating technical and scale inefficiencies in data<br />
envelopment analysis”, Management Science, Vol. 30, No.<br />
9, 1984, pp. 1078-1092.<br />
[3] Charnes, A., Cooper, W.W., Golany, B., Seiford, L.M.,<br />
Stutz, J., “Foundations <strong>of</strong> data envelopment analysis and<br />
Pareto-Koopmans empirical production functions”. <strong>Journal</strong><br />
<strong>of</strong> Econometrics, Vol. 30, No. 9, 1985, pp. 91-107.<br />
[4] Fare, R., Grosskopf, S., Intertemporal Production<br />
Frontiers: With Dynamic DEA. Kluwer Academic, Boston,<br />
MA, 1996.<br />
[5] Liang, L., Yang, F., Cook, W. D., Zhu, Joe., DEA models<br />
for supply chain efficiency evaluation, Annals <strong>of</strong><br />
Operations Research, Vol. 145, No. 1, 2006, pp. 35-49.<br />
[6] Andersen, P., Petersen, N.C., “A procedure for ranking<br />
efficient units in data envelopment analysis”, Management<br />
Science, Vol. 39, No. 10, 1993, pp. 1261–1264.<br />
[7] Cooper, W.W., Seiford L. M., Zhu Joe (Eds.), Data<br />
envelopment analysis, Kluwer Academic <strong>Publisher</strong>s,<br />
London, 2004, pp. 1-2.<br />
[8] Cooper, W.W., Seiford L. M., Zhu Joe (Eds.), Data<br />
envelopment analysis, Kluwer Academic <strong>Publisher</strong>s,<br />
London, 2004, pp. 539-564.<br />
© 2011 ACADEMY PUBLISHER<br />
[9] Charnes, A., Cooper, W.W., “Programming with linear<br />
fractional functionals”, Naval Research logistics quarterly,<br />
vol. 9, No. 3-4, 1962, pp. 181-185.<br />
[10] Farrell M.J., “The measurement <strong>of</strong> productive efficiency”,<br />
<strong>Journal</strong> <strong>of</strong> Royal Statistic Society. Series A, Vol. 120, No.<br />
3, 1957, 253-281.<br />
[11] Cooper, W.W., Seiford L. M., Zhu Joe (Eds.), Data<br />
envelopment analysis, Kluwer Academic <strong>Publisher</strong>s,<br />
London, 2004, pp. 8-13.<br />
[12] Cooper W. W., Seiford L. M., Tone Kaoru (Eds.), Data<br />
envelopment analysis: A comprehensive text with models,<br />
Applications, references, and DEA-Solver S<strong>of</strong>tware,<br />
Kluwer academic publishers, Boston, 2000, pp. 136-138.<br />
[13] Banker, R. D., Cooper, W. W., Seiford, L. M., Robert, M.<br />
Thrall, Zhu Joe, ‘Returns to scale in different DEA<br />
models’, European <strong>Journal</strong> <strong>of</strong> Operational Research, vol.<br />
154, No. 2, 2004, pp. 345–362.<br />
[14] Thierry Post, Jaap Spronk, Performance benchmarking<br />
using interactive data envelopment analysis, European<br />
<strong>Journal</strong> <strong>of</strong> Operational Research, Vol. 115, No. 3, 1999,<br />
pp. 472-487.<br />
Jing Han is an associate pr<strong>of</strong>essor in School <strong>of</strong> Economics and<br />
Management at Huainan Vocational & Technical College. Her<br />
major field <strong>of</strong> study includes electronic commerce, and<br />
enterprise management (E-mail: hanjing623@163.com).
1862 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
An Optimal Inventory Control Model for a<br />
Supply Chain with Shortage Constraints<br />
Yinkuan Gu<br />
Management School/Anhui University <strong>of</strong> Technology, Maanshan, China<br />
Email: first.author@hostname1.org<br />
Hongxia Zhang<br />
Management School/Anhui University <strong>of</strong> Technology, Maanshan, China<br />
Abstract—The article studies, with the level constraints in<br />
short supply, the inventory decision model <strong>of</strong> the minimum<br />
total annual cost <strong>of</strong> the supply chain which, composed <strong>of</strong> a<br />
single supplier and multiple buyers, involving supplier’s<br />
lead time as a decision variable, replacing the cost <strong>of</strong><br />
shortages with the level in short supply, and has solved the<br />
difficult problem in the practice.<br />
Index Terms—Level in short supply, supply chain, lead time<br />
I. INTRODUCTION<br />
To meet the needs <strong>of</strong> customers timely, businesses<br />
must maintain higher inventory levels to avoid shortages.<br />
However, high inventory levels are <strong>of</strong>ten associated with<br />
high inventory costs, many companies strive to reduce<br />
production or lead time cycle, and thus have a<br />
corresponding reduction in inventory.<br />
The current study on lead time and inventory decisions<br />
mostly focused on individual enterprises. Liao, C.J., and<br />
Shyu, C.H.(1992) , for the perpetual inventory system,<br />
divided the activities such as the procurement, order<br />
processing, production, transportation, storage test, <strong>of</strong> the<br />
lead time period into n-independent component parts and<br />
each part has its own different time limit as well as<br />
operating costs, to analyze the best lead time and reorder<br />
point. Ben-Daya, M, and Raouf, A. set the EOQ into Liao<br />
and Shyn’s study. Ouyang, LY., Yeh, NC., and Wu, KS.,<br />
discussed the mixture inventory model with backorders<br />
and lost sales while some customers not wish to wait in<br />
the situation <strong>of</strong> out <strong>of</strong> stock. For being quite difficult to<br />
assess the unit costs for out <strong>of</strong> stock in practice, Ouyang,<br />
LY. and Chuang, BR.(2000) replaced the shortage cost<br />
with the level <strong>of</strong> out stock as the parameters <strong>of</strong> measuring<br />
shortage. However, the shortening <strong>of</strong> the lead time<br />
depends on the improvement and cooperation <strong>of</strong> the<br />
upstream and downstream <strong>of</strong> the supply chain. For this,<br />
Ben-Daya, M. and Hariga, M. (2004) studied the<br />
integrated single vendor single buyer model with<br />
stochastic demand and variable lead time to measure the<br />
optimal inventory for the perpetual inventory system.<br />
This paper studies, based on the aforementioned<br />
literature, with the level constraints in short supply, the<br />
inventory decision model <strong>of</strong> the minimum total annual<br />
cost <strong>of</strong> the supply chain which, being composed <strong>of</strong> a<br />
© 2011 ACADEMY PUBLISHER<br />
doi:10.4304/jcp.6.9.1862-1867<br />
single supplier and multiple buyers, involving supplier's<br />
lead time as a decision variable, replacing the cost <strong>of</strong><br />
shortages with the level <strong>of</strong> short supply.<br />
II. ASSUMPTIONS AND PARAMETERS<br />
A. Assumptions<br />
Underlying assumptions <strong>of</strong> the proposed model in this<br />
paper as follows:<br />
(a) The buyers and suppliers are based on the periodic<br />
inventory system;<br />
(b) The suppliers take a batch production methods,<br />
common distribution strategy, and supply in the same<br />
cycle;<br />
(c)The production cycle <strong>of</strong> the suppliers is integral<br />
multiples to the same provision cycle above;<br />
(d)The production rate, delivery time, inventory<br />
holding costs, unit order number, transportation costs,<br />
order activity cost <strong>of</strong> the suppliers known as a fixed<br />
constant;<br />
(e)The amount <strong>of</strong> the buyer’s demand <strong>of</strong> the lead time<br />
is the same as the amount <strong>of</strong> the supplier’s requirement in<br />
the production cycle for a random variable, and following<br />
the normal distribution;<br />
(f)The target level <strong>of</strong> the suppliers and the buyers’<br />
order are the average demand <strong>of</strong> the lead time plus the<br />
safety stock quantity.<br />
B. Parameters<br />
The parameters and their symbols <strong>of</strong> the model used as<br />
follows:<br />
n = The total number <strong>of</strong> the buyers;<br />
di = The demand per unit time <strong>of</strong> the buyer I, average<br />
2<br />
as<br />
d σ<br />
i<br />
di<br />
, variance as<br />
D = The demand per unit time <strong>of</strong> the suppler, average<br />
as<br />
n<br />
∑ di<br />
i=<br />
1 , variance as<br />
n<br />
∑<br />
i=<br />
1<br />
σ<br />
2<br />
d i<br />
;<br />
P = Productivity <strong>of</strong> the suppler(P>D);<br />
T = The buyer’s common replenishment cycle;<br />
L = The supplier’s delivery time to the buyer’s orders;<br />
K = The shipping times <strong>of</strong> the suppliers in each<br />
production cycle;
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1863<br />
xi = The demand <strong>of</strong> the buyer i <strong>of</strong> the warranty period<br />
(T+L),a random variable, following the probability<br />
density function fR+L(xi), average as<br />
2<br />
d i<br />
( T + L)<br />
σ d ( T + L)<br />
i<br />
variance as<br />
;<br />
y = The demand <strong>of</strong> a production cycle <strong>of</strong> the suppler, a<br />
random variable, following the probability density<br />
function fKT(xi),average as<br />
n<br />
KT ∑<br />
i=<br />
1<br />
2<br />
d i<br />
n<br />
KT ∑ di<br />
i=<br />
1 ,variance as<br />
σ<br />
;<br />
Ri = The target level <strong>of</strong> replenishment <strong>of</strong> the buyer I;<br />
Zi = The safety factor <strong>of</strong> the <strong>of</strong> the lead time to the<br />
buyer i, set as a decision variable;<br />
Rv = The target level <strong>of</strong> the supplier’s production;<br />
Zv = The safety factor <strong>of</strong> the <strong>of</strong> the lead time to the<br />
supplier, set as a decision variable;<br />
αi = The shortage upper limit <strong>of</strong> the buyer i;<br />
αv = The shortage upper limit <strong>of</strong> the supplier;<br />
L0 = The buyer’s lead time at the time <strong>of</strong> the system;<br />
C (L) = The increased crashing cost <strong>of</strong> shortening the<br />
delivery time, a non-increasing function for L,and C (L0)<br />
=0;<br />
F = Basic ordering and transportation costs (USD /<br />
times);<br />
Fi = The ordering and transportation costs <strong>of</strong> the buyer<br />
i (USD / times);<br />
A = Suppliers’ batch adjustment costs(USD / times);<br />
hi = The holding costs <strong>of</strong> a unit <strong>of</strong> inventory <strong>of</strong> the<br />
buyer i (USD / times/unit product/unit time);<br />
hv = The holding costs <strong>of</strong> a unit <strong>of</strong> inventory <strong>of</strong> the<br />
suppler (USD / times/a unit product/unit time);<br />
ECi = The expected total inventory cost per unit time<br />
<strong>of</strong> the buyer I;<br />
ECv = The expected total inventory cost per unit time<br />
<strong>of</strong> the suppler;<br />
ETC = The expected total inventory cost per unit time<br />
<strong>of</strong> the supply chain;<br />
III. MODEL ANALYSIS AND SOLUTION<br />
A. Model analysis<br />
Based on the assumptions and parameter settings<br />
above, this paper establishes the following models:<br />
The shortage level <strong>of</strong> the order cycle <strong>of</strong> the buyer i<br />
(shortage probability)<br />
f T + L ( xi<br />
) dx = Φ(<br />
Z i )<br />
=<br />
∫Ri ∞<br />
The shortage level <strong>of</strong> the production cycle <strong>of</strong> the<br />
supplier (shortage probability)<br />
f ( y)<br />
dy = Φ(<br />
Z )<br />
= ∫∞ Rv<br />
KT<br />
v<br />
By the previous assumptions, then there are<br />
Φ(Zi) αi,Φ(Zv) αv,i = 1,2,…,n<br />
Items <strong>of</strong> the related costs associated with the buyer,<br />
include the expected order cost, transport cost, expected<br />
holding cost. The order and transport cost <strong>of</strong> the buyer i is<br />
© 2011 ACADEMY PUBLISHER<br />
,<br />
Fi/T, the average inventory level <strong>of</strong> the buyer i can be<br />
estimated as:<br />
( Ri − d i L)<br />
+ ( R i − d i L − d iT<br />
)<br />
d iT<br />
= R i − d i L −<br />
2<br />
2 (1)<br />
Where, the target level <strong>of</strong> replenishment <strong>of</strong> the buyer i<br />
Ri<br />
= di<br />
( T + L)<br />
+ Ziσ<br />
d T + L<br />
i<br />
set as<br />
, then the expected<br />
stock holding cost per unit time <strong>of</strong> the buyer i can be<br />
⎛<br />
⎞<br />
⎜<br />
d T<br />
h + Z T + L ⎟<br />
⎜<br />
i ⎟<br />
estimated as ⎝<br />
⎠<br />
d i<br />
i<br />
i σ<br />
2<br />
. In addition, while<br />
the suppliers distribute in a joint way, the buyers share<br />
the cost <strong>of</strong> order and transportation per times, and then<br />
the cost <strong>of</strong> order and transportation per unit time any<br />
buyer burdened is F/T.<br />
To the related storage costs <strong>of</strong> the suppler, the<br />
adjustment costs per unit time is A/KT, according to the<br />
study <strong>of</strong> Ben-Daya and Hariga (2004)①, the average<br />
inventory level <strong>of</strong> the suppler can be estimated as:<br />
n<br />
n ⎡<br />
⎤<br />
∑ d iT<br />
⎢ ∑ d i<br />
⎥<br />
n<br />
i = 1<br />
i = 1 ⎢ ( 2 − K ) + K − 1⎥<br />
+ R v − KT ∑ d i<br />
2 ⎢ P<br />
⎥<br />
i = 1<br />
⎢<br />
⎣<br />
⎥<br />
⎦<br />
(2)<br />
For the target stock level <strong>of</strong> the suppler is<br />
n<br />
n<br />
2<br />
v = KT∑<br />
di<br />
+ Z v KT∑σ<br />
di<br />
i=<br />
1<br />
i=<br />
1<br />
R<br />
, then the expected<br />
inventory carrying cost per unit time <strong>of</strong> the suppler is:<br />
n<br />
n<br />
⎧ ⎡<br />
⎤⎫<br />
⎪∑<br />
d iT<br />
∑<br />
⎪<br />
⎪<br />
⎢ d i<br />
⎥<br />
n<br />
i=<br />
1 i=<br />
1<br />
⎪<br />
2<br />
h ⎨ ⎢ ( ) ⎥<br />
v 2 − K + K − 1 ⎬ + Z v KT ∑ σ d i<br />
⎪ 2 ⎢ P<br />
⎥⎪<br />
i=<br />
1<br />
⎪ ⎢<br />
⎥<br />
⎩ ⎣<br />
⎦⎪⎭<br />
Considering that the lead time <strong>of</strong> the suppler can be<br />
adjusted by the extra crashing cost, then the lead time can<br />
be regard as a decision variable, so the crash cost per unit<br />
time C(L)/T <strong>of</strong> shortening the lead time should be<br />
considered. To the C(L)/T, to be set as a step function in<br />
the paper, it is, the crash cost will not change in a certain<br />
range, if beyond the certain range, it will be higher, then<br />
the function can designed as:<br />
⎧r0<br />
L = L0<br />
⎪<br />
r1<br />
L1<br />
≤ L < L0<br />
C(<br />
L)<br />
= ⎨<br />
⎪M<br />
M<br />
⎪<br />
⎩rb<br />
Lb<br />
≤ L < Lb<br />
−1<br />
Where, ri (i = 0,1,…,b) is a parameter for the crash<br />
cost, Lb is the shortest lead time.<br />
In summary, there are: the total cost <strong>of</strong> the system per<br />
unit time = the expected inventory total cost <strong>of</strong> the buyer<br />
per unit time + the expected inventory total cost <strong>of</strong> the<br />
suppler per unit time + the crash cost <strong>of</strong> shortening the<br />
lead time per unit time. It is:<br />
①Ben-Daya, M. and Hariga, M., Integrated single vendor single buyer<br />
model with stochastic demand and variable lead time[J], International<br />
<strong>Journal</strong> <strong>of</strong> Production Economics, 92, 1,2004: 75-80.
1864 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
n<br />
C<br />
( )<br />
( L)<br />
ETC K,<br />
T , L,<br />
Z,<br />
Z v = ∑ ECi<br />
+ EC v +<br />
T<br />
i=<br />
1<br />
⎧ n<br />
n<br />
n<br />
⎡ ⎛ ⎞ ⎛ ⎞⎤<br />
⎫<br />
⎪hv<br />
∑ d i ⎢ ⎜ ∑ d i ⎟ ⎜ 2∑<br />
d ⎟<br />
n<br />
i ⎥ n ⎪<br />
1 ⎡<br />
⎤ ⎪ i=<br />
1<br />
=<br />
=<br />
⎪<br />
∑ ( ) ⎨ ⎢ ⎜ i 1 ⎟ + ⎜ i 1<br />
− ⎟ d<br />
⎥<br />
ihi<br />
= ⎢F<br />
+ Fi<br />
+ C L ⎥ + T K 1 −<br />
1 + ∑ ⎬<br />
T ⎣ = ⎦ ⎪ ⎢ ⎜ ⎟ ⎜ ⎟<br />
i 1<br />
2<br />
P P ⎥ i=<br />
1 2 ⎪<br />
⎪ ⎢ ⎜ ⎟ ⎜ ⎟⎥<br />
⎩ ⎣ ⎝ ⎠ ⎝ ⎠⎦<br />
⎪<br />
⎭<br />
n<br />
+ h Z<br />
n<br />
2<br />
+ hvZ<br />
v KT∑<br />
σ d + ∑ hiZ<br />
iσ<br />
d T + L<br />
i<br />
i<br />
i=<br />
1 i=<br />
1<br />
(3)<br />
By the previous assumption, there are<br />
Φ(Zi)≦αi,Φ(Zv)≦αv,i = 1,2,…,n<br />
Equation (3) is the basic model established by this<br />
paper, the corresponding parameters to the minimum<br />
ETC are the optimal solution.<br />
To solve, the formula (3) is added with the slack<br />
variable Si2 (i= 1,2,…,n), then the following Lagrange<br />
function can be established as:<br />
n<br />
2 2 1 ⎡ A ⎤<br />
ETC(<br />
K,<br />
T,<br />
L,<br />
Z,<br />
Zv<br />
, λ,<br />
λv,<br />
S , Sv<br />
) = ⎢F<br />
+ ∑Fi<br />
+ + C(<br />
L)<br />
T<br />
⎥<br />
⎣ i=<br />
1 K ⎦<br />
⎧ n<br />
n<br />
n<br />
⎡ ⎛ ⎞ ⎛ ⎞⎤<br />
⎫<br />
⎪hv∑<br />
di<br />
⎢ ⎜ ∑di<br />
⎟ ⎜ 2∑di<br />
⎟⎥<br />
n ⎪<br />
⎪ i=<br />
1<br />
=<br />
=<br />
⎪<br />
+ ⎨ ⎢ ⎜ i 1<br />
− ⎟ + ⎜ i 1<br />
− ⎟ d<br />
⎥ ihi<br />
T K 1<br />
1 + ∑ ⎬<br />
⎪ 2 ⎢ ⎜ P ⎟ ⎜ P ⎟⎥<br />
i=<br />
1 2 ⎪<br />
⎪ ⎢ ⎜ ⎟ ⎜ ⎟⎥<br />
⎩ ⎣ ⎝ ⎠ ⎝ ⎠⎦<br />
⎪<br />
⎭<br />
v<br />
v<br />
KT<br />
n<br />
∑<br />
i=<br />
1<br />
2<br />
σ +<br />
di<br />
n<br />
∑<br />
h Z σ<br />
i i di<br />
i=<br />
1<br />
T + L<br />
n<br />
2<br />
2<br />
+ ∑λi<br />
[ αi<br />
−1+<br />
Φ(<br />
Zi<br />
) + Si<br />
] + λv<br />
[ αv<br />
−1+<br />
Φ(<br />
Zv<br />
) + Sv<br />
]<br />
i=<br />
1<br />
(4)<br />
Where, λ is the Lagrange multiplier, and λ= (λ1,<br />
λ2,…, λn) ≥ 0,λv ≥0 ,Z = ( Z1,Z2,…, Zn ),S = ( S12,<br />
S22,…, Sn2).<br />
Equation (4) is the single-stage supply chain<br />
inventory decision model with shortage constraints which<br />
the paper established.<br />
B. Solving process<br />
The solution process as follow:<br />
Take the formula (4) the second derivative for L, and<br />
get:<br />
n<br />
∑<br />
i=<br />
1<br />
( )<br />
( ) 3<br />
h<br />
2<br />
2<br />
iZ iσ<br />
di<br />
∂ ETC 1 ∂ C L<br />
= −<br />
2<br />
2<br />
∂L<br />
T ∂L<br />
4 T + L<br />
For the C(L) is a step function, to L in each L<br />
level range there are ∂2C(L)/∂L2 = 0, so to each L level<br />
range there are ∂2ETC/∂L2 < 0, this means that, with the<br />
specific K, T, the optimal solution to L is at the endpoint<br />
<strong>of</strong> the certain range. And for the slack variable Si2 is 0,<br />
the prerequisites <strong>of</strong> the minimum <strong>of</strong> the expected total<br />
cost per unit time ETC is that the first derivative is equal<br />
to zero, that is:<br />
∂ETC<br />
= α i −1<br />
+ Φ(<br />
Z i ) = 0<br />
∂λi<br />
(5)<br />
∂ETC<br />
= α v −1<br />
+ Φ(<br />
Z v ) = 0<br />
∂λv<br />
(6)<br />
∂ETC<br />
= hiσ<br />
d T + L − λiφ(<br />
Z i ) = 0<br />
i ∂Z<br />
i<br />
(7)<br />
n<br />
∂ETC<br />
2<br />
= hv<br />
KT∑<br />
σ d − λvφ(<br />
Z v ) = 0<br />
i<br />
∂Z<br />
v<br />
i=<br />
1<br />
(8)<br />
© 2011 ACADEMY PUBLISHER<br />
By the formula (5) can get:<br />
Zi* = Φ-1(1-αi), i = 1,2,…,n (9)<br />
By the formula (6) can get:<br />
Zv* = Φ-1(1-αv) (10)<br />
Where, Φ-1(x)is a standard normal inverse function.<br />
By the formula (7) can get :<br />
hiσ<br />
d T + L<br />
*<br />
i λi<br />
=<br />
> 0, i = 1,<br />
2,<br />
⋅⋅<br />
⋅,<br />
n<br />
φ(<br />
Zi<br />
)<br />
(11)<br />
*<br />
λ =<br />
n<br />
2<br />
hv KT∑<br />
σ di<br />
i=<br />
1<br />
> 0<br />
By the formula (8) can get:<br />
v<br />
φ(<br />
ZV<br />
) (12)<br />
In addition, take the formula (4) the second derivative<br />
respectively for Zi, Zv, and can get:<br />
2<br />
∂ ETC ⎛<br />
= −λ<br />
⎜<br />
2 i − Zi<br />
∂Z<br />
⎜<br />
i ⎝<br />
2<br />
Zi<br />
1 − ⎞<br />
2 e ⎟ = λiZ<br />
i<br />
2π<br />
⎟<br />
⎠<br />
2<br />
Zi<br />
1 −<br />
2 e > 0<br />
2π<br />
2<br />
∂ ETC ⎛<br />
= −λ<br />
⎜<br />
2 v − Zv<br />
∂Z<br />
⎜<br />
v ⎝<br />
2<br />
Zv<br />
1 − ⎞<br />
2 e ⎟ = λvZv<br />
2π<br />
⎟<br />
⎠<br />
2<br />
Zv<br />
1 −<br />
2 e > 0<br />
2π<br />
Put the Z*, Zv*, λ*, λv* which get from the formula<br />
(9), (10), (11), (12) into the formula (4) can get:<br />
ETC<br />
* * * *<br />
( K,<br />
T Z , Z , λ , λ )<br />
+ h Z<br />
*<br />
v v<br />
n<br />
*<br />
∑λi<br />
i=<br />
1<br />
v<br />
KT<br />
n<br />
∑<br />
i=<br />
1<br />
v<br />
1 ⎛<br />
= ⎜F<br />
+<br />
T ⎝<br />
2<br />
di<br />
+<br />
∑<br />
i=<br />
1<br />
n<br />
*<br />
∑hi<br />
Zi<br />
σ di<br />
i=<br />
1<br />
A ⎞<br />
Fi<br />
+ ⎟<br />
K ⎠<br />
⎧ n<br />
n<br />
n<br />
⎡ ⎛ ⎞ ⎛ ⎞⎤<br />
⎪hv<br />
∑d<br />
i ⎢ ⎜ ∑d<br />
i ⎟ ⎜ 2∑d<br />
i ⎟⎥<br />
⎪ i=<br />
1<br />
=<br />
=<br />
⎨ ⎢ ⎜ i 1 ⎟ + ⎜ i 1<br />
+ T K 1−<br />
−1⎟⎥<br />
+<br />
⎪ 2 ⎢ ⎜ P ⎟ ⎜ P ⎟⎥<br />
⎪ ⎢ ⎜ ⎟ ⎜ ⎟⎥<br />
⎩ ⎣ ⎝ ⎠ ⎝ ⎠⎦<br />
σ<br />
T + L<br />
* *<br />
*<br />
[ α −1+<br />
Φ(<br />
Z ) ] + λ [ α −1+<br />
Φ(<br />
Z ) ]<br />
n<br />
n<br />
∑<br />
i=<br />
1<br />
⎫<br />
⎪<br />
dihi<br />
⎪<br />
⎬<br />
2 ⎪<br />
⎪<br />
⎭<br />
+<br />
i<br />
i v v<br />
v<br />
(13)<br />
With the specific K, take the formula (13) the first<br />
derivative, set it as 0 and can get:<br />
n<br />
n<br />
n<br />
⎡ ⎛ ⎞ ⎛ ⎞⎤<br />
h<br />
2<br />
1<br />
⎢ ⎜ ⎟ ⎜ ⎟<br />
∂ ⎛ ⎞ ∑d<br />
∑d<br />
∑d<br />
n<br />
v i<br />
i<br />
i<br />
ETC<br />
A<br />
⎥<br />
i=<br />
1 ⎢ ⎜ i=<br />
1<br />
1<br />
1 ⎟+<br />
⎜ i=<br />
= − ⎜F<br />
+ + ⎟+<br />
−<br />
−1⎟⎥<br />
2 ∑Fi<br />
K<br />
∂T<br />
T ⎝ 1 ⎠ 2 ⎢ ⎜ ⎟ ⎜ ⎟<br />
i=<br />
K<br />
P P ⎥<br />
⎢ ⎜ ⎟ ⎜ ⎟⎥<br />
⎣ ⎝ ⎠ ⎝ ⎠⎦<br />
F +<br />
n<br />
∑<br />
T<br />
+<br />
F +<br />
i<br />
i=<br />
1<br />
2<br />
n<br />
∑<br />
i=<br />
1<br />
A<br />
K<br />
d Z ihi<br />
+<br />
2<br />
*<br />
v<br />
n<br />
∑<br />
n<br />
∑<br />
2<br />
*<br />
K σd<br />
hiZ<br />
iσ<br />
i<br />
di<br />
i=<br />
1 i=<br />
1 + = 0<br />
2 T 2 T + L<br />
n ⎡ ⎛<br />
hv∑<br />
di<br />
⎢ ⎜<br />
i=<br />
1<br />
= ⎢K⎜1<br />
−<br />
2 ⎢ ⎜<br />
⎢ ⎜<br />
⎣ ⎝<br />
n<br />
∑<br />
n<br />
∑<br />
i=<br />
1<br />
*<br />
d h Z i i v<br />
i=<br />
1<br />
+ +<br />
2<br />
2<br />
K σ di<br />
i=<br />
1<br />
2 T<br />
+<br />
By the formula (14) can get:<br />
n<br />
P<br />
∑<br />
d<br />
i<br />
⎞ ⎛<br />
⎟ ⎜ 2<br />
⎟ + ⎜<br />
⎟ ⎜<br />
⎟ ⎜<br />
⎠ ⎝<br />
n<br />
∑<br />
i=<br />
1<br />
n<br />
∑<br />
i=<br />
1<br />
P<br />
d<br />
*<br />
h Z σ<br />
i<br />
i<br />
i<br />
⎞⎤<br />
⎟⎥<br />
−1⎟⎥<br />
⎟⎥<br />
⎟⎥<br />
⎠⎦<br />
di<br />
2 T + L<br />
(14)
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1865<br />
n<br />
⎧ n<br />
⎛ A ⎞<br />
⎡ ⎛<br />
2⎜<br />
F + ∑ Fi<br />
+ ⎟ ⎪hv<br />
∑d<br />
i<br />
=<br />
⎪<br />
⎢ ⎜<br />
⎝ i 1 K ⎠ 2 i=<br />
1 = ⎨ ⎢K⎜1<br />
−<br />
3<br />
T T ⎪ 2 ⎢ ⎜<br />
⎪ ⎢ ⎜<br />
⎩ ⎣ ⎝<br />
+ Z<br />
*<br />
v<br />
K<br />
n<br />
∑<br />
i=<br />
1<br />
⎧ n ⎡ ⎛<br />
⎪hv<br />
∑d<br />
i<br />
⎪<br />
⎢ ⎜<br />
2 i=<br />
1 ≥ ⎨ ⎢K⎜1<br />
−<br />
T ⎪ 2 ⎢ ⎜<br />
⎪ ⎢ ⎜<br />
⎩ ⎣ ⎝<br />
n<br />
σ T<br />
2<br />
di<br />
∑<br />
i=<br />
1<br />
3<br />
−<br />
2<br />
n ⎞ ⎛ ⎞⎤<br />
di<br />
⎟ ⎜ 2∑di<br />
⎟⎥<br />
⎟ ⎜ i=<br />
1 + −1⎟⎥<br />
+<br />
P ⎟ ⎜ P ⎟⎥<br />
⎟ ⎜ ⎟<br />
⎠ ⎝ ⎠<br />
⎥<br />
⎦<br />
n<br />
∑<br />
i=<br />
1<br />
+<br />
n<br />
*<br />
∑hi<br />
Zi<br />
σ di<br />
i=<br />
1<br />
n ⎞ ⎛ ⎞⎤<br />
di<br />
⎟ ⎜ 2∑d<br />
i ⎟⎥<br />
⎟ ⎜ i=<br />
1 + −1⎟⎥<br />
+<br />
P ⎟ ⎜ P ⎟⎥<br />
⎟ ⎜ ⎟<br />
⎠ ⎝ ⎠<br />
⎥<br />
⎦<br />
n<br />
3<br />
−<br />
n<br />
2 2<br />
*<br />
d T + ∑hi<br />
Z<br />
i<br />
i σ di<br />
i=<br />
1<br />
− ( T + L)<br />
n<br />
∑<br />
i=<br />
1<br />
1<br />
2<br />
T<br />
⎫<br />
dihi<br />
⎪<br />
⎬<br />
2 ⎪<br />
⎪⎭<br />
n<br />
∑<br />
i=<br />
1<br />
−1<br />
⎫<br />
dihi<br />
⎪<br />
⎬<br />
2 ⎪<br />
⎪⎭<br />
3<br />
*<br />
−<br />
+ Zv<br />
K∑σ<br />
( T + L)<br />
2<br />
i=<br />
1<br />
(15)<br />
Take the formula (13) the second derivative for T and<br />
can get:<br />
n<br />
n<br />
n<br />
⎛ A ⎞ *<br />
2<br />
*<br />
2⎜<br />
F +<br />
2 ∑Fi<br />
+ ⎟ Zv<br />
K∑σ<br />
d T ∑Zi<br />
h<br />
i<br />
i di<br />
∂ ETC * * * *<br />
= 1<br />
= 1<br />
= 1<br />
, , , =<br />
⎝ i K ⎠<br />
i<br />
i<br />
Z<br />
−<br />
−<br />
2 i Zv<br />
λ λv<br />
3<br />
∂T<br />
T<br />
4<br />
4<br />
3<br />
2<br />
σ ( T + L)<br />
Put the formula (16) into (15) then can get:<br />
⎧ n<br />
n<br />
n<br />
⎡ ⎛ ⎞ ⎛ ⎞⎤<br />
2<br />
⎪hv<br />
2<br />
* * * * 2 ⎪<br />
⎢ ⎜ ⎟ ⎜ ⎟<br />
∂<br />
∑di<br />
∑di<br />
∑di<br />
ETC<br />
⎥<br />
i=<br />
1<br />
1<br />
1<br />
, , ,<br />
1<br />
1<br />
2<br />
⎨ ⎢ ⎜ i=<br />
⎟+<br />
⎜ i=<br />
Z =<br />
−<br />
− ⎟⎥<br />
i Zv<br />
λ λv<br />
K<br />
+<br />
∂T<br />
T ⎪ 2 ⎢ ⎜ P ⎟ ⎜ P ⎟⎥<br />
⎪ ⎢ ⎜ ⎟ ⎜ ⎟⎥<br />
⎩ ⎣ ⎝ ⎠ ⎝ ⎠⎦<br />
3<br />
2<br />
(16)<br />
n 3<br />
−<br />
n 3<br />
3<br />
*<br />
2 2<br />
*<br />
−<br />
+ Z<br />
+<br />
( + ) 2<br />
v K∑σ<br />
d T ∑hi<br />
Ziσ<br />
d T L > 0<br />
i<br />
i<br />
4 i=<br />
1 4 i=<br />
1<br />
This means that the solution T which get form the<br />
formula (14) must be a local minimum, furthermore,<br />
since the values <strong>of</strong> this second derivative is always<br />
positive, means for equation (13), with the particular K,<br />
the optimal solution to T get by formula (14) was the only<br />
solution.<br />
3<br />
n<br />
∑<br />
i=<br />
1<br />
⎫<br />
dihi<br />
⎪<br />
⎬<br />
2 ⎪<br />
⎪⎭<br />
In summary, the following algorithm can be taken to<br />
calculate the optimal solution to K, T, L, Z, Zv, λ, λv <strong>of</strong><br />
the model:<br />
Step 1, set K=1, L=Li, i=0, 1,…,b.<br />
Step 2, get the solution to Z*, ZV* by the (9) and (10).<br />
Step 3, get the T*(K, Li) by the (14).<br />
Step 4, put the Z*, ZV*, T*(K, Li), Li into equation<br />
(11) and (12) and get λ*(K, Li), λv*(K, Li).<br />
Step 5, put the Z*, ZV*, T*(K, Li), Li, λ*(K, Li),<br />
λv*(K, Li) into the formula (13) to get<br />
( )<br />
*<br />
*<br />
*<br />
*<br />
*<br />
ETCi ( K,<br />
Li<br />
) = ETCT<br />
( K,<br />
Li<br />
) , Z ( K,<br />
Li<br />
) , Zv<br />
( K,<br />
Li<br />
) , λ ( K,<br />
Li<br />
) , λv<br />
( K,<br />
Li<br />
) , Li<br />
, K<br />
Step 6, set ETCi(K, Li) = Min[ETCi(K, Li)], i=0,1,…<br />
,b.<br />
Step 7, if K=1, set ETCs = ETC(K, L), K=K+1, go<br />
back to step 3, then if ETC(K, L)< ETCs, ETCs =<br />
ETC(K, L), K=K+1, go back to step 3, otherwise, set<br />
K*=K-1, L*=L, T*=T*(K*, L), λ*=λ*(K*, L),<br />
λv*=λv*(K*, L), ETC*= ETC s, and the solution is over.<br />
IV. MODEL APPLICATION EXAMPLE<br />
Set the C(L) as:<br />
⎧0<br />
L = 0.<br />
02<br />
⎪<br />
5 0.<br />
01 ≤ L < 0.<br />
02<br />
C(<br />
L)<br />
= ⎨<br />
⎪11<br />
0.<br />
005 ≤ L < 0.<br />
01<br />
⎪<br />
⎩18<br />
0.002 ≤ L < 0.<br />
005<br />
Other parameters shown in Table 1 and Table 2 as<br />
below:<br />
Table 1 The buyer’s parameters<br />
d = 6000 unit/year d = 5000 unit/year d = 10000 unit/year<br />
1<br />
2<br />
σ = 600 unit/year σ = 800 unit/year σ = 900 unit/year<br />
d1<br />
d<br />
2<br />
F1 = 100 USD/times F2 = 150 USD/times F3 = 80 USD/times<br />
h1 = 5 USD/unit/year h2 = 4 USD/unit/year h3 = 4.5 USD/unit/year<br />
α1 = 0.01 α2 = 0.01 α3 = 0.01<br />
F = 100 USD/times<br />
Table 2 The supplier’s parameters<br />
P = 28000 unit/year A = 200 USD/patch hv = 3 USD/unit/year αv = 0.01<br />
The results by calculation shown in Table 3 and Table 4 as below:<br />
K<br />
L * (K)<br />
T * (K)<br />
Z1 *<br />
Z2 *<br />
Z3 *<br />
Zv *<br />
λ1 * (K)<br />
λ2 * (K)<br />
λ3 * (K)<br />
λv * (K)<br />
ETC * (K)<br />
© 2011 ACADEMY PUBLISHER<br />
Table 3 The solution to the parameters<br />
1 2 3 4 5<br />
0.002 0.002 0.002 0.002 0.002<br />
0.07338 0.06390 0.05934 0.05626 0.05387<br />
2.326 2.326 2.326 2.326 2.326<br />
2.326 2.326 2.326 2.326 2.326<br />
2.326 2.326 2.326 2.326 2.326<br />
2.326 2.326 2.326 2.326 2.326<br />
30905.4 28896.3 27879.1 27169 26606.9<br />
32965.8 30822.7 29797.7 28980.3 28380.7<br />
41722.3 39010 37636.8 36678.2 35919.3<br />
41023.7 54138.3 63897.1 71838.6 78596.4<br />
23140.2 23101.9 23695.7 24392 25096.5<br />
3<br />
d<br />
3
1866 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
Table 4 The optimal solutions to the model<br />
K<br />
* L * T * Buyer 1<br />
R1 S1<br />
Buyer 2<br />
R2 S2<br />
Buyer 3<br />
R3 S3<br />
supplier<br />
Rv Sv<br />
ETC<br />
2 0.002 0.0639 754 1 808 1 1197 1 3803 2 23101.9<br />
Note: S1, S2, S3 for the buyer’s expected shortages in the replenishment cycle, Sv for the supplier’s expected shortages in the<br />
delivery period.<br />
From the table 3 we can get the conclusion that the [8] Reve, T., and Johansen, E. Organizational Buying in the<br />
total cost curve shows the convex trends: The optimal Offshore Oil Industry. Industry Marketing Management,<br />
shipping times per production cycle for the manufacturer<br />
is 2 times. However, compared to the case <strong>of</strong> not<br />
considering the shortening delivery time (L*=L0), the<br />
best common replenishment cycle <strong>of</strong> the buyer T* is<br />
reduced by the 0.06414 years to 0.06390 years, the best<br />
replenishment lead time <strong>of</strong> the buyer L* is reduced to<br />
1982,(11):275-282.<br />
[9] Silver, E. and Peterson, P., Decision System for Inventory<br />
Management and Production Planning, 2nd ed., Wiley,<br />
New York, 1985.<br />
[10] Miller, D. and Dorge, C. Psychological and Traditional<br />
Determinants <strong>of</strong> Structure Administrative Science<br />
Quarterly, 1986, 31: 539-560.<br />
0.002 years. In addition, the table 4 shows that, all the [11] Malone, W., Yates, J. and Benjamin, I. Electronic Markets<br />
safety factors Z1*, Z2*, Z3* <strong>of</strong> the lead time <strong>of</strong> the buyer and Electronic Hierarchies. Communications <strong>of</strong> the ACM,<br />
1, 2, 3, are 2.326, 99% service level, that is, the buyers’ 1987,30(6):484-497.<br />
replenishment target level per cycle is 753, 808, 1197<br />
units, the expected shortages per cycle is 1 unit. The<br />
safety factor Zv* <strong>of</strong> the supplier’s lead time is 2.326 too,<br />
that is, the supplier’s replenishment target level per cycle<br />
is 3803 units, the expected shortages per cycle is 2 units.<br />
At the time, the optimal total cost <strong>of</strong> the supply chain per<br />
[12] Novack, A., and Simco, W. The Industrial Procurement<br />
Process: A Supply Chain Perspective. <strong>Journal</strong> <strong>of</strong> Business<br />
Logistics, 1991,12,(1):145-167.<br />
[13] Liao, C.J., and Shyu, C.H., An analytical determine <strong>of</strong> lead<br />
time with normal demand, International <strong>Journal</strong> <strong>of</strong><br />
Operations and Production Management, 1991, l7, 4:115-<br />
124.<br />
unit time is $23101.9, go down 2.2% compared to case <strong>of</strong> [14] Ballou, R. H., Business Logistics Management, 3rd ed.,<br />
the supplier not allow to shorten the delivery.<br />
Prentice-Hall, Englewood Cliffs, NJ, 1992.<br />
[15] Davis, T. Effective Supply Chain Management. Sloan<br />
V. CONCLUSION<br />
Man-agement Review (Summer), 1993: 35-46.<br />
[16] Lee, H. L. and C. Billington, Material management in<br />
To the supply chain management, the lead time has a decentralized supply chains. Operations<br />
great impact on inventory management performance. Research,1993,41(5): 835-847.<br />
However, the majority <strong>of</strong> supply chain inventory decision<br />
model regarded the lead time as being fixed, this does not<br />
correspond with the practice. Innovations <strong>of</strong> this paper is<br />
that, set the lead time as a variable, with the constraints <strong>of</strong><br />
meeting different shortages, explores the inventory<br />
decision model <strong>of</strong> a supply chain which composed <strong>of</strong> a<br />
single supplier and multiple buyers. While the earlier<br />
[17] Gerwin, D. Manufacturing Flexibility: A Strategic<br />
Perspective. Management Science, 1993, 39(4): 395-410.<br />
[18] Li, K., Shyu, T. and Adiga, S. A Heuristic Rescheduling<br />
Algorithm for Computer-Based Production Scheduling<br />
Systems. International <strong>Journal</strong> <strong>of</strong> Production Research,<br />
1993, 31(8): 815- 1826.<br />
[19] Sethi, V. and Carraher, S.M., Developing measures for<br />
assessing the organizational impact <strong>of</strong> information<br />
calculations appear more complicated, if programmed, technology: a comment on Mahmood and Soon’s paper.<br />
the application will be very simple. The inadequacy <strong>of</strong> Decision Sciences, 1993, 24: 867-77.<br />
the study is that the model in this article are not compared<br />
with other models and sentenced to the merits, this is also<br />
the further research content for the author.<br />
[20] Banerjee, A. and Banerjee, S., A coordinated order-up-to<br />
inventory control policy for a single supplier and multiple<br />
buyers using electronic data interchange, International<br />
<strong>Journal</strong> <strong>of</strong> Production Economics, 1994, 35, 1-3: 85-91.<br />
REFERENCES<br />
[21] Ben-Daya, M, and Raouf, A., Inventory models involving<br />
lead time as a decision variable, The <strong>Journal</strong> <strong>of</strong> the<br />
[1] Cronbach, L. J., & Meehl, P. E..Construct validity in<br />
psychological tests. Psychological Bulletin,<br />
1955,52:281302.<br />
[2] Duncan, B. Characteristics <strong>of</strong> Organizational<br />
Environmental and Perceived Environmental Uncertainty.<br />
Administrative Science Quarterly,1972, 17: 313-327.<br />
[3] Lawshe, H. A Quantitative Approach to Contect Validity.<br />
Personnel Psychology, 1975, 28: 563-575.<br />
[4] Lawshe, H. A Quantitative Approach to Contect Validity.<br />
Personnel Psychology, 1975, 28: 563-575.<br />
[5] Whybark, C., and William, J. Material Requirement<br />
Planning under Uncertainty. Decision Science, 1976,17(4):<br />
595-606.<br />
[6] Mintzberg, T.Patterns in Strategy Formation. Management<br />
Science,1978,(24):934-948.<br />
[7] Churchill, A. A Paradigm for Developing Better Measures<br />
<strong>of</strong> Marketing Constructs. <strong>Journal</strong> <strong>of</strong> Marketing Research,<br />
1979,16(1):64-73.<br />
Operational Research Society, 1994,45(5): 579-582.<br />
[22] Tersine, R. J., Principles <strong>of</strong> Inventory and Materials<br />
Management, 4th ed., 1994.<br />
[23] Kotteaku, G., Laios, G. and Moschuris, J. The influence <strong>of</strong><br />
product complexity on the purchasing structure.<br />
International <strong>Journal</strong> <strong>of</strong> Management Science, 1995,<br />
23(1):27-39.<br />
[24] Krajewski, L.J. and Ritzman, L.P., Operations<br />
Management: Strategy and Analysis. Addison Wesley,<br />
Reading, MA, 1996.<br />
[25] Ouyang, LY., Yeh, NC., and Wu, KS., Mixture inventory<br />
model with backorders and lost sales f or variable lead<br />
time, <strong>Journal</strong> <strong>of</strong> Operational Research Society, 1996,47:<br />
829-832.<br />
[26] Washington, D.C., Statistical Abstract <strong>of</strong> the United States<br />
U.S. Bureau <strong>of</strong> the Census: 1996, 116th ed., 443, 1996.<br />
[27] Fisher, M., What is the Right Supply Chain for Your<br />
Product . Harvard Business Review, 1997, 75(2): 105-116.<br />
© 2011 ACADEMY PUBLISHER
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1867<br />
[28] Calinescu, A., Efstathiou, J., Schim J., and Bermejo, J.<br />
Applying and Assessing Two Methods for Measuring<br />
Complexity in Manufacturing. The <strong>Journal</strong> <strong>of</strong> the<br />
Operational Research Society, 1998, 49: 723-733.<br />
[29] Hair, F., Anderson, E., Tatham, L., and Black, C.<br />
Multivariate Data Analysis. Prentice Hall, New Jersey,<br />
1998.<br />
[30] Shawnee, V., C. Roger and D. Cornelia. Supply chain<br />
flexibility: an empirical study. <strong>Journal</strong> <strong>of</strong> Supply Chain<br />
Management,1999,35(3):16-24.<br />
[31] Vickery, V., Roger, C. and Cornelia, D. Supply Chain<br />
Flexibility: An Empirical Study. <strong>Journal</strong> <strong>of</strong> Supply Chain<br />
Management,1999, 35(3):16-24.<br />
[32] Khurana, A. Managing Complex Production Processes.<br />
Sloan Management Review, 1999, 40(2): 85-97.<br />
[33] Zhu, L. and Soh, C. FMS Job-Shop scheduling under<br />
disruptions with consideration <strong>of</strong> time and sequence<br />
deviation. IEEE International Symposium on Intelligent<br />
Control/Intelligent Systems and Semiotics, 1999: 138-143.<br />
[34] Ouyang, LY. and Chuang, BR., Stochastic Inventory<br />
Model Involving Variable Lead Time with a Service Level,<br />
Yugoslav <strong>Journal</strong> <strong>of</strong> Operations Research, 2000, 10, 1: 81-<br />
98.<br />
[35] Chopra, S. and Meindl, P. Supply Chain Management:<br />
Strategy, Planning, and Operation. Prentice-Hall, New<br />
Jersey, 2001.<br />
[36] Subramaniam, C., and Shaw, M. A Study <strong>of</strong> the Value and<br />
the Impact <strong>of</strong> B2B E-Commerce: The Case <strong>of</strong> Web-Based<br />
Procurement. International <strong>Journal</strong> <strong>of</strong> Electronic<br />
Commerce, 2002, 6(4): 19-40.<br />
[37] Gatignon, H., Tushman, L., Smith, W., and Anderson, P.A<br />
Structural Approach to Assessing Innovation: Construct<br />
© 2011 ACADEMY PUBLISHER<br />
Development <strong>of</strong> Innovation Locus, Type, and<br />
Characteristics. Management Science,2002,48(9):1103-<br />
1122.<br />
[38] Taha, H. A., Operations Research, 2003.<br />
[39] Yang, B. Y., Watkins, K. E., & Marsick, V. J.The<br />
construct <strong>of</strong> the learning organization: Dimensions,<br />
measurement, and validation. Human Resource<br />
Development Quarterly, 2004, (15): 31-55.<br />
[40] Ben-Daya, M. and Hariga, M., Integrated single vendor<br />
single buyer model with stochastic demand and variable<br />
lead time, International <strong>Journal</strong> <strong>of</strong> Production Economics,<br />
2004, 92, 1:75-80.<br />
[41] Yao, Yuliang ;Dong, Yan ;Dresner, Martin, Managing<br />
supply chain backorders under vendor managed inventory:<br />
An incentive approach and empirical analysis, European<br />
<strong>Journal</strong> <strong>of</strong> Operational Research, 2010, 6,31(2):350-359.<br />
[42] Ahmadi Javid, Amir ;Azad, Nader, Incorporating location,<br />
routing and inventory decisions in supply chain network<br />
design, Transportation Research - Part E the Logistics and<br />
Transportation Review,2010,46,5:582-597.<br />
[43] Kogan, K ;Lou, S ;Tapiero, C S ;Shnaiderman, M, Supply<br />
Chain With Inventory Review and Dependent Demand<br />
Distributions: Dynamic Inventory Outsourcing, IEEE<br />
Transactions on Automation Science and Engineering,<br />
2010,7,2:197-207.<br />
[44] Battini, D. ;Gunasekaran, A. ;Faccio, M. ;Persona, A.<br />
;Sgarbossa, F., Consignment stock inventory model in an<br />
integrated supply chain, International <strong>Journal</strong> <strong>of</strong><br />
Production Research, 2010, 119, 2: 477-500.<br />
[45] Harish Krishnan;Ralph A. Winter, Inventory Dynamics<br />
and Supply Chain Coordination, Management<br />
Science,2010,56,1:141.
1868 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
Variable Selection for Credit Risk Model Using<br />
Data Mining Technique<br />
Kuangnan Fang<br />
Department <strong>of</strong> Planning and statistics/Xiamen University, Xiamen, China<br />
Email: ruiqwy@163.com<br />
Hong Huang *<br />
Economics Department/Hefei Normal University, Hefei, China<br />
Email: HH6@263.net<br />
Abstract—With the emergence <strong>of</strong> the current financial crisis,<br />
societies see the increasing importance <strong>of</strong> credit risks<br />
management in financial institutions. Four mainstream<br />
credit risk rating models have been developed, however,<br />
their applicability in the Taiwan market is yet to be<br />
evaluated. In this paper, six major credit risk models,<br />
including Merton Option Pricing Model,Discriminant<br />
Analysis Model, Logistic Regression (Logit) Model, Probit<br />
Model, Survival Analysis Model, and Artificial Neural<br />
Network Model were examined, in order to identify the<br />
common variables applicable to each model. The common<br />
variables were then applied to each respective model<br />
directly. Using Transition Matrix and mapping methods to<br />
estimate long term default probability, for developing<br />
appropriate credit risk model with the estimated default<br />
probability.<br />
Index Terms—Credit Default Risk; Logit; Logistic<br />
Regression Model<br />
I. INTRODUCTION<br />
In recent years, with the development <strong>of</strong> global credit<br />
portfolio management, continuous innovations in<br />
financial credit derivatives and financial statistical<br />
techniques, the growth on awareness <strong>of</strong> credit risks<br />
among financial institutions and regulatory authorities,<br />
both practical and theoretical research and development<br />
<strong>of</strong> credit risk evaluation models are given high<br />
importance and under vigorous progress. Seeing the<br />
vitality <strong>of</strong> considering credit risks in financial institutions,<br />
The New Basel Capital Accord focuses on strengthening<br />
the risk management mechanism <strong>of</strong> banks by requiring<br />
banks to establish a sound internal risk assessment<br />
mechanism and to increase the responsibility <strong>of</strong> the<br />
exte ∗ rnal supervisory bodies. The new accord encourages<br />
financial institutions to establish their own credit rating<br />
mechanisms; however, it has allowed flexibility in<br />
choosing which credit risk model to use. At present, there<br />
exists several developed credit risk models; each has its<br />
own theoretical basis and advantages. Further discussion<br />
is required to investigate whether a particular model is<br />
applicable to the Taiwan market, or, in other words,<br />
∗ Corresponding author <strong>of</strong> this article.<br />
© 2011 ACADEMY PUBLISHER<br />
doi:10.4304/jcp.6.9.1868-1874<br />
whether it is applicable globally or it should be adjusted<br />
according to local factors.<br />
With the flexibility towards credit risks allowed in the<br />
new accord, in this paper, we shall analyze the six major<br />
credit risk models, including Merton Option Pricing<br />
Model, Discriminant Analysis Model, Logit Model,<br />
Probit Model, Survival Analysis Model and Artificial<br />
Neural Network Model, to identify common variables<br />
applicable to each model based on the financial<br />
statements <strong>of</strong> companies in Taiwan and market data. The<br />
common variables can then be applied to each respective<br />
model directly, in order to establish an appropriate credit<br />
risk model with the estimated default probability.<br />
II. MAJOR CREDIT RISK MODELS<br />
A. Credit Metrics Model<br />
Credit Metrics Model was developed by J.P. Morgan<br />
in 1997. It mainly uses the technique <strong>of</strong> migration<br />
analysis and Value-at-Risk to look at the credit risks<br />
arising from credit ratings changes <strong>of</strong> credit assets in the<br />
investment portfolio.<br />
Credit Metrics Model mainly depends on historical<br />
average default rates and the credit rating transition<br />
matrix. First, it estimates the probability <strong>of</strong> transitions<br />
between risk groups based on historical data, and at the<br />
same time establishes the correlation between credit<br />
ratings and the value <strong>of</strong> a debtor company's asset, so as to<br />
determine the joint migration behavior <strong>of</strong> credit qualities<br />
among the asset portfolios. Then, portfolio default loss<br />
distribution can be generated by looking at the market<br />
value changes <strong>of</strong> asset portfolio in the Monte Carlo<br />
simulation <strong>of</strong> quality transitions. Eventually, the value <strong>of</strong><br />
a single loan or loan portfolio can be calculated. The<br />
model has high applicability as it can be applied to a wide<br />
variety <strong>of</strong> financial products, such as bonds, loans, loan<br />
commitments, accounts receivable, letters <strong>of</strong> credit, as<br />
well as financial derivatives. However, it emphasizes the<br />
assumption that all counterparties within the same risk<br />
group have the same degree <strong>of</strong> credit risk. In addition, in<br />
determining the credit transition matrix probability, the<br />
model does not adjust properly according to the<br />
prevailing economic conditions. Therefore, there are
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1869<br />
<strong>of</strong>ten gaps between estimation results and empirical<br />
results.<br />
B. KMV Model<br />
The KMV model is proposed by KMV Corporation<br />
based on the Merton Model. It defines the "distance to<br />
default" which indicates the distance between a<br />
company's asset value and the default point. The greater<br />
the distance, the smaller the default probability will be.<br />
On the other hand, the smaller the distance, the greater<br />
the default probability <strong>of</strong> the company's assets will be. In<br />
other words, default will occur when the company's asset<br />
value is lower than the default point. However, different<br />
from Merton model, KMV discovers that the company<br />
has refinancing abilities in real practices; therefore<br />
default may not necessarily occur when asset value is<br />
lower than the book value <strong>of</strong> liabilities. According to<br />
KMV, the real default point is usually somewhere<br />
between the value <strong>of</strong> total liabilities and the value <strong>of</strong><br />
current liabilities. For normalization, the distance-todefault<br />
is indicated as the number <strong>of</strong> standard deviations<br />
between the company's asset value and the default point.<br />
Then, by mapping the distance-to-default to the Expected<br />
Default Frequency (EDF), the EDF can be calculated.<br />
KMV Corporation has accumulated a large database<br />
which is used to estimate correlations between default<br />
probabilities and corporate defaults. Based on these<br />
correlations, credit ratings transition matrix and default<br />
loss distribution <strong>of</strong> the debtor can then be further derived.<br />
Instead <strong>of</strong> relying on the credit ratings transition matrix,<br />
the KMV approach tracks the market conditions and<br />
incorporates the company's financial data and market data<br />
in the model to accurately grasp the credit risk changes <strong>of</strong><br />
the asset components. In addition, the accuracy <strong>of</strong> the<br />
prediction from the model is enhanced by its ability <strong>of</strong><br />
directly calculating the EDF <strong>of</strong> the company. However,<br />
the model assumes the company's asset value changes<br />
follow the normal distribution and does not consider the<br />
volatility <strong>of</strong> liabilities.<br />
C. Credit Risk+ Model<br />
Credit Risk+ is a default model proposed by Credit<br />
Suisse Financial Products (CSFP) in 1996. It is mainly<br />
based on an actuarial approach to derive the loss<br />
distribution <strong>of</strong> bonds or loans portfolio, and calculate the<br />
credit loss provision. The basic hypothesis is that default<br />
loss occurs when many debtors default, and each debtor's<br />
default probability is the same and very small. Therefore,<br />
the number <strong>of</strong> defaults in the asset portfolio can be<br />
estimated in accordance with the Poisson distribution,<br />
while the default probabilities depend upon a gammadistributed<br />
set <strong>of</strong> risk factors and will change over time.<br />
The model is based on a basic assumption that the<br />
number <strong>of</strong> defaults in the portfolio follows a Poisson<br />
distribution, and uses the volatility <strong>of</strong> default probabilities<br />
to reflect the influences <strong>of</strong> default correlation. Through<br />
statistical analysis <strong>of</strong> default rates and recovery rates <strong>of</strong><br />
defaulted loans, loans <strong>of</strong> common default loss<br />
characteristics are put under same groups to derive the<br />
probability function <strong>of</strong> loss distribution. Then the future<br />
loss distribution <strong>of</strong> the portfolio will be estimated and<br />
© 2011 ACADEMY PUBLISHER<br />
eventually, the expected and non-expected losses <strong>of</strong> the<br />
portfolio can be obtained. The model makes no<br />
assumption for the reason <strong>of</strong> default risks and requires<br />
small amount <strong>of</strong> data. It has also taken into consideration<br />
<strong>of</strong> volatility <strong>of</strong> default probabilities in the process <strong>of</strong><br />
calculation. However, the model assumes credit<br />
exposures are fixed and regarded as a constant. The<br />
model also does not take into account the risks <strong>of</strong> rating<br />
changes.<br />
D. Credit Portfolio View Model<br />
The basic theories <strong>of</strong> Credit Portfolio View were<br />
published by McKinsey & Company in 1997. The main<br />
characteristics <strong>of</strong> the model are that it assumes the<br />
probabilities <strong>of</strong> default occurrence and credit quality<br />
changes are closely related to the overall economic<br />
conditions. In general, many credit risk models assume<br />
that default occurrence is a result <strong>of</strong> individual financial<br />
health <strong>of</strong> the specific company. However, empirical<br />
findings show that the probabilities <strong>of</strong> default and rating<br />
migration <strong>of</strong> a company fluctuate with the business cycle.<br />
When economic conditions worsen, the default<br />
probability <strong>of</strong> a company default increases accordingly,<br />
and vice versa. In other words, credit cycles and<br />
economic cycles are closely correlated. The model<br />
mainly uses the following process to assess the credit risk<br />
<strong>of</strong> a company: set up a multi-factor model which measure<br />
systematic risks to determine the economic conditions;<br />
then evaluate the default probability <strong>of</strong> a company with<br />
the Logit Model. By modeling the relationship between<br />
credit ratings transition matrix and macroeconomic<br />
factors such as economic growth rate, default loss<br />
distribution is derived. The model assumes that default<br />
probabilities are related to the overall economic<br />
conditions, which is in line with the reality. In the<br />
process <strong>of</strong> calculation <strong>of</strong> credit risk, it uses the actual<br />
discrete distribution <strong>of</strong> the portfolio, which is more<br />
accurate than using continuous distribution, and is able to<br />
assess the credit risks <strong>of</strong> liquid and non-liquid assets at<br />
the same time. However, the selection <strong>of</strong> economicfinancial<br />
factors may be subjectively influenced, and<br />
important economic factors could be missed out in the<br />
evaluation process, resulting in overestimation or<br />
underestimation.<br />
III. RESEARCH METHODOLOGY<br />
For the purpose <strong>of</strong> accuracy and applicability, we first<br />
use six models, namely Merton Option Pricing Model,<br />
Discriminant Analysis Model, Regression Analysis<br />
Model, Logit Model and Probit Model, Survival Analysis<br />
Model, Artificial Neural Network Model, to establish a<br />
credit risk scoring model with the best variables set and<br />
common variables set. Among the six models, only<br />
Merton Option Pricing Model uses the market approach,<br />
while the other five models uses the actual approach. The<br />
common characteristic <strong>of</strong> actual approach models is that<br />
they require historical financial data for modeling. The<br />
selection <strong>of</strong> variables to be used in the model is another<br />
concern. We shall first select the variables that can be<br />
input into the model, then, among these selected ones,
1870 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
choose the best variables set using statistical methods,<br />
and apply the common variables to each model for<br />
comparing their results <strong>of</strong> differences.<br />
Having derived results from the above evaluation<br />
model, in order to find out a reasonable default<br />
probability, a bank will usually use a quantitative<br />
approach to modeling. When using a quantitative<br />
approach for modeling, attention should first be made to<br />
whether the selected variables are suitable for estimating<br />
default probabilities. The bank must prove that the<br />
selected estimation variables have significant correlation<br />
with default probabilities. It should adopt a statistical<br />
method to prove if the selected variables have significant<br />
explanatory power <strong>of</strong> default probabilities. To this end,<br />
the most common statistical method is to build a scoring<br />
system based on the regression approach. After the<br />
scoring system is established, the bank must rank and<br />
grade the rating <strong>of</strong> each exposure <strong>of</strong> its investment or<br />
loan portfolio. According to the New Accord, there<br />
should be at least seven grades <strong>of</strong> rating so as to prevent<br />
over-concentration <strong>of</strong> risks. In this paper, we first<br />
establish the required scoring model, then quantified the<br />
ratings by mapping method to derive default<br />
probabilities; the results are validated with benchmarking.<br />
IV. EMPIRICAL ANALYSIS<br />
A. Sampling<br />
In this study, sample selection criterion is that the<br />
company has to be publicly listed as <strong>of</strong> December 2010.<br />
Accordingly, credit clients' data between January 2001<br />
and December 2010 have been collected from banks in<br />
Taiwan as samples. The financial information used in this<br />
study are mainly combined statements, supplemented by<br />
individual statements. 10,032 observations have been<br />
collected, excluding data with omission,which include<br />
285 default cases and 9747 normal companies. In order to<br />
apply to the model, we classify the samples into training<br />
samples and valid samples. The sample distribution is<br />
summarized in table 1.<br />
Table 1: Sample distribution<br />
training<br />
samples<br />
Valid<br />
samples<br />
Total<br />
number<br />
normal<br />
companies<br />
5,604 4,143 9,747<br />
default cases 153 132 285<br />
Total number 5,757 4,275 10,032<br />
B. Selection <strong>of</strong> Variables<br />
(a).Selecting Common Variables<br />
There are more than a hundred variables generated<br />
from a company's financial statement analysis; however,<br />
it is doubtful if each variable can be used to explain the<br />
default occurrence <strong>of</strong> the company. Therefore, we will<br />
first make reference to the variables selection <strong>of</strong> famous<br />
research institutions in Taiwan and around the world, as<br />
well as those adopted by representative papers.<br />
The industry characteristics and sampling quantities<br />
adopted by the Taiwan Corporate Credit Risk Index<br />
(TCRI), which evaluates public companies (non-<br />
© 2011 ACADEMY PUBLISHER<br />
financial), conform to the research requirements <strong>of</strong> this<br />
paper. Therefore, we have made reference to the<br />
variables used in the TCRI rating. According to the<br />
TCRI, a good company should be pr<strong>of</strong>itable, with asset<br />
management efficiency, sound financial planning and a<br />
market leader. Accordingly, we use the following four<br />
dimensions <strong>of</strong> financial indicators: pr<strong>of</strong>itability,<br />
efficiency, security and size.<br />
The Falkenstein (2000) model uses variables from six<br />
dimensions (pr<strong>of</strong>itability, security, size, liquidity,<br />
efficiency and growth ability) and compares the<br />
correlation between financial ratios and default<br />
probability under each dimension to choose the most<br />
suitable variables to be used in the model. Finally, 10<br />
financial ratios are chosen to build the statistical model.<br />
In 1968, Edward I Altman used a number <strong>of</strong> variables<br />
to conduct estimation for company failures. There were<br />
22 financial ratios used for validation, including liquidity,<br />
pr<strong>of</strong>itability, financial leverage, repayment capability and<br />
efficiency. Eventually, the five ratios with best predictive<br />
power were selected for the statistical modeling.<br />
According to the empirical experience <strong>of</strong> TCRI in<br />
credit rating, the credit risks <strong>of</strong> a company are not<br />
completely reflected in the financial ratios and many risks<br />
are actually reflected in non-financial data. Therefore, in<br />
this paper we will also consider the following variables:<br />
opinions <strong>of</strong> accountants, related-parties purchase-sales<br />
ratio, directors' pledge ratio, P/E ratio, P/B ratio and<br />
compound Return on Equity.<br />
Due to various reasons, a company may adopt financial<br />
ratios to make its books look better. If this is the case, we<br />
may not find out the real situation about the company by<br />
judging the financial ratios only. For example, a<br />
company may borrow in the name <strong>of</strong> its subsidiary by<br />
endorsing the loan. In light <strong>of</strong> the consideration that<br />
financial ratios may not reflect the real stories, in this<br />
research, we also calculate the "adjusted" financial ratios,<br />
including: recurring net pr<strong>of</strong>it, debt-to-equity ratio, longterm<br />
pr<strong>of</strong>itability indicators.<br />
(b). Selecting best variables for each model<br />
Even after the above variables screening, we still come<br />
up with a large number <strong>of</strong> variables. Each <strong>of</strong> these<br />
variables may not necessarily has explanatory power<br />
about our sample companies; besides, if there are too<br />
many variables included in the model, the model will<br />
become too complicated, and collinearity problem among<br />
variables may arise, leading to unreasonable estimation <strong>of</strong><br />
parameters. In addition, the variables that can explain<br />
default probabilities may vary among different models.<br />
Therefore, in the following, we will evaluate the variables<br />
suitable for each model so that models with the best<br />
statistical explanatory power can be built.<br />
Regarding Logit Model, Probit Model and Survival<br />
Analysis, in the process <strong>of</strong> variables selection, first we<br />
put the independent variables into two groups, depending<br />
on whether the company is defaulting or not; then we use<br />
the SPSS (s<strong>of</strong>tware for quantitative data analysis) to carry<br />
out two-tailed T-tests for independent samples. At the<br />
confidence coefficient <strong>of</strong> 0.05, mean differences are<br />
tested and variables with significant differences (P-
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1871<br />
Value
1872 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
Table 3: Table <strong>of</strong> Common Variables<br />
Dimension Selected Variance<br />
ROA (Before Interest<br />
Pr<strong>of</strong>itability<br />
and Tax)<br />
Compound ROA<br />
Repayment Capability Quick Ratio<br />
Activity Total Asset Turnover<br />
Growth Revenue Growth<br />
Financial Structure Debt / Equity<br />
A good independent variable shall have significant<br />
explanatory power on the dependent variables. We want<br />
to conduct a statistical analysis to validate if the selected<br />
variables have good explanatory power over the default<br />
probability. The most common method used for such<br />
purpose is regression; therefore, we will use regression to<br />
Selected Variables<br />
validate the variables. As to the estimation <strong>of</strong> default<br />
probabilities, we have made reference to the research by<br />
Xue Ren-rui, Liu Ying-feng.After we have obtained the<br />
default probabilities, we have conducted simple<br />
regression analysis to the variables using the weighted<br />
least squares method, so as to find out if each variable has<br />
significant explanatory power to the default probabilities.<br />
Table 4 is the analysis results. From Table 4, we can see<br />
that other than revenue growth, all variables can<br />
significantly explain the default probabilities. Through<br />
the above selection <strong>of</strong> variables, from the variables that<br />
we have selected arbitrarily, we can find out those that<br />
are proven by statistical inference to be closely related to<br />
default probabilities.<br />
Table 4: Analysis <strong>of</strong> simple regression <strong>of</strong> variables<br />
Coefficie<br />
nt<br />
T-<br />
value<br />
P-<br />
value<br />
Includ<br />
/Exclud<br />
ROA<br />
(Before Interest and Tax)<br />
-0.0066 -3.77 0.00 Includ<br />
Compound ROA -0.0069 -2.70 -0.01 Includ<br />
Quick Ratio -0.0012 -5.81 -0.00 Includ<br />
Total Asset Turnover -0.0066 -3.77 -0.00 Includ<br />
Revenue Growth -0.0004 -1.49 0.14 Exclud<br />
Debt / Equity 0.0012 4.11 0.00 Includ<br />
As "Industry" is a dummy variable, regression cannot<br />
be used to verify its correlation with the default<br />
probabilities. In this research we assume it is a<br />
significant variable without testing.<br />
In order to test if collinearity exists among the selected<br />
variables, we also look at the VIF values. The VIF values<br />
<strong>of</strong> the variables are calculated and listed in Table 5.<br />
Table 5: Table <strong>of</strong> VIF <strong>of</strong> Selected Variables<br />
Selected Variables<br />
ROA<br />
VIF Value Included/Excluded<br />
(Before Interest and<br />
Tax)<br />
2.3122 Included<br />
Compound ROA 2.1243 Included<br />
Quick Ratio 1.3848 Included<br />
Total Asset Turnover 1.2623 Included<br />
Debt / Equity 1.3717 Included<br />
From Table 5, only Return on Assets Ratio (Before<br />
Interest and Tax) and Compound Return on Assets Ratio<br />
generate higher VIF values. Based on the usual standard<br />
that VIF less than 10 will not jeopardize the parameter<br />
estimates, we infer that there are no multi-collinearity<br />
issues with the above variables.<br />
(d). Validation<br />
We have conducted regression to the best variables and<br />
common variables <strong>of</strong> each model ,we can define the<br />
accuracy ratio <strong>of</strong> each model as the numbers <strong>of</strong><br />
companies whose regression results are in conformity<br />
with physical facts divided the numbers <strong>of</strong> observations<br />
(10,032) .In order to verify the efficiency <strong>of</strong> the variables<br />
in the models, we have listed out the accuracy ratios <strong>of</strong><br />
the models with the best variables set and common<br />
variables set respectively in Table 6 and Table 7:<br />
Table 6: Accuracy Ratio <strong>of</strong> Each Model with the Best Variables Set<br />
Model<br />
Merton<br />
Artificial<br />
Accuracy<br />
Logit Probit Discriminant Hazard<br />
Option<br />
Neural<br />
Ratio<br />
Regression Regression Analysis Ratio<br />
Pricing<br />
Network<br />
In-Sample N/A* 0.9566 0.9556 1.0000 0.8044 1.0000<br />
Out-Sample 0.4667 0.9111 0.9289 0.9378 0.7778 0.8756<br />
Note: As the Merton model does not require In-Sample parameters, such data is not available<br />
Model<br />
Accuracy Ratio<br />
Table 7: Accuracy Ratio <strong>of</strong> Each Model with the Common Variables Set<br />
Merton<br />
Option<br />
Pricing<br />
Logit<br />
Regression<br />
Probit<br />
Regression<br />
Discriminant<br />
Analysis<br />
Hazard Ratio<br />
Artificial<br />
Neural<br />
Network<br />
In-Sample N/A* 0.9578 0.9378 0.9589 0.7956 1.0000<br />
Out-Sample 0.5467 0.9200 0.9211 0.9289 0.7889 0.8067<br />
© 2011 ACADEMY PUBLISHER
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1873<br />
From the validation results, we can see that the<br />
accuracy ratios <strong>of</strong> using common variables set are similar<br />
to the accuracy ratios using best variables set. This<br />
demonstrates it is feasible to establish a common<br />
variables set which can be applied to different models.<br />
C.Credit default probabilities<br />
During validation, we have obtained the default<br />
probabilities <strong>of</strong> each sample company under different<br />
models. Once we have a representative probability rate<br />
(score) for each sample, we can carried out the grading<br />
with the scores.<br />
Regarding the estimation <strong>of</strong> default probabilities, the<br />
most direct approach is to acquire the ratings <strong>of</strong> the<br />
samples in different periods, and obtain the long run<br />
average estimated default probabilities under each grade<br />
by building a transition matrix.<br />
Take Logit Model as an example; based on the rating<br />
results in 2005 & 2006 and information <strong>of</strong> default<br />
occurrence <strong>of</strong> the companies in 2006, we can build a<br />
transition matrix as in Table 8.<br />
Table 8: 2005-2006 Transition Matrixes<br />
1 2 3 4 5 6 7 8 9 10 Default<br />
1 72% 16% 8% 0% 0% 3% 0% 1% 0% 0% 0%<br />
2 9% 53% 26% 3% 5% 2% 2% 2% 0% 0% 0%<br />
3 12% 10% 39% 25% 14% 2% 0% 3% 3% 2% 0%<br />
4 0% 4% 19% 33% 23% 17% 4% 0% 0% 0% 0%<br />
5 0% 0% 5% 15% 41% 26% 8% 5% 0% 0% 0%<br />
6 0% 2% 2% 3% 14% 34% 29% 14% 2% 0% 2%<br />
7 0% 0% 0% 0% 3% 24% 31% 29% 3% 0% 9%<br />
8 0% 0% 0% 2% 0% 7% 11% 40% 28% 7% 4%<br />
9 0% 0% 0% 0% 0% 1% 1% 22% 46% 26% 5%<br />
10 0% 0% 0% 0% 0% 2% 0% 0% 12% 65% 22%<br />
The data in transition matrix represents the probability<br />
that credit quality migrate from some rate to another rate<br />
after a year. This probability can be calculated as follows:<br />
P<br />
i, j<br />
=<br />
n<br />
n<br />
1, j<br />
0, j<br />
n 0, j is the number <strong>of</strong> companies with credit rating i at<br />
n 1, j<br />
is the number <strong>of</strong> companies with credit rating j<br />
If data deficiencies result in banks failing to estimate<br />
default probabilities, we need quantify the internal ratings<br />
by use <strong>of</strong> other approaches. By Carey、Hrycay (2001),<br />
Mapping method can be used to calculatedefault<br />
probabilities.<br />
Based on the above mentioned internal rating results,<br />
using the rating results <strong>of</strong> Logit Model as indicators, we<br />
find out the default probabilities from internal ratings by<br />
various mapping methods including judgmental,<br />
mechanical and weighted average mappings.<br />
t=0,<br />
at t=1.<br />
Table 9: Simulation Mapping Results<br />
Actual Data Mapping<br />
Internal<br />
Rating<br />
No. <strong>of</strong><br />
Companies<br />
No. <strong>of</strong><br />
Defaulting<br />
Companies<br />
Default<br />
Probability<br />
(PD)<br />
Median <strong>of</strong><br />
Logit<br />
Regression<br />
Rating<br />
PD that<br />
corresponds<br />
to the median<br />
Average <strong>of</strong><br />
Logit<br />
Model<br />
Ratings<br />
Weighted<br />
Average<br />
PD<br />
1 1089 0 0.00% 1 0.00% 1.27 0.04%<br />
2 978 0 0.00% 2 0.00% 2.39 0.29%<br />
3 942 3 0.31% 3 0.00% 3.30 0.55%<br />
4 882 3 0.34% 4 0.67% 4.16 0.66%<br />
5 948 12 1.27% 5 0.98% 5.09 0.81%<br />
6 1032 15 1.45% 6 0.58% 5.77 1.22%<br />
7 936 30 3.21% 7 1.76% 6.66 2.58%<br />
8 1116 48 4.30% 8 5.36% 7.55 4.58%<br />
9 1044 69 6.61% 9 7.80% 8.69 7.77%<br />
10 1065 105 14.37% 10 14.53% 9.81 13.39%<br />
Under Judgmental mapping, at first we map internal<br />
grades to external grades subjectively, then take default<br />
probabilities from external ratings as probabilities from<br />
internal ratings. Under mechanical mapping, we sort the<br />
company in each internal grade by corresponding external<br />
grades, and then take the median <strong>of</strong> average default<br />
probabilities in external grades as default probabilities in<br />
© 2011 ACADEMY PUBLISHER<br />
each internal grade. Under weighted average mapping,<br />
we take the weighted average <strong>of</strong> average default<br />
probabilities in external grades corresponded to internal<br />
grades as default probabilities in each internal grade. Due<br />
to Judgmental mapping is lack <strong>of</strong> logical basis, we just<br />
calculate the default probabilities with mechanical
1874 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
mapping and weighted average mapping respectively (see<br />
Table9).<br />
We can see that it is more suitable to use median to<br />
compute the actual default probability for the safe grades<br />
Logit<br />
Probit<br />
No. <strong>of</strong><br />
Co.<br />
(1~4), while it is better to use weighted average to<br />
estimate actual default probability for the risky grades<br />
(8~10).<br />
Table 10: Table <strong>of</strong> Benchmark Ratings<br />
1 2 3 4 5 6 7 8 9 10<br />
1 363 95% 3% 2% 0% 0% 0% 0% 0% 0% 0%<br />
2 326 5% 91% 4% 0% 0% 0% 0% 0% 0% 0%<br />
3 314 0% 4% 93% 3% 0% 0% 0% 0% 0% 0%<br />
4 294 0% 0% 5% 90% 4% 1% 0% 0% 0% 0%<br />
5 316 0% 0% 0% 3% 92% 5% 0% 0% 0% 0%<br />
6 344 0% 0% 0% 1% 5% 91% 3% 0% 0% 0%<br />
7 312 0% 0% 0% 0% 3% 4% 89% 3% 1% 0%<br />
8 372 0% 0% 0% 0% 0% 0% 1% 94% 5% 0%<br />
9 348 0% 0% 0% 0% 0% 0% 0% 5% 90% 5%<br />
10 355 0% 0% 0% 0% 0% 0% 0% 1% 11% 88%<br />
To validate the default probabilities, we adopt<br />
benchmark comparisons for empirical explanations.<br />
Continue with the above analysis using the results<br />
corresponding to the medians under mapping, we use<br />
Logit Model as benchmarks for ratings comparison.<br />
Then carry out the benchmarks comparison to the Probit<br />
Model. The data represents the distribution <strong>of</strong> the number<br />
<strong>of</strong> some rate under logit model. For example, when the<br />
rate is 1 under logit model, 95%number <strong>of</strong> company<br />
belongs to 1 under probit model; only 5% number <strong>of</strong><br />
company is other rate. The results are shown in Table 10.<br />
From Table 10, we can see that the comparison results<br />
for each grade are acceptable. Consequently, we can<br />
build a credit scoring model based on these findings.<br />
V. CONCLUSION<br />
The effectiveness <strong>of</strong> credit risks management relies on<br />
whether it can operate with the local environment.<br />
Therefore, choosing the variable set that fits in with the<br />
local conditions is critical to the performance <strong>of</strong> a credit<br />
rating model. But most <strong>of</strong> researchers always choose<br />
corresponding variables for different model. In this paper,<br />
we have adopted Merton Option Pricing Model,<br />
Discriminant Analysis Model, Regression Analysis<br />
Models (Logit Model and Probit Model), Survival<br />
Analysis Model and Artificial Neural Network Model in<br />
finding the common variables set for application in credit<br />
risks management models.Our findings show that five<br />
variables, namely Return on Assets Ratio (Before Interest<br />
and Tax), Compound Return on Assets Ratio, Quick<br />
Ratio,Total Asset Turnover, Debt-to-EquityRatio, are<br />
applicable to different credit risk rating<br />
models.Moreover, this paper estimates long term average<br />
default probability by using Transition Matrix and<br />
Mapping method. Validation findings also show that they<br />
have good forecasting abilities. Such findings help to<br />
simplify the application <strong>of</strong> credit risks management<br />
models and better adapt the models to the local conditions<br />
in China. We believe that the research methods presented<br />
in this paper can also be applied to other countries or<br />
regions around the worlds. It can serve as a good<br />
© 2011 ACADEMY PUBLISHER<br />
reference for establishing credit risks management<br />
models that fit with the local conditions.<br />
ACKNOWLEDGMENT<br />
This work was supported by the Fundamental Research<br />
Funds for the Central Universities (2010221040), China<br />
National Social Science Fund (09AZD045), Ministry <strong>of</strong><br />
Education for Humanities and Social Sciences<br />
(08JA630004), Anhui Provincial Natural Science<br />
Research Project for Universities (KJ2010A072) and<br />
China National Bureau <strong>of</strong> Statistics Fund (2009LZ045).<br />
We would like to thank the editor, associate editor, and<br />
referees for careful review and insightful comments,<br />
which have led to significant improvement <strong>of</strong> the article.<br />
REFERENCES<br />
[1] Basel Committee on Banking Supervision (1999), Credit<br />
Risk Modeling: Current Practice and Application, Bank for<br />
International Settlements.<br />
[2] Basel Committee on Banking Supervision (2003),<br />
Consultative Document: The New Basel Capital Accord,<br />
Bank for International Settlements.<br />
[3] Carey, Mark and M. Hrycay (2001), Parameterzing credit<br />
risk models with rating data, <strong>Journal</strong> <strong>of</strong> Banking & Finance<br />
25, p. 197-270<br />
[4] Division <strong>of</strong> Banking Supervision and Regulation (1998),<br />
Bank Holding Company Supervision Manual, Board <strong>of</strong><br />
Governors <strong>of</strong> the Federal Reserve System<br />
[5] Division <strong>of</strong> Banking Supervision and Regulation(2003),<br />
Draft Supervisory Guidance on Internal Ratings-Based<br />
Systems for Corporate Credit, Board <strong>of</strong> Governors <strong>of</strong> the<br />
Federal Reserve System<br />
[6] Engelmann, Bernd, E. Hayden and D. Tasche (2003),<br />
Testing Rating Accuracy, Risk Falkenstein, E., A. Boral,<br />
and L. Carty (2000), RiskCalc Private Model: Moodys<br />
Default Model for Private Firms, Moodys Investors<br />
Service.<br />
[7] Ferguson Jr., Roger W. (2003), Basel II: Some Issues for<br />
Implementation,Basel Sessions 2003 Speech, Institute <strong>of</strong><br />
International Finance, New York .<br />
[8] Keenan, S. C. and J. R. Sobehart (2001), Performance<br />
Measures for Credit Risk Models, Moody’s Risk<br />
Management Services Research Report.
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1875<br />
Corporate-, Product-, and User-Image<br />
Dimensions and Purchase Intentions<br />
The Mediating Role <strong>of</strong> Cognitive and Affective Attitudes<br />
Xian Guo Li*, Xia Wang, Yu Juan Cai<br />
Department <strong>of</strong> Marketing<br />
School <strong>of</strong> Business, Renmin University <strong>of</strong> China, Beijing, China<br />
Abstract—This study investigates the effects <strong>of</strong> corporate-,<br />
product- and user image dimensions on purchase intentions,<br />
with cognitive and affective attitudes as mediator. A<br />
questionnaire survey was conducted with convenience<br />
sample. The results demonstrate significant effects <strong>of</strong> three<br />
brand image dimensions on purchase intention. In addition,<br />
the cognitive and affective attitudes fully or partially<br />
account for the relationship. This study contributes to the<br />
understanding <strong>of</strong> the assessment <strong>of</strong> the relationship between<br />
brand image dimensions and purchasing behavior.<br />
Implications for brand management are also discussed.<br />
Index Terms—Corporate Image, Product Image, User<br />
Image, Purchase Intention<br />
I. INTRODUCTION<br />
Brand image has been an important concept in<br />
consumer behavior research since the early 1950s. Both<br />
marketing researchers and marketers have long advocated<br />
the use <strong>of</strong> a clearly defined brand image as a basis for<br />
market success. A well-communicated brand image<br />
enables consumers to identify the needs satisfied by the<br />
brand and thereby differentiate the brand from its<br />
competitors [1, 2]. In fact, developing a brand image<br />
strategy has been described as the first and most vital step<br />
in positioning a brand and driving brand equity in the<br />
marketplace [3-5].As the growing importance <strong>of</strong> brand<br />
image strategy in marketing, a research issue evolved that<br />
how the brand image perceptions affect consumers<br />
purchasing behavior.<br />
Brand image is the most efficient way to talk to<br />
consumers via translating the different benefits about a<br />
brand. One common mistake brand strategists make is<br />
having too narrowed a view <strong>of</strong> the brand and only<br />
focusing some attributes when creating a brand’s image<br />
[6]. Given consumers’ perceptions may not be product<br />
specific; brand image is a multi-dimensional construct.<br />
The image <strong>of</strong> a brand can be described as having three<br />
contributing sub-images, the image <strong>of</strong> the provider <strong>of</strong> the<br />
product/service, or corporate image; the image <strong>of</strong> the<br />
user; and the image <strong>of</strong> the product/service itself [4]. Thus,<br />
two questions arise: what makes an effective brand and,<br />
Corresponding author: Xian Guo Li is with the Department <strong>of</strong><br />
Marketing, School <strong>of</strong> Business, Renmin University <strong>of</strong> China (Phone:<br />
86-13910602316; Email: rdlxg@126.com)<br />
© 2011 ACADEMY PUBLISHER<br />
doi:10.4304/jcp.6.9.1875-1879<br />
second, how the company can effectively communicate to<br />
consumers with different brand image strategy?<br />
In the aforementioned studies, however, relatively little<br />
empirical evidence has been provided for the effects <strong>of</strong><br />
these dimensions on purchase intention. Especially for<br />
the Chinese markets, as brand image perception varies<br />
across culture [7], the effects <strong>of</strong> brand image dimensions<br />
need to be examined further. Thus, the purpose <strong>of</strong> this<br />
study is to investigate the predicting roles <strong>of</strong> corporate-,<br />
product-, and user-image on purchase intention in the<br />
context <strong>of</strong> Chinese mobile-phone market, and the<br />
mediating role <strong>of</strong> cognitive and affective attitude are also<br />
examined in this study.<br />
This paper is structured as follows. We first review the<br />
literature on key issues involving brand image,<br />
associations with purchasing behavior, and the mediating<br />
role <strong>of</strong> attitude. The data and methods <strong>of</strong> the study follow<br />
in the section. Empirical evidence on the effects <strong>of</strong> brand<br />
image dimensions on purchasing intention is provided,<br />
and the mediating role <strong>of</strong> cognitive and affective attitude<br />
is also demonstrated. The paper ends with conclusions<br />
and implications. It is expected that this study will<br />
provide a more thorough understanding <strong>of</strong> building a<br />
company’s brand image strategy focusing on three brand<br />
image dimensions in Chinese mobile-phone industry.<br />
II. LITERATURE REVIEW<br />
A. Brand Image<br />
There has been general agreement that brands—at least<br />
some brands—do have images, defined as the<br />
associations linked to a brand [4], or perceptions about a<br />
brand as reflected by the brand associations held in<br />
consumer memory [3]. When consumers see a particular<br />
brand, the brand association is any idea caused by that<br />
certain brand, including feelings, experiences, appraisals,<br />
and brand positioning [3]. The brand image perception<br />
varies across categories, brands [4] and culture [7], thus<br />
need to be investigated in multi-cultures, especially for<br />
Chinese markets.<br />
Brand image is a complex constructs and can be made<br />
<strong>of</strong> several dimensions [8]. Brand association is the mutual<br />
combination <strong>of</strong> informational nodes and come from all<br />
possible forms, and may reflect product characteristics or<br />
independent characteristics outside the product [3]. Biel
1876 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
[4] suggests brand image has three components: corporate<br />
image, image <strong>of</strong> the user and image <strong>of</strong> the product. While<br />
Hsieh et al [9] extends the product image with corporate<br />
image and country image, and inspect the relationship<br />
between product-, corporate-, country- image and<br />
purchase behavior, which was also verified in multicultures<br />
[10].<br />
Park et al [2] states brand image incorporates the<br />
functional, experiential and symbolic benefits to the<br />
consumer; many brands <strong>of</strong>fer a mixture <strong>of</strong> symbolic,<br />
functional, and experiential benefits. A brand with<br />
functional benefits is one designed to satisfy consumers'<br />
needs to solve consumption-related problems. A brand<br />
with symbolic (or social) benefits is one designed to<br />
fulfill consumers' desires for self-enhancement, role<br />
position, group membership, or ego identification.<br />
Finally, a brand with experiential benefits is one designed<br />
to fulfill consumers' desires for sensory pleasure, variety,<br />
or cognitive stimulation.<br />
The current study extends Hsieh’s et al [9] study by<br />
following Biel’s [4] definition and adopts the three<br />
previously mentioned brand concepts as corporate image,<br />
product image, and user image. The brand theorists<br />
suggest that what a person knows about a company can<br />
influence perceptions <strong>of</strong> the company's products, e.g. the<br />
corporate ability associations and corporate social<br />
responsibility associations will influence consumers’<br />
beliefs about and attitudes towards the products <strong>of</strong> that<br />
company [11], thus corporate brand image may affect the<br />
product evaluations, and the relationship is moderated by<br />
perceived risk [12]. The product image is related to the<br />
benefits attached to the products. As the symbolic,<br />
functional, and experiential benefits <strong>of</strong> the products have<br />
been proved to lead to brand preference [6], the product<br />
image will also influence the product evaluations. The<br />
user image refers to whether the brand personality is<br />
congruent with the consumers [13]. If the brand<br />
personality fit the consumers’ self-concept, the product<br />
may receive a high evaluation.<br />
With regard to the performance <strong>of</strong> brand image, Aaker<br />
[1] claims that brand association aid in acquiring or<br />
handling information, creating positive attitudes or<br />
feelings, positioning brand and differentiating it from<br />
competitors as well as creating value for the company.<br />
Empirical evidence suggests that brand image has<br />
positive influence on brand-extension attitude [8].<br />
Moreover, Krishnan [5] demonstrates that compared to<br />
brands with low equity, high equity brands will have<br />
greater number <strong>of</strong> positive associations, more unique<br />
associations from competing brands, fewer unique<br />
associations from the category, and more associations<br />
from direct experiences and word-<strong>of</strong>-mouth, which has<br />
directly verified the relationship between association<br />
pattern <strong>of</strong> brand image and brand equity.<br />
B. Mediating Role <strong>of</strong> Consumers Attitude<br />
Brand associations in the study <strong>of</strong> Keller [3] are<br />
classified into three major categories with respect to their<br />
level <strong>of</strong> abstraction (i.e., attribute, benefit, and overall<br />
brand attitude). Here, attribute refers to descriptive<br />
features that characterize a product or service, benefit is<br />
© 2011 ACADEMY PUBLISHER<br />
the personal value that consumers attach to the product or<br />
service, and brand attitude is consumers' overall<br />
evaluation <strong>of</strong> the brand [9]. Ideally, in consumers'<br />
memory, brand-image perception should encompass all<br />
three types <strong>of</strong> brand associations. However, given the<br />
entailed complexity, most <strong>of</strong> the studies incorporate only<br />
benefit associations as the key elements [6, 9]. The<br />
corporate-, product-, and user image <strong>of</strong> this study also<br />
embrace only the benefit associations <strong>of</strong> brand image.<br />
Thus, overall brand attitude as an important part <strong>of</strong> brand<br />
association should be investigated further.<br />
Brand attitudes are important because they <strong>of</strong>ten form<br />
the basis for consumer behavior (e.g., brand choice).<br />
Though different models <strong>of</strong> brand attitudes have been<br />
proposed, one widely accepted approach is based on a<br />
multi-attribute formulation in which brand attitudes are a<br />
function <strong>of</strong> the associated attributes and benefits that are<br />
salient for the brand [3]. According to the theory <strong>of</strong><br />
planned behavior, there are three conceptually<br />
independent determinants <strong>of</strong> intention: attitude toward the<br />
behavior, subjective norm and perceived behavioral<br />
control. As a general rule, the more favorable the attitude,<br />
the stronger should be an individual’s intention to<br />
perform the behavior under consideration [14]. According<br />
to the planned behavior theory, attitudes develop<br />
reasonably from the beliefs people hold about the object<br />
[14]. Thus the attitude may mediate the relationship<br />
between brand image beliefs and purchase intention.<br />
The basic theory <strong>of</strong> planned behavior model was<br />
expanded to include the separation <strong>of</strong> affective and<br />
cognitive predictors <strong>of</strong> attitude towards purchase<br />
intention. The majority <strong>of</strong> social psychology literature<br />
suggests that attitudes are composed <strong>of</strong> cognitive,<br />
affective, and behavioral parts. This multidimensional<br />
view <strong>of</strong> attitude implies that consumers’ willingness to<br />
buy may be influenced by cognitive and affective<br />
antecedents [15]. Following this, we propose that<br />
purchase intention can be predicted by cognitive and<br />
affective attitudes.<br />
C. Research Questions and Hypotheses<br />
Drawing from the previous literature and field<br />
observation, we set up the research questions in an<br />
attempt to explore the relationship between brand image<br />
dimensions and purchase intention. Generally, the social<br />
responsibility and consumer concern <strong>of</strong> an enterprise will<br />
increase the consumers’ willingness to buy their products,<br />
and a well product or service image may increase the<br />
consumers’ brand usage. Meanwhile, consumer would<br />
like to buy the products with congruent personality. Thus,<br />
we hypothesize that:<br />
H1a: Corporate image has positive influence on<br />
purchase intentions.<br />
H1b: Product image has positive influence on purchase<br />
intentions.<br />
H1c: User image has positive influence on purchase<br />
intentions.<br />
According to the planned behavior model, the attitude<br />
may mediate the relationship between beliefs and<br />
intention. Thus with regard to this study, we separate the
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1877<br />
two components <strong>of</strong> cognitive attitude and affective<br />
attitude, and hypothesize that:<br />
H2a: Cognitive attitude will mediate the relationship<br />
between corporate image and purchase intention.<br />
H2b: Cognitive attitude will mediate the relationship<br />
between product image and purchase intention.<br />
H2c: Cognitive attitude will mediate the relationship<br />
between user image and purchase intention.<br />
H3a: Affective attitude will mediate the relationship<br />
between corporate image and purchase intention.<br />
H3b: Affective attitude will mediate the relationship<br />
between product image and purchase intention.<br />
H3c: Affective attitude will mediate the relationship<br />
between user image and purchase intention.<br />
III. DATA AND METHOD<br />
A. Instruments<br />
All our measures employ items from multiple-item<br />
scales that have been tested and used in previous studies.<br />
The dependent variables <strong>of</strong> purchase intention gathered<br />
from the work <strong>of</strong> Dodds et al [16]. The purchase intention<br />
was measured on four items as “The likelihood <strong>of</strong> buying<br />
products <strong>of</strong> this brand is very high”, “I would consider<br />
buying products <strong>of</strong> this brand”, “The probability that I<br />
would like to buy products <strong>of</strong> this brand is very high”,<br />
and “My willingness to buy this product is very high”.<br />
The predictor variables <strong>of</strong> three brand image<br />
dimensions were measured on multi-items modified from<br />
Zhuohao et al [17] and Xiucheng and Jie [18]. The<br />
corporate image was measured as “The innovation and<br />
update <strong>of</strong> the products <strong>of</strong> this corporate is strong”, “The<br />
corporate <strong>of</strong> this brand care for customer very much”, and<br />
“The corporate <strong>of</strong> this brand have a well impression”.<br />
The product image was measured with the following<br />
indicators: Function, Style, Durability, and Quality. And<br />
the user image was measured with the following<br />
statements. “I can easily imagine this brand as a person”,<br />
“This brand have a strong personality”, and “The<br />
personality <strong>of</strong> this brand matches with mine”.<br />
The mediator variables <strong>of</strong> cognitive attitude and<br />
affective attitude were adapted from the study <strong>of</strong><br />
Verplanken et al [19]. The affective attitudes were:<br />
favorable, pleasant, comfortable, exciting, and attractive.<br />
The cognitive attitudes were good, wise, positive, useful,<br />
and worthy. All <strong>of</strong> the items were evaluated on a sevenpoint<br />
Likert scale ranging from “strongly disagree” to<br />
“strong degree”.<br />
B. Participations<br />
The study was investigated in mobile-phone market,<br />
because the products have the following characteristic:<br />
(1) It is a hedonic and utilitarian products; (2) It is a highinvolvement<br />
characteristic; and (3) It is widely adopted in<br />
Chinese markets. Data was collected from three<br />
universities in Beijing. The final effective sample size<br />
was 268.<br />
© 2011 ACADEMY PUBLISHER<br />
IV. RESULTS<br />
A. The Effect <strong>of</strong> Brand Image Dimensions on Purchase<br />
Intention<br />
We expect that the brand image dimensions have<br />
positive dimensions have positive influence on purchase<br />
intention. The coefficients <strong>of</strong> brand image dimensions<br />
and purchase intention were estimated and presented in<br />
table 1. All <strong>of</strong> the coefficients for corporate image,<br />
product image and user image were significantly positive,<br />
indicating positive relationships between brand image<br />
dimensions and purchase intention. H1 to H3 were<br />
supported. The standardized path coefficients <strong>of</strong> product<br />
image leading to purchasing intention showed a relative<br />
stronger relationship than the two other dimensions,<br />
indicating the product image should be given more<br />
attention in the context <strong>of</strong> this study.<br />
TABLE I. REGRESSION RESULTS OF BRAND IMAGE<br />
DIMENSIONS ON PURCHASE INTENTION<br />
B SE B β<br />
Constant 1.044*** .195<br />
Corporate Image .094* .047 .104<br />
Product Image .423*** .046 .451<br />
User Image .202*** .052 .212<br />
*p
1878 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
partially explain the relationship between product image<br />
and purchase intention.<br />
TABLE II. THE MEDIATING EFFECT OF COGNITIVE<br />
ATTITUDE ON THE RELATIONSHIP BETWEEN CORPORATE<br />
IMAGE AND PURCHASE INTENTION<br />
Step Predictor:<br />
Corporate Image<br />
B SE B β<br />
1 Corporate Image .411*** .043 .456<br />
2 Corporate Image .434*** .036 .540<br />
3 Corporate Image .035 .035 .039<br />
Cognitive Attitude .867*** .043 .772<br />
4 Sobel Z 10.32***<br />
*p
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1879<br />
previous studies, which suppose corporate image, product<br />
image, and user image will affect the consumers’<br />
willingness to purchase.<br />
Results <strong>of</strong> this study indicates that building brand<br />
image <strong>of</strong> mobile-phone industry should focus more on the<br />
product image to leading consumers’ buying decisions,<br />
however, other dimensions <strong>of</strong> corporate and user image<br />
should not be ignored as they have positive relationship<br />
with consumers’ purchasing behavior.<br />
Besides that, the mediating roles <strong>of</strong> cognitive and<br />
affective attitudes were also examined in the study and<br />
reveal that attitudes can fully or partially account for the<br />
relationships between brand image dimensions and<br />
purchase intentions.<br />
The contribution <strong>of</strong> this study is to empirically<br />
investigate the effects <strong>of</strong> brand image dimensions on<br />
purchasing behavior, and examine the influence route<br />
deeply from the view <strong>of</strong> planned behavior model, which<br />
may contribute to understanding the relationships<br />
between brand image dimensions and performance, and<br />
make a supplementary for the planned behavior model<br />
further. The managerial implication is that help the<br />
enterprises comprehend the three dimensions <strong>of</strong> brand<br />
image, and make appropriate marketing campaigns.<br />
Limitation <strong>of</strong> this study includes the lack <strong>of</strong> category<br />
specific investigation as the contribution <strong>of</strong> the three<br />
dimensions <strong>of</strong> brand image varies by product category<br />
and by brand [4]. Sampling frames is coming from the<br />
students in the university. The convenience sample may<br />
limit the generalizability <strong>of</strong> this study. Other variables<br />
such as subjective norm should also be controlled in the<br />
study, as they may affect the consumers’ willingness to<br />
buy [16]. Direction <strong>of</strong> further research is to conduct<br />
research in other categories and increase the<br />
generalizability <strong>of</strong> the study.<br />
REFERENCE<br />
[1] Aaker, D. (1991). Managing Brand Equity. Ontario: The<br />
Free Press.<br />
[2] Park, C.W., Jaworski, B.J., Maclnnis, D.J. (1986).<br />
“Strategic Brand Concept-Image Management”, <strong>Journal</strong><br />
<strong>of</strong> Marketing, 50(4): 135-145.<br />
[3] Keller, K.L. (1993). “Conceptualizing, measuring, and<br />
managing customer based brand equity”, <strong>Journal</strong> <strong>of</strong><br />
Marketing, 57 (1), 1-22.<br />
[4] Biel, A. (1992). “How Brand Image Drives Brand Equity”,<br />
<strong>Journal</strong> <strong>of</strong> Advertising Research, 32(6): 6-12.<br />
[5] Krishnan, H. S. (1996). “Characteristics <strong>of</strong> memory<br />
associations: A consumer-based brand equity perspective”,<br />
International <strong>Journal</strong> <strong>of</strong> Research in Marketing, 3, 389-<br />
405.<br />
[6] Salciuviene, L., Lee, K., Yu, C. (2007). “The Impact <strong>of</strong><br />
© 2011 ACADEMY PUBLISHER<br />
Brand Image Dimensions on Brand Preference”,<br />
Economic and Management, 12, 464-469.<br />
[7] Park, H., Rabolt, N.J. (2009). “Cultural Value,<br />
Consumption Value, and Global Brand Image: A Cross-<br />
National Study”, Psychology & Marketing, 26(8): 714-<br />
735.<br />
[8] Salinas, E.M., and Perez, J.M.P. (2009). “Modeling the<br />
brand extensions' influence on brand image”, <strong>Journal</strong> <strong>of</strong><br />
Business Research, 62: 50-60.<br />
[9] Hsieh, M., Pan, S., Setiono, R. (2004). “Product-,<br />
Corporate-, and Country Image Dimensions and Purchase<br />
Behavior: A Multicountry Analysis”, <strong>Journal</strong> <strong>of</strong> the<br />
<strong>Academy</strong> <strong>of</strong> Marketing Science, 32(3): 251-270.<br />
[10] Chung, J., Pysarchik, D.T., and Hwang, S. (2009),<br />
“Effects <strong>of</strong> Country-<strong>of</strong>-Manufacture and Brand Image on<br />
Korean Consumers’ Purchase Intention”, <strong>Journal</strong> <strong>of</strong><br />
Global Marketing, 22, 21-41.<br />
[11] Brown, T.J., Dacin, P.A. (1997). “The Company and the<br />
Product: Corporate Associations and Consumer Product<br />
Responses”, <strong>Journal</strong> <strong>of</strong> Marketing, 61(1): 68-84.<br />
[12] Gurhan-Canli, Z., and Batra, R. (2004). “When Corporate<br />
Image Affects Product Evaluations: The Moderating Role<br />
<strong>of</strong> Perceived Risk”, <strong>Journal</strong> <strong>of</strong> Marketing Research, 41(2):<br />
197-205.<br />
[13] Sirgy, J. (1982). “Self-concept in Consumer Behavior: A<br />
Critical Review”, <strong>Journal</strong> <strong>of</strong> Consumer Research, 9, 287-<br />
300.<br />
[14] Ajzen, I. (1991), “The Theory <strong>of</strong> Planned Behavior, ”<br />
Organizational Behavior and Human Decision Processes,<br />
50, 179-211.<br />
[15] Zajonc, R.B., Markus, H. (1982), “Affective and<br />
Cognitive Factors in Preferences”, <strong>Journal</strong> <strong>of</strong> Consumer<br />
Research, 9(2): 123-131.<br />
[16] Dodds, W., Monroe, K.B., Grewal, D. (1991), “Effects <strong>of</strong><br />
Price, Brand, and Store Information on Buyers' Product<br />
Evaluations”, <strong>Journal</strong> <strong>of</strong> Marketing Research, 28: 307-319.<br />
[17] Zhuohao, C., Zhi, L., and Qingyun, J. (2006), “How Does<br />
Brand Personality Influence Consumer’s Attitudes ? A<br />
Study from the Perspective <strong>of</strong> Consumer Brand<br />
Cognition”, <strong>Journal</strong> <strong>of</strong> Marketing Science, 2(2): 103-116<br />
(In Chinese).<br />
[18] Xiucheng, F., and Jie, C. (2002), “Measurement <strong>of</strong> Brand<br />
Image: A Brand Identity-Based Integrated Model and<br />
Empirical Study”, Nankai <strong>Journal</strong> (Philosophy and Social<br />
Science Edition), 3, 65-71.<br />
[19] Verplanken, B., H<strong>of</strong>stee, G., Janssen, H.J.W. (1998),<br />
“Accessibility <strong>of</strong> Affective versus Cognitive Components<br />
<strong>of</strong> Attitudes”, European <strong>Journal</strong> <strong>of</strong> Social Psychology, 28,<br />
23-35.<br />
[20] Baron, R.M., & Kenny, D. A. (1986). “The Moderator -<br />
Mediator Variable Distinction in Social Psychological<br />
Research: Conceptual, Strategic, and Statistical<br />
Considerations”, <strong>Journal</strong> <strong>of</strong> Personality and Social<br />
Psychology, 51(6):1173-1182.
1880 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
A Microcomputer-Based Predictive Digital<br />
Current Programmed Control System for Threephase<br />
PWM Rectifier<br />
Zhongjiu Zheng<br />
College <strong>of</strong> Electrical Engineering, Dalian University <strong>of</strong> Technology, Dalian, 116024, P .R. China<br />
zhengzhongjiu@163.com<br />
Gu<strong>of</strong>eng Li, and Ninghui Wang<br />
College <strong>of</strong> Electrical Engineering, Dalian University <strong>of</strong> Technology, Dalian, 116024, P .R. China<br />
gu<strong>of</strong>enli@dlut.edu.cn, ninghuiw@263.net<br />
Abstract—The paper describes a microcomputer control<br />
system, which uses the floating-point digital signal processor<br />
TMS320LF2407 from Texas Instruments, for three-phase<br />
PWM rectifier. It could effectively eliminate harmonic<br />
distortion <strong>of</strong> line currents and provides power factor<br />
correction. Moreover, it can be save electrical energy and<br />
reduction <strong>of</strong> production cost. In the control system, the<br />
predictive current control in two-dimensional (α-β)<br />
stationary frame, makes the input current following the<br />
phase voltage in phase to get unity power factor; and space<br />
vector pulse wide modulation (SVPWM) generates the<br />
modulation wave. Finally, the three-phase PWM rectifier<br />
using the proposed control system is designed in<br />
Simulink/Matlab and executed in laboratory prototype, and<br />
the results are provided to verify the proposed control<br />
system in the end <strong>of</strong> the paper.<br />
Index Terms—PWM rectifier; Predictive digital current<br />
control; Space vector pulse wide modulation; Unity power<br />
factor<br />
I. INTRODUCTION<br />
In knowledge economy era, research in the field <strong>of</strong><br />
power electronics has taken a great interest in the power<br />
quality, such as power supply efficiency, saving electrical<br />
energy, economical, reliability, volume, and weight.<br />
Traditional uncontrolled three-phase rectifiers have been<br />
widely used in the industrial complexes, but the<br />
disadvantages are severity energy losses, high cost, big<br />
volume and weight, and introducing massive harmonic<br />
currents into the grid that does not fulfill the new<br />
standards for the electric grid.<br />
With the development <strong>of</strong> digital signal processors<br />
(DSP) control devices and the IGBT power devices, DSPbased<br />
controller for three-phase PWM rectifiers have<br />
been proposed in some papers [1]-[6] in which general<br />
purpose and floating-point DSPs are used. This technique<br />
uses a floating-point DSP to effectively eliminate system<br />
harmonics and it also provides power factor correction.<br />
Moreover, it can be save electrical energy and reduction<br />
<strong>of</strong> production cost.<br />
PWM rectifiers [7]-[11] as a non-polluting and<br />
economical equipment are going to be more popular<br />
because <strong>of</strong> several advantages described as:<br />
© 2011 ACADEMY PUBLISHER<br />
doi:10.4304/jcp.6.9.1880-1885<br />
- reduction <strong>of</strong> production cost;<br />
- saving electrical energy;<br />
-Low harmonic distortion <strong>of</strong> line currents;<br />
- Regulation <strong>of</strong> input power factor to unity;<br />
- Adjustment and stabilization <strong>of</strong> output DC voltage;<br />
-Bi-directional power flow;<br />
The objective <strong>of</strong> this paper is to present a economical<br />
predictive digital current control strategy <strong>of</strong> three-phase<br />
PWM rectifier based on modern floating-point digital<br />
signal processor (DSP) which facilitates the work on<br />
s<strong>of</strong>tware development. The proposed predictive digital<br />
current control system operates with constant switching<br />
frequency using space-vector modulation (SVM).The<br />
control system include predictive current algorithm,<br />
SVPWM control algorithm, proportional integral (PI)<br />
algorithm and so on. The predictive current control make<br />
the input current following the phase voltage in phase to<br />
get unity power factor; SVPWM generates the six via<br />
modulation wave; and the PI regulator keep the output<br />
voltage constant . In this way, the whole system <strong>of</strong> PWM<br />
rectifiers is obtained when the control algorithm and<br />
PWM generation are carried out using a digital signal<br />
processor (DSP) with minimal external hardware.<br />
II. MODELING FOR THE THREE-PHASE PWM RECTIFIER<br />
The three-phase boost PFC rectifiers is consisted <strong>of</strong> six<br />
switches with anti-paralleled diodes as shown in Figure 1<br />
.This topologies is ideally applicable to DC-linked AC<br />
Figure 1. The main circuit <strong>of</strong> three-phase PWM rectifier.
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1881<br />
motor drives since it draws sinusoidal input current ,and<br />
controls the DC bus voltage .Moreover ,its capability <strong>of</strong><br />
bi-directional power flow allows reverse operation ,which<br />
is especially advantageous for three-phase high power<br />
factor PWM rectifier.<br />
Assume that the three-phase voltage is symmetrical,<br />
stable and interior resistance is zero; three-phase loop<br />
resistance R and L are the same value respectively;<br />
switching loss and on-state voltage is neglectable;<br />
affection <strong>of</strong> distributing parameters is neglectable<br />
;switching frequency <strong>of</strong> the rectifier is high enough .<br />
The parameters in Figure 1 are listed below.<br />
e , e , e Phase voltage;<br />
a b c<br />
i , i , i Phase current;<br />
a b c<br />
U , U , U Voltage between leg midpoint and N<br />
AN BN CN<br />
point;<br />
L Input inductance;<br />
R Equivalent resistance <strong>of</strong> the loop;<br />
C Capacitance <strong>of</strong> the dc bus;<br />
S<br />
V dc Output DC voltage;<br />
i o Load current;<br />
R Load resistance;<br />
L<br />
We can define switch function as follows,<br />
S i<br />
⎧1<br />
= ⎨<br />
⎩0<br />
i phase upper switch is on<br />
i phase bottom switch is on<br />
(1)<br />
,i=a,b,c<br />
Hence, the mathematic model <strong>of</strong> PWM rectifier is:<br />
⎡<br />
⎢<br />
− R<br />
⎢<br />
⎢ 0<br />
A = ⎢<br />
⎢<br />
⎢ 0<br />
⎢<br />
⎢<br />
⎣ sa<br />
•<br />
Z X = A X + U<br />
0<br />
− R<br />
0<br />
s<br />
b<br />
0<br />
0<br />
− R<br />
s<br />
c<br />
− ( s<br />
a<br />
− ( s<br />
b<br />
− ( s<br />
[ L,<br />
L,<br />
L C ]<br />
Z ,<br />
S<br />
c<br />
1<br />
−<br />
3<br />
1<br />
−<br />
3<br />
1<br />
−<br />
3<br />
0<br />
∑<br />
k<br />
k = a,<br />
b,<br />
c<br />
∑<br />
k<br />
k = a,<br />
b,<br />
c<br />
∑<br />
k<br />
k = a,<br />
b,<br />
c<br />
(2)<br />
⎤<br />
s )<br />
⎥<br />
⎥<br />
⎥ (3)<br />
s )<br />
⎥<br />
⎥<br />
s ) ⎥<br />
⎥<br />
⎥<br />
⎦<br />
= (4)<br />
[ ] T<br />
e e , e i<br />
U = , −<br />
a<br />
, (5)<br />
b<br />
[ ] T<br />
i , i , i ,<br />
X V<br />
a<br />
b<br />
c<br />
c<br />
dc<br />
o<br />
= . (6)<br />
III. CONTROL SYSTEM FOR THREE-PHASE PWM RECTIFIER<br />
The control system schematic diagram <strong>of</strong> the threephase<br />
PWM rectifier is shown in figure 2. The control<br />
system adopts predictive current control in twodimensional<br />
(α-β) stationary frame ,pulse wide<br />
© 2011 ACADEMY PUBLISHER<br />
modulation mode is based on space vector ,DC voltage<br />
control adopts conventional PI controller. This method<br />
keeps the fast response merit. Real current can follow<br />
reference current in one switching period and switching<br />
frequency keeps constant .In addition ,parameters<br />
selection is simple for there is only one PI controller in<br />
the system .<br />
A. Principle <strong>of</strong> Predictive Digital Current Control<br />
Write mathematical model <strong>of</strong> three-phase PWM<br />
rectifier in three-dimensional stationary (a-b-c) frame as<br />
eα<br />
eβ<br />
iβ<br />
iα *<br />
uαN *<br />
uβN *<br />
I amp<br />
Figure 2. Predictive current control configuration <strong>of</strong> three-phase PWM<br />
rectifier based on (α-β) stationary frame.<br />
follow .<br />
⎧<br />
⎪U<br />
⎪<br />
⎨U<br />
⎪<br />
⎪<br />
U<br />
⎪⎩<br />
AN<br />
BN<br />
CN<br />
= e<br />
a<br />
= e<br />
b<br />
= e<br />
c<br />
dia<br />
− ( L + Ria<br />
)<br />
dt<br />
dib<br />
− ( L + Rib<br />
)<br />
dt<br />
dic<br />
− ( L + Ric<br />
)<br />
dt<br />
The mathematic model in the two-dimensional<br />
stationary (α-β) frame can be obtained by applying the<br />
following α-β transformation as seen in equation (8).<br />
Vdc<br />
*<br />
Vdc<br />
(7)<br />
⎡ 1 1 ⎤<br />
⎢<br />
1 − −<br />
2<br />
⎥<br />
T =<br />
2 2<br />
αβ / abc ⎢<br />
⎥ (8)<br />
3 ⎢<br />
3 3<br />
0 − ⎥<br />
⎢⎣<br />
2 2 ⎥⎦<br />
Expression (9) is the mathematic model in stationary<br />
(α-β) frame:<br />
⎧<br />
⎪U<br />
⎨<br />
⎪U<br />
⎪⎩<br />
αN<br />
βN<br />
= e<br />
= e<br />
α<br />
β<br />
di<br />
− ( L<br />
dt<br />
di<br />
− ( L<br />
dt<br />
α<br />
β<br />
+ Ri<br />
α<br />
+ Ri<br />
β<br />
)<br />
)<br />
(9)
1882 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
Judging from (9), Uα N , U β N are the only variable to<br />
control AC current i α , i β respectively. Average (9) from<br />
t k to k + 1<br />
t , it derives<br />
∫ + 1 tk<br />
1 diα<br />
U α N = eα<br />
− ( L + Riα<br />
) dt<br />
T tk<br />
dt<br />
S<br />
L<br />
R t<br />
e − [ iα<br />
( tk<br />
+ 1)<br />
− iα<br />
( tk<br />
)] − iα<br />
dt<br />
T<br />
T tk<br />
= ∫ + k 1<br />
α (10)<br />
S<br />
∫ + 1 tk<br />
1 diβ<br />
U β N = eβ<br />
− ( L + Riβ<br />
) dt<br />
T tk<br />
dt<br />
S<br />
L<br />
R t<br />
e − [ iβ<br />
( tk<br />
+ 1)<br />
− iβ<br />
( tk<br />
)] − iβ<br />
dt<br />
T<br />
T tk<br />
= ∫ + k 1<br />
β (11)<br />
S<br />
Here, ( U α N , U β N ) , ( e α , e β ) stand for<br />
average value <strong>of</strong>( Uα N , U β N ),( e α , e β )in one<br />
control period respectively. S<br />
T = k + 1<br />
S<br />
S<br />
t - t k . Assume U αN<br />
, U β N are the same with the reference voltage<br />
U α , U β in each period and omit R, it can derive<br />
*<br />
N<br />
*<br />
N<br />
⎧<br />
⎪<br />
U<br />
⎨<br />
⎪U<br />
⎪⎩<br />
*<br />
αN<br />
*<br />
βN<br />
= e<br />
= e<br />
α<br />
β<br />
L<br />
− [ i ( t α<br />
TS<br />
L<br />
− [ i ( t β<br />
T<br />
S<br />
k + 1<br />
k + 1<br />
) − i ( t<br />
α<br />
) − i ( t<br />
β<br />
k<br />
k<br />
)]<br />
)]<br />
(12)<br />
Assume grid current can track reference current in one<br />
period that means iα t ) =<br />
*<br />
iα t ) and<br />
( k + 1<br />
( k + 1<br />
*<br />
iβ t ) = iβ ( t + 1)<br />
, thus (13) can be written as<br />
( k + 1<br />
k<br />
⎧<br />
⎪<br />
U<br />
⎨<br />
⎪U<br />
⎪⎩<br />
*<br />
αN<br />
*<br />
βN<br />
= e<br />
= e<br />
α<br />
β<br />
L<br />
− [ i<br />
TS<br />
L<br />
− [ i<br />
T<br />
S<br />
α<br />
β<br />
*<br />
*<br />
( t<br />
( t<br />
k+<br />
1<br />
k+<br />
1<br />
) − i ( t<br />
α<br />
k<br />
) − i ( t<br />
β<br />
k<br />
)<br />
)<br />
(13)<br />
From the expression (13) ,we can see that the same<br />
variables have been decoupled .And the variable<br />
* *<br />
( U α N , U β N ) is the reference voltage vector U ref which<br />
is the input value <strong>of</strong> the SVPWM algorithm.<br />
B. SVPWM ALGORITHM<br />
As shown in Figure 1 ,there are eight possible<br />
combinations <strong>of</strong> on and <strong>of</strong>f states <strong>of</strong> the upper power<br />
transistors .So there are six kinds <strong>of</strong> active state ,i.e. ,nonzero<br />
vectors ,and two kinds <strong>of</strong> zero vector ( U and 000<br />
© 2011 ACADEMY PUBLISHER<br />
U ) .The eight basic space vectors defined by the<br />
111<br />
combination <strong>of</strong> the switches are shown in Figure 3 .In<br />
order to make the input current phase track the input<br />
voltage phase ,and keep output DC voltage constant ,the<br />
SVPWM technique is used to approximate the reference<br />
voltage vector U ref .The following presents how to use<br />
β<br />
*<br />
U βN<br />
Figure 3. Basic space vectors.<br />
d<br />
U0<br />
U<br />
0<br />
U60(110)<br />
°<br />
60<br />
d<br />
U60<br />
U<br />
60<br />
*<br />
U αN<br />
U ref<br />
α<br />
U0(100)<br />
Figure 4. Projection <strong>of</strong> the Reference Voltage Vector in Sector Ⅰ.<br />
the fast algorithm to obtain the SVPWM signal according<br />
to the reference voltage vector based on DSP .<br />
(a). Determination <strong>of</strong> the sector<br />
*<br />
*<br />
U and U are converted to a balanced three-phase<br />
α β<br />
N<br />
N<br />
quantities V ref 1 , V ref 2 and ref 3<br />
V according to the<br />
following inverse CLARKE transformation :<br />
⎧<br />
⎪V<br />
⎪<br />
⎨V<br />
⎪<br />
⎪<br />
⎪V<br />
⎩<br />
ref 1<br />
ref 2<br />
ref 3<br />
= u<br />
rβ<br />
− u<br />
=<br />
− u<br />
=<br />
rβ<br />
rβ<br />
+<br />
2<br />
− 3 * u<br />
2<br />
3 *<br />
u<br />
rα<br />
rα<br />
(14)<br />
From (14), the following decisions can be made on the<br />
variable N information:<br />
V > 0 then a=1, else a=0<br />
If ref 1<br />
If V ref 2 > 0 then b=1, else b=0<br />
If V ref 3 > 0 then c=1, else c=0<br />
The variable N is defined as : N = 4*c + 2*b + a ;
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1883<br />
The sector in which U ref is depends on variable N<br />
.The corresponding relations between variable N and<br />
sector are shown in table Ⅰ.<br />
The Figure 4 show the projection <strong>of</strong> the reference<br />
voltage vector in sector Ⅰ.<br />
TABLE I.<br />
CORRESPONDING RELATIONS BETWEEN VARIABLE N AND<br />
SECTOR<br />
N 1 2 3 4 5 6<br />
Sector Ⅱ Ⅵ Ⅰ Ⅳ Ⅲ Ⅴ<br />
(b) Calculation <strong>of</strong> the durations<br />
*<br />
uα N<br />
*<br />
and uβ N represent the normalized (α , β)<br />
components <strong>of</strong> U ref with respect to the maximum phase<br />
voltage( V dc<br />
3<br />
). We can obtain the duty ratio by the<br />
following formula:<br />
⎡<br />
⎡d<br />
⎤ 0<br />
x ⎢<br />
⎢ ⎥<br />
⎢<br />
⎢ ⎥ 3<br />
= ⎢<br />
⎢<br />
d y ⎥<br />
⎢ 2<br />
⎢ ⎥<br />
⎢ 3<br />
⎢⎣<br />
d ⎥⎦<br />
⎢−<br />
z<br />
⎣ 2<br />
1<br />
⎤<br />
⎥ *<br />
⎥ ⎡u<br />
α N ⎤<br />
1<br />
(15)<br />
⎥ ⎢ ⎥<br />
2 ⎥ ⎢ * ⎥<br />
1 ⎥ ⎣u<br />
βN<br />
⎦<br />
2<br />
⎥<br />
⎦<br />
For different sectors the value for duty ratio<br />
( d , d ) in terms <strong>of</strong> ( d x , d y , d z ) are listed in<br />
U k<br />
Table Ⅱ.<br />
U k + 60<br />
TABLE II.<br />
CORRESPONDING RELATIONS BETWEEN ( d U<br />
, d k U<br />
) AND<br />
k + 60<br />
( d x , d y , d z ) IN DIFFERENT SECTORS<br />
Sector Ⅰ Ⅱ Ⅲ Ⅳ Ⅴ Ⅵ<br />
d - z d z d d x -d x<br />
- d y d y<br />
U k<br />
d U k + 60<br />
d x y<br />
According to the duty ratio, we can get the<br />
corresponding SVPWM signal for the reference voltage<br />
U by DSP .Form the above motioned , this simplified<br />
ref<br />
d -d y<br />
and fast algorithm for SVPWM avoid the nonlinear<br />
operations and improve the calculation speed and<br />
accuracy.<br />
IV. SIMULATION AND EXPERIMENTAL RESULTS<br />
In order to evaluate the PWM rectifier performances,<br />
using the proposed predictive digital current control<br />
system operates with space-vector modulation (SVM),<br />
simulation and prototype model have been carried out<br />
using the following parameters: output power is 900W,<br />
© 2011 ACADEMY PUBLISHER<br />
d z - z d - d x<br />
input phase voltage e =50V , output DC voltage<br />
V dc =150V, equivalent resistance <strong>of</strong> the loop R =0.5Ω,<br />
input inductance L =8mH, DC-link capacitor<br />
C =4700µF, and the switching frequency<br />
S<br />
f =5KHz.<br />
s<br />
Simulation has been completed by MATLAB s<strong>of</strong>tware<br />
.The predictive current control make the input current<br />
following the phase voltage in phase to get unity power<br />
factor .And the PI regulator to keep the output voltage<br />
constant .According to simulation model ,A three-phase<br />
boost-type PWM rectifier with DSP control system is<br />
implemented and tested in the laboratory .In the prototype<br />
model ,DSP micro-controller (TMS320LF2407A) and<br />
IPM (pm100cla60) are employed ;the AD conversion<br />
,CLARKE transformation ,PI regulator , predictive<br />
current controller ,and the fast algorithm for SVPWM are<br />
implemented in the s<strong>of</strong>tware procedures .<br />
Both the simulation and experiment ,we have got the<br />
desired results .A sinusoidal input current in same phase<br />
with the corresponding input phase voltage is obtained as<br />
shown in Figure 5 and Figure 6.We can see that the<br />
power factor is nearly unity .Figure 7 and Figure 8 show<br />
the input current <strong>of</strong> phase A ,B ,C which are phase<br />
separation <strong>of</strong> 2π/3 and the THD
1884 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
Figure 7. The simulation waveforms input current <strong>of</strong> phase A,B,C.<br />
Figure 8. The experimental waveforms <strong>of</strong> input current <strong>of</strong> phase A,B,C<br />
(current : 10A/div).<br />
V. CONCLUSIONS<br />
This paper has presented a economical three-phase<br />
PWM rectifier, which use the predictive digital current<br />
programmed control system. The proposed predictive<br />
current control strategy operates with constant switching<br />
frequency using SVM. The simulation and experimental<br />
results have proved excellent performance <strong>of</strong> the<br />
proposed predictive digital current programmed control<br />
system. Below is the features and advantages:<br />
1) It will reduce the production cost <strong>of</strong> the three-phase<br />
PWM rectifier;<br />
2) It is vital importance in non-polluting and energy<br />
conservation;<br />
3) The PWM rectifier realizes the low harmonic and<br />
unit power factor ;<br />
4) The utilization efficiency <strong>of</strong> DC voltage will be<br />
higher close to 1 ;<br />
REFERENCES<br />
[1] B.R. Lin and T.Y. Yang, “Three-phase AC/DC converter<br />
with high power factor,” IEE Proc.-Electr. Power Appl.,<br />
vol.152, no. 3,pp.757-764, May. 2005.<br />
[2] C.-T. Pan and Y-H. Liao, “Modeling and coordinate<br />
control <strong>of</strong> circulating currents in parallel three-phase boost<br />
rectifiers,” IEEE Trans. Ind. Electron., vol. 54, no. 2, pp.<br />
825–838, Apr. 2007.<br />
[3] Soo-Bin Han, Nam-Sep Choi, and Gyu-Hyeong,<br />
“Modeling and analysis <strong>of</strong> static and dynamic<br />
characteristics for buck-type three-phase PWM rectifier by<br />
circuit DQ transformation,” IEEE Transactions on Power<br />
electronics, vol. 13, no. 2, pp. 323–336, Mar.1998.<br />
© 2011 ACADEMY PUBLISHER<br />
Figure 9. The simulation waveforms <strong>of</strong> output DC voltage.<br />
Figure 10. The experimental waveforms <strong>of</strong> output DC voltage<br />
(DC voltage : 25V/div ).<br />
[4] Yuri Shtessel, Simon Baev, Haik Biglari, “Unity Power<br />
Factor Control in Three-Phase AC/DC Boost Converter<br />
Using Sliding Modes,” IEEE TRANSACTIONS ON<br />
INDUSTRIAL ELECTRONICS, vol. 55, no. 11, pp.3874-<br />
3882,Nov.2008.<br />
[5] Ivo Barbi and Flabio Alberto Bardemaker Batista, “Space<br />
Vector Modulation for Two-Level Unidirectional PWM<br />
Rectifiers,” IEEE TRANSACTIONS ON POWER<br />
ELECTRONICS, vol. 25, no. 1, pp.178-187,Jan. 2010.<br />
[6] T. Jin and K. M. Smedley, “A universal vector controller<br />
for four-quadrant three-phase power converters,” IEEE<br />
Trans. Circuits Syst. I, Reg. Papers, vol. 54, no. 2, pp. 377–<br />
390, Feb. 2007.<br />
[7] Bhim Singh, BrijN.Singh, Ambrish Chandra,Kamal Al-<br />
Haddad, Ashish Pandey, Dwarka P. Kothari, “A Review <strong>of</strong><br />
Three-Phase Improved Power Quality AC–DC<br />
Converters,” IEEE Transactions on Power electronics, vol.<br />
51, no. 3, pp.641-660,Jun.2004.<br />
[8] Ana Vladan Stankovic and Ke Chen, “A New Control<br />
Method for Input–Output Harmonic Elimination <strong>of</strong> the<br />
PWM Boost-Type Rectifier Under Extreme Unbalanced<br />
Operating Conditions,” IEEE TRANSACTIONS ON<br />
INDUSTRIAL ELECTRONICS, vol.56, no.7, pp.2420-<br />
2430,Jul .2009.<br />
[9] Sergio Vazquez, Juan Antonio Sanchez, Juan Manuel<br />
Carrasco, Jose Ignacio Leon, and Eduardo Galvan, “A<br />
Model-Based Direct Power Control for Three-Phase Power<br />
Converters,” IEEE TRANSACTIONS ON INDUSTRIAL<br />
ELECTRONICS, vol.55, No.4,pp1647-1657, Apr. 2008.<br />
[10] Yongsug Suh and Thomas A. Lipo, “Control Scheme in<br />
Hybrid Synchronous Stationary Frame for PWM AC/DC<br />
Converter Under Generalized Unbalanced Operating<br />
Conditions,” IEEE TRANSACTIONS ON INDUSTRY
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1885<br />
APPLICATIONS, vol. 42, no. 3,pp.825-835, May/Jun.<br />
2006.<br />
[11] Wang Jiuhe, Li Huade, Wang Liming, “Direct Power<br />
Control System <strong>of</strong> Three Phase Boost Type PWM<br />
Rectifiers” Proceedings <strong>of</strong> the CSEE, vol.26 ,no.18,pp.54-<br />
60, Sep. 2006.<br />
Zhongjiu Zheng was born in<br />
Heilongjiang, China, in 1981. He<br />
received the B.S. degree in Information<br />
and Communication Engineering from<br />
Dalian University <strong>of</strong> Technology<br />
(DUT), Dalian, China, in 2003, where<br />
he is currently working toward the<br />
combined M.S./Ph.D. degrees in the<br />
area <strong>of</strong> Electrical and Electronics<br />
Engineering in DUT.<br />
His research interests include three-phase power factor<br />
correction, digital control <strong>of</strong> switching power converters, power<br />
converter topologies, and uninterrupted power supply system.<br />
Gu<strong>of</strong>eng Li was born in Heilongjiang,<br />
China, in 1968. He received the B.S.<br />
degree in Physics from Harbin Normal<br />
University, Harbin, China, in 1990, the<br />
M.S degree from Northeast Normal<br />
University, Shenyang, China, in 1993,<br />
and the Ph.D. degree from Dalian<br />
University <strong>of</strong> Technology, Dalian,<br />
China, in 2000.<br />
From 1993 to 1997, he was a lecturer<br />
with the Physics Department, Northeast Normal University.<br />
Since 2000, he was a lecturer in Dalian University <strong>of</strong><br />
Technology. Currently, he is a Pr<strong>of</strong>essor and Director <strong>of</strong> the<br />
Special Power Supplies Research Institute, Dalian University <strong>of</strong><br />
Technology. His research interests include special power<br />
supply, environmental engineering, static electricity, and ship<br />
electrified transmission automation.<br />
Ninghui Wang was born in Jilin, China,<br />
in 1954. He received the B.S. degree in<br />
Physics from Northeast Normal<br />
University, Shenyang, China, in 1981.<br />
From 1972 to 1991,he was a engineer in<br />
Northeast Normal University. Since<br />
1991, he works in Dalian University <strong>of</strong><br />
Technology (DUT). Currently, he is a<br />
Pr<strong>of</strong>essor <strong>of</strong> DUT, and Standing<br />
Director <strong>of</strong> China Power Supply Society.<br />
His research interests include Theory and new technology <strong>of</strong><br />
electrical engineering, mechano-electronic, and preparation <strong>of</strong><br />
magnesium oxide.<br />
© 2011 ACADEMY PUBLISHER
1886 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
Supply Chain Coordination under Return Policy<br />
with Asymmetric Information about Cost <strong>of</strong><br />
Reverse Logistics Operations<br />
Ting Long Zhang<br />
Institute <strong>of</strong> Economics and Management/Anhui Normal University, WuHu, China<br />
saztl@mail.ustc.edu.cn<br />
Abstract—In this paper, we study return policy and supply<br />
chain coordination in a channel <strong>of</strong> one supplier and one<br />
retailer. The paper assumes that unsold merchandise should<br />
been refunded to the supplier by the retailer. The retailer<br />
knows the cost <strong>of</strong> reverse logistics operations but the<br />
supplier has to estimate it. The contract menu under<br />
asymmetric reverse logistics cost information between<br />
supply chain members was designed and discussed. The goal<br />
<strong>of</strong> the supplier’s contract is to coordinate the channel and<br />
then get more pr<strong>of</strong>it. The problem is analyzed as a<br />
Stackelberg game in which the supplier declares a contract<br />
menu with return price and wholesale price to the retailer<br />
and requires the retailer report the cost <strong>of</strong> reverse logistics.<br />
Then the retailer reports the cost and gets the<br />
corresponding contract. The optimal solutions <strong>of</strong> the<br />
contract menu are derived, and numerical examples are<br />
presented to illustrate the properties <strong>of</strong> the contract menu.<br />
Index Terms—-supply chain; return policy; reverse<br />
logistics; asymmetric information<br />
I. INTRODUCTION<br />
Because <strong>of</strong> the existence <strong>of</strong> multiple decision makers<br />
in supply chain, the decisions that are locally optimal can<br />
be globally inefficient. It is well documented in marketing<br />
and economics literature that uncoordinated decisions lead<br />
to “double marginalization”, which is one <strong>of</strong> the causes <strong>of</strong><br />
channel inefficiency [1],[2]. Coordination among<br />
suppliers and retailers is a very important strategic issue in<br />
supply chain management. Coordination between<br />
independent firms in a supply chain relationship has<br />
gained much attention recently and many studies have<br />
been presented. In order to provide compatible incentives<br />
to improve the supply chain performance and achieve the<br />
win-win solution, some types <strong>of</strong> supply chain contracts<br />
have been discussed. For instance, see return policies [3],<br />
revenue sharing[4], quantity discount [5], quantity<br />
flexibility [6], sales rebate[7]. See[8]for excellent reviews.<br />
The goal <strong>of</strong> these contracts is to coordinate supply chain,<br />
which means that the total pr<strong>of</strong>it <strong>of</strong> the decentralized<br />
supply chain will be equal to that achieved under a<br />
centralized system.<br />
In this paper, the focus is on combined contract <strong>of</strong><br />
wholesale price and return policy. Wholesale price is a<br />
fundamental decision for supply chain coordination in<br />
distribution channel. The “Quantity discount” is a popular<br />
© 2011 ACADEMY PUBLISHER<br />
doi:10.4304/jcp.6.9.1886-1890<br />
method used to stimulate the retailer to order [9]. [10]<br />
shows that in complex supply chain linear quantity<br />
discount alone cannot coordinate supply chain.<br />
The return policy is also called as buyback policy for<br />
many cases and in many researches. This is<br />
understandable since almost all return policies incur<br />
buyback price. That is why vast researches pay more<br />
attention to buyback price. [3]demonstrates that a policy<br />
to <strong>of</strong>fer full credit to the buyer for a partial return <strong>of</strong> goods<br />
may achieve channel coordination and the supplier can get<br />
any percentage <strong>of</strong> channel pr<strong>of</strong>it by setting proper<br />
wholesale price and buyback price. There are many<br />
restrictions about setting. For example, when the retail<br />
price which affects is endogenous, the buyback contract<br />
no longer coordinates the supply chain [8]. For more<br />
complex setting, the other contract will with buyback to<br />
improve the performance[7],[10].<br />
Though both buyback policy and return policy decide<br />
buyback price, there are remarkable difference between<br />
them. That is, return policy incurs reverse logistics, but<br />
buyback policy may not incur reverse logistics. Most<br />
studies about buyback policy or return policy do not<br />
consider reverse logistics. Recently, more attention is<br />
devoted to the logistics <strong>of</strong> return policies. [11] shows that<br />
the development <strong>of</strong> e-commerce in electronic market<br />
increases the value <strong>of</strong> surplus products. [12] reviews<br />
return contract and illustrates the decision <strong>of</strong> contract<br />
when consider cost <strong>of</strong> the return good. [13] investigates a<br />
supply chain consider the forward logistics and reverse<br />
logistics simultaneously.<br />
Most studies to date on return policy have assumed that<br />
the salvage value is same for all member <strong>of</strong> supply chain.<br />
But for some products, such as electronic products and<br />
books, the excess products should been reprocessed or<br />
delivered to another substitute channel if resend them to<br />
the supplier. Therefore, the value <strong>of</strong> excess goods may be<br />
higher for the supplier than for the retailer. This is a<br />
foundation assumption for return policy. On the other<br />
hand, the paper investigating buyback contract considers<br />
the logistics is rare. Furthermore, the majority <strong>of</strong> supply<br />
chain coordination researches assume a symmetric<br />
information situation. Because <strong>of</strong> the variety and<br />
complexity <strong>of</strong> logistics activities, the accurate cost<br />
accounting is troublesome. It is difficult to estimate the<br />
expenses <strong>of</strong> the returned purchase when the return
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1887<br />
activities are managed by the retailer. This paper considers<br />
an asymmetric information about the cost <strong>of</strong> returned<br />
goods.<br />
The paper proceeds as follows. The next section<br />
presents the assumptions and notations. In Section 3, the<br />
integrated model is discussed firstly. In Section 4, the<br />
return policy under symmetric information situation is<br />
investigated. Section 5 focuses on the return policy for an<br />
asymmetric information relationship. Section 6 gives the<br />
numerical analysis. Section 7 summarizes the findings.<br />
II. ASSUMPTIONS AND NOTATIONS<br />
The demand D is a random within [0, b ] . We denote<br />
by f, F , µ the density function, distribution function<br />
y<br />
<strong>of</strong> D , respectively. Let E( y) = ∫ xf( x) dx . The retail<br />
−∞<br />
price p and the supplier cost c are exogenous variable<br />
and the wholesale price <strong>of</strong> the supplier w is endogenous<br />
variable. The salvage values <strong>of</strong> the supplier and the<br />
retailer are different and denoted by m v and v r ,<br />
respectively. In this paper, we assume the retailer takes<br />
back work and pays the logistic cost, denoted by l .<br />
If vr ≥vm − l , from the supply chain point <strong>of</strong> view,<br />
returning goods is unreasonable. This paper<br />
assumes vr < vm −l and considers an asymmetric<br />
information about the cost <strong>of</strong> reverse logistics. We<br />
assume the real value <strong>of</strong> l is the retailer’s private<br />
knowledge and we call this retailer l -type retailer for<br />
convenience in presentation. The supplier does not make<br />
sure the type <strong>of</strong> the retailer, but he deems the value <strong>of</strong> l is<br />
either l with probability <strong>of</strong> ρ or l with probability<br />
<strong>of</strong>1− ρ . The buyback contract is a practical method for<br />
the supplier to share risks and losses <strong>of</strong> the retailer. We<br />
denote r as the buyback price, which is the decision<br />
variable <strong>of</strong> the supplier as well as w . In asymmetric<br />
information situation, the supplier should <strong>of</strong>fer retailers a<br />
menu <strong>of</strong> returns policies trading <strong>of</strong>f l -type retailer<br />
with l -type retailer. The one goal <strong>of</strong> the supplier’s<br />
contract is to coordinate the supply chain and the other is<br />
to maximize the supplier pr<strong>of</strong>it.<br />
Let l -type retailer’s ordering size is Ql () , the<br />
expected surplus and sale are Ol () and Sl () . Simply<br />
calculating gives<br />
Ql ()<br />
Ol () = ∫ F( xdx ) , Sl () = Ql () −Ol<br />
() (1)<br />
0<br />
The total expected pr<strong>of</strong>it <strong>of</strong> the channel is<br />
∏ m+ r() l = ( p−c) Q() l −( p− vm+ l) O() l (2)<br />
The pr<strong>of</strong>it <strong>of</strong> l -type retailer is<br />
∏ r () l = ( p−w) Q() l −( p− r+ l) O() l<br />
(3)<br />
The pr<strong>of</strong>it <strong>of</strong> the supplier is<br />
∏ () l = ( w−c) Q() l −( r− v ) O() l<br />
(4)<br />
m m<br />
III. THE INTEGRATED MODEL<br />
The goal <strong>of</strong> this paper is to develop a return policy to<br />
coordinate the supply chain. The coordination <strong>of</strong> supply<br />
© 2011 ACADEMY PUBLISHER<br />
chain means the decision in decentralized enable the<br />
channel to obtain the same pr<strong>of</strong>its as a vertically<br />
integrated firm’s. In order to give a benchmark for follows<br />
discussion, in this section, we first focus on an integrated<br />
structure in which both the supplier and the retailer agree<br />
to take decisions to maximize the total channel pr<strong>of</strong>its<br />
(joint pr<strong>of</strong>it maximization).<br />
We denote the optimal order size and the maximum<br />
expected pr<strong>of</strong>it <strong>of</strong> the channel by * *<br />
Q () l , m r() l ∏ + . Using<br />
Leibniz’s rule to obtain the first and second derivatives<br />
shows that m r() l ∏ + is concave. The sufficient optimality<br />
condition is the well-known formula:<br />
*<br />
F( Q ( l)) = ( p− c)/( p+ l− vm).<br />
(5)<br />
Using the relationship<br />
Q<br />
∞<br />
xf( x) dx = µ − xf( x) dx<br />
∫ ∫<br />
0<br />
and substituting from (5) into (2) and simplifying gives<br />
the optimal expected pr<strong>of</strong>it:<br />
* *<br />
∏ m+ r() l = ( p+ l− vm) E( Q ()). l<br />
(6)<br />
* *<br />
Proposition 1. Q () l and m r() l ∏ + decrease as l increases<br />
*<br />
Pro<strong>of</strong>. From (5), we have ∂Q ()/ l ∂ l < 0.<br />
Taking the first-<br />
*<br />
order derivative <strong>of</strong> m r() l ∏ + , one has<br />
* * * *<br />
∂ ∏m+ r()/ l ∂ l = E( Q ()) l − Q () l F( Q ()) l < 0.<br />
The higher l means the higher the cost, thus this<br />
conclusion is intuitional.<br />
For the convenience in presentation is follows<br />
subsections, let<br />
*<br />
Q () l<br />
*<br />
O () l = ∫ F( x) dx.<br />
(7)<br />
0<br />
IV. THE RETURN POLICY UNDER SYMMETRIC<br />
INFORMATION SITUATION<br />
For the sake <strong>of</strong> comparing, before investigate the<br />
asymmetric information situation, now we discuss the<br />
problem <strong>of</strong> channel coordination by return policy with<br />
common knowledge about l . When the supplier know<br />
the retailer’s cost l , the supplier first declares the<br />
wholesale price w and buyback price r . The retailer, as<br />
s<br />
the follower sets the decision <strong>of</strong> ordering size Q () l . It is<br />
straightforward to find that only if r > vm+ l , then the<br />
retailer sends back the excess goods.<br />
Using the same method gives<br />
s<br />
F( Q ( l)) = ( p− w)/( p+ l− r).<br />
(8)<br />
s<br />
l -type retailer’s expected pr<strong>of</strong>it, denoted by r () l ∏ , is<br />
s s<br />
∏ r ( l) = ( p+ l− r) E( Q ( l))<br />
(9)<br />
From (5) and (8), we get the observation as in Proposition<br />
2.<br />
Proposition 2. If wrsatisfy ,<br />
w= βc+ (1 − β) p, r = (1 − β)( p+ l) + βvm<br />
, (10)<br />
where l /( l + c −vm) ≤β≤1 the combined contract <strong>of</strong> wholesale price and buyback<br />
policy can coordinate the supply chain and has the<br />
follows properties:<br />
Q
1888 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
(i) vm≤r ≤ w,<br />
∂r/ ∂ w> 0, ∂r/ ∂ vm> 0, ∂r/ ∂ l > 0;<br />
s *<br />
(ii) ∏ () l = β∏ s<br />
*<br />
(), l ∏ () l = (1 −β) ∏ (); l<br />
r m+ r m m+ r<br />
(iii) 0≤1 −β≤( c− vm)/( l+ c− vm)<br />
Proposition 2 shows that when the supplier takes the<br />
contract (10) to coordinate the supply chain, the buyback<br />
price the supplier <strong>of</strong>fers is higher than his savage value.<br />
Furthermore, the buyback price should increase with the<br />
increase <strong>of</strong> wholesale price, the supplier’s savage value<br />
and the retailer’s logistics cost. If this buyback does not<br />
incur reverse transportation, the dominant supplier can<br />
get any percentage <strong>of</strong> the channel pr<strong>of</strong>it. However, when<br />
the retailer return the excess products, the logistics cost<br />
will impose restrictions on the supplier’s pr<strong>of</strong>it, and the<br />
higher l leads to the lower pr<strong>of</strong>it.<br />
V. THE RETURN POLICY UNDER ASYMMETRIC<br />
INFORMATION SITUATION<br />
Assume now the supplier cannot confirm the retailer<br />
is l -type or l -type. That is, the supplier has only the<br />
retailer’s reported cost value, denoted by l f . There is no<br />
reason to assume that the retailer will report honestly<br />
“ l f ” as the same as the real value. In this section, we<br />
discuss how to develop a policy to coordinate the supply<br />
chain. The key factor for success in coordinating the<br />
supply chain is to make the l -type retailer report<br />
honestly, i.e. lf= l Accordance with “the revelation<br />
principle”[14], the contract should satisfies “incentive<br />
compatibility constraint” and “participate constraint”,<br />
simultaneously.<br />
From Proposition 2, consider the contract<br />
* *<br />
menu{(<br />
w, r, Q ( l)),( w, r, Q ( l ))} :<br />
w= βc+ (1 − β) p, r = (1 − β)( p+ l) + βvm,<br />
where l /( l − vm + c)<br />
≤ β ≤1<br />
. (11)<br />
w= βc+ (1 − β) p, r = (1 − β)( p+ l) + βv<br />
where l /( l − vm + c)<br />
≤ β ≤1<br />
The supplier declares the contract menu and requires the<br />
retailer report her cost <strong>of</strong> logistics l . He should <strong>of</strong>fer l -<br />
*<br />
type retailer the contract ( wrQ , , ( l )) and l -type retailer<br />
the contract<br />
that<br />
by<br />
*<br />
( wrQ , , ( l )) . It is straightforward to find<br />
* *<br />
{( w, r, Q ( l)),( w, r, Q ( l))} can be substituted<br />
* *<br />
{( , Q ( l)),( , Q ( l))}<br />
β β . On base <strong>of</strong> analysis in<br />
subsection 4, we get the retailer’s pr<strong>of</strong>it.<br />
For l -type retailer, if she reports honestly, i.e. lf= l,<br />
as<br />
then her pr<strong>of</strong>it, denoted by ∏ ( β,<br />
l)<br />
, is<br />
r<br />
as<br />
*<br />
r m r<br />
∏ ( β, l) = β ∏ + ( l)<br />
; (12)<br />
if she reports dishonestly, i.e. lf= l , then her pr<strong>of</strong>it,<br />
as<br />
denoted by ∏ ( β,<br />
l)<br />
, is<br />
© 2011 ACADEMY PUBLISHER<br />
r<br />
∏ = ∏ +∆ . (13)<br />
as<br />
* *<br />
r ( β, l) β m+ r(<br />
l) lO ( l)<br />
m<br />
For l - type retailer, if she reports honestly, i.e. lf= l ,<br />
as<br />
then her pr<strong>of</strong>it, denoted by ∏ ( β,<br />
l)<br />
, is<br />
r<br />
as<br />
*<br />
r ( , l) m r(<br />
l)<br />
∏ β = β ∏ + , (14)<br />
if she reports dishonestly, i.e. lf as<br />
denoted by ∏ ( β,<br />
l)<br />
, is<br />
= l , then her pr<strong>of</strong>it,<br />
r<br />
as<br />
* *<br />
∏ r ( β, l) = β ∏m+ r(<br />
l) −∆ lO ( l).<br />
(15)<br />
The incentive compatibility constraint is<br />
as as as as<br />
∏r ( β, l) −∏ r ( β, l) > 0, ∏r ( β, l) −∏ r ( β,<br />
l)<br />
> 0.<br />
(16)<br />
Let<br />
*<br />
∏m+ r()<br />
l ∆lO() l ∆lO()<br />
l<br />
all (,) = , bll (,) = , cll (,) = .<br />
* * *<br />
∏m+ r() l ∏m+ r() l ∏m+<br />
r()<br />
l<br />
(17)<br />
The condition <strong>of</strong> (16) is same as<br />
βall (,) + bll (,) ≥ β ≥ βall<br />
(,) + cll (,) . (18)<br />
We assume that the reserved pr<strong>of</strong>it <strong>of</strong> the retailer is the<br />
pr<strong>of</strong>it which she can get in decentralized setting without<br />
the supplier’s buyback policy. In this model, the supplier<br />
set the optimal wholesale price which maximizes his<br />
0<br />
pr<strong>of</strong>it. We denote the retailer pr<strong>of</strong>it by ∏ r . Hence, the<br />
participate constraint is<br />
as<br />
0 as<br />
0<br />
∏r ( β,<br />
l)<br />
≥∏r and ∏r ( β , l)<br />
≥∏ r . (19)<br />
From (11) and (19), let<br />
0 ⎧ ∏r<br />
l ⎫<br />
β = Max ,<br />
0 ⎨ *<br />
⎬<br />
⎩∏m+ r()<br />
l l+ c−vm⎭ . (20)<br />
0 ⎧⎪ ∏r<br />
l ⎫⎪<br />
β 0 = Max ⎨ ,<br />
*<br />
⎬<br />
⎩⎪∏ m+ r() l l+ c−vm⎭⎪ Therefore, the constraint condition <strong>of</strong> (11) and (19) can<br />
been simplified as<br />
0 0 , β ≥ β β ≥ β . (21)<br />
Let<br />
β = all (,) β<br />
1 0 + cll (,), β = all (,) β<br />
2<br />
0 + bll (,) . (22)<br />
After making clear the conditions which ensures the<br />
retailer report honestly and achieves the coordination <strong>of</strong><br />
the supply chain, we now discuss the supplier’s decisions<br />
<strong>of</strong> β, β which optimizes the supplier’s pr<strong>of</strong>it. The<br />
problem <strong>of</strong> the supplier is<br />
as<br />
* *<br />
Max ∏ m = ρ(1 −β) ∏ m+ r( l) + ( 1 −ρ) (1 −β) ∏m+<br />
r(<br />
l)<br />
. (23)<br />
s..(19) t and (21)<br />
0≤β≤ 1 and l ≥ l,<br />
then<br />
Lemma. If 0<br />
* * *<br />
m+ r m+ r<br />
∏ () l +∆lO () l ≤∏ () l and β ≤ 1<br />
2<br />
Pro<strong>of</strong>. Given l , the derivative <strong>of</strong><br />
*<br />
∏<br />
* *<br />
() l +∆lO () l ≤∏ () l with respect to l is<br />
m+ r m+ r<br />
[<br />
*<br />
m+ r( l) *<br />
lO ( l)]/ l<br />
* * * * * *<br />
∂∏ +∆ ∂<br />
= E( Q ()) l −Q () l F( Q ()) l −[ E( Q ()) l −Q<br />
() l F( Q ())] l
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1889<br />
and the derivative <strong>of</strong><br />
respect to l is<br />
* * *<br />
EQ ( ( l)) − Q( lFQ ) ( ( l))<br />
with<br />
* * * *<br />
∂[ E( Q ()) l −Q () l F( Q ())]/ l ∂ l =− F( Q ()) l < 0.<br />
Thus,<br />
∂∏ +∆ ∂ ≤ , which means<br />
* *<br />
[ m+ r( l) lO ( l)]/ l 0<br />
∏ () l +∆lO() l ≤∏ () l<br />
.Because<br />
* *<br />
m+ r m+ r<br />
* * *<br />
β all (,) bll (,) [ () ()]/ ()<br />
2<br />
m+ rl lO l m+ rl<br />
2<br />
≤ + < ∏ +∆ ∏ , so<br />
β ≤ 1 .<br />
Eq.(23) shows that the supplier’s problem is a Liner<br />
Program (LP) and all value coefficient are negative. If the<br />
restricts <strong>of</strong> β ≥ β is omitted, the feasible region<br />
0<br />
for β , β is illustrated as in Fig.1. Due to the properties <strong>of</strong><br />
LP, the nearest point to origin <strong>of</strong> coordinates is the<br />
solution <strong>of</strong> (23).<br />
Proposition 3. The solution <strong>of</strong> the supplier’s problem,<br />
* *<br />
denoted by β , β , is<br />
(i) if β ≤ β , then<br />
0 1<br />
* *<br />
= and β = β 0<br />
β β<br />
1<br />
(as in Fig.1),<br />
(ii) if β ≤ β ≤ β , then β = β<br />
1 0 2<br />
0 and β = β 0 (as in<br />
Fig.2),<br />
*<br />
(3) if β < β ≤ all (,) + bll (,) , then β = β<br />
2 0<br />
0 and<br />
*<br />
β = [ β − bll ( , )]/ all ( , ) (as in Fig.3),<br />
0<br />
*<br />
(4) if β > all (,) + bll (,) , then (23) has no feasible<br />
0<br />
solution.<br />
0 *<br />
If ∏r / ∏m+ r( l) ≥ l/( l+ c− vm)<br />
, it is straightforward to get<br />
β () l −T<br />
(,) l l<br />
0<br />
1<br />
0 0 *<br />
∏r ⎪⎧ ∏r ∏m+<br />
r()<br />
l l ⎪⎫<br />
∆lO()<br />
l<br />
= −Max ,<br />
* ⎨ * * ⎬−<br />
*<br />
∏m+ r() l ⎪⎩∏m+ r() l ∏m+ r() l l+ c−v ∏m<br />
r()<br />
l<br />
m ⎪⎭<br />
+<br />
.<br />
0 *<br />
Proposition 4. If ∏ / ∏ ( l) ≥ l/( l+ c− v ) , then<br />
β ≤ β and β = β β = β<br />
0<br />
1<br />
* *<br />
1,<br />
r m+ r m<br />
0<br />
I. NUMERICAL ANALYSIS<br />
In this section, we assume D is uniform distribution<br />
within [ 0,b ] .<br />
By simply operation, one gets<br />
p+ l−vm all (,) =<br />
p+ l+∆l− v<br />
∆l<br />
, bll (,) =<br />
p+ l−v ,<br />
p+ l−vm cll (,) =<br />
( p+ l+∆l−v )<br />
m m<br />
m<br />
2<br />
⎧p+ l−vm l ⎫<br />
β = Max ,<br />
0 ⎨ ⎬<br />
⎩4( p− vr) c+ l−vm ⎭<br />
.<br />
⎧p+ l+∆l− vm l+∆l ⎫<br />
β 0 = Max ⎨ ,<br />
⎬<br />
⎩ 4( p− vr) c+ l+∆l−vm ⎭<br />
© 2011 ACADEMY PUBLISHER<br />
*<br />
β<br />
β<br />
1<br />
β<br />
0<br />
β<br />
β<br />
2<br />
β<br />
1<br />
β<br />
all (,) + bll (,)<br />
β<br />
0<br />
β<br />
2<br />
β<br />
1<br />
C<br />
A<br />
B<br />
β 0<br />
A<br />
C<br />
β 0<br />
β = all (,) β + bll (,)<br />
β = all (,) β + cll (,)<br />
Figure 1. β ≤ β<br />
β = all (,) β + bll (,)<br />
A<br />
C<br />
β 0<br />
0 1<br />
β = all (,) β + cll (,)<br />
Figure 2. β ≤ β ≤ β<br />
1 0 2<br />
β = all (,) β + bll (,)<br />
D<br />
β = all (,) β + cll (,)<br />
β − bll (,)<br />
0 β =<br />
all (,)<br />
Figure 3. β < β ≤ al (,) l + b(,) l l<br />
2 0<br />
The sensitivity analysis <strong>of</strong> β β are listed in Tables 1-<br />
*<br />
3. Tables 1-3 show that β , β increase as ∆l increase.<br />
The greater ∆ l means more uncertainty <strong>of</strong> the supplier<br />
about the retailer’s type and it is disadvantageous for the<br />
* *<br />
supplier. Table 2 shows that β , β decrease as c<br />
increases. The higher cost make the supplier ask for the<br />
higher percentage to ensure enough pr<strong>of</strong>it. The<br />
* ,<br />
*<br />
*<br />
E<br />
1<br />
E<br />
1<br />
E<br />
1<br />
β<br />
β<br />
β
1890 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
higher c and the higher ∆l should decrease the supplier’s<br />
*<br />
pr<strong>of</strong>it. Table 4 shows that β , β increase as p increases,<br />
this is easy to understand because the higher retailer price<br />
leads to the higher reserved pr<strong>of</strong>it.<br />
(i) c = 40, p = 80, v = 25, v = v −∆ v, l = 2, l = l+∆ l<br />
*<br />
*<br />
*<br />
m r m<br />
Table 1. β , β vary with ∆v and ∆ l<br />
∆ l<br />
*<br />
β<br />
*<br />
β<br />
*<br />
β<br />
*<br />
β<br />
*<br />
β<br />
∆ v = 12<br />
∆ v = 16<br />
∆ v = 20<br />
2 0.225 0.214 0.219 0.210 0.219 0.210<br />
4 0.282 0.285 0.282 0.285 0.282 0.285<br />
6 0.329 0.347 0.329 0.347 0.329 0.347<br />
8 0.364 0.4 0.364 0.4 0.364 0.4<br />
10 0.390 0.444 0.390 0.444 0.390 0.444<br />
(ii) c = 40,50,60, p = 80, v = 25, v = 10, l = 2, l = l+∆ l<br />
*<br />
*<br />
Table 2. β , β vary with c and ∆ l<br />
∆ l<br />
*<br />
β<br />
*<br />
β<br />
m r<br />
*<br />
β<br />
c = 40<br />
c = 50<br />
c = 60<br />
2 0.219 0.210 0.219 0.210 0.219 0.210<br />
4 0.283 0.285 0.218 0.217 0.218 0.217<br />
6 0.329 0.347 0.233 0.242 0.217 0.225<br />
8 0.364 0.4 0.264 0.285 0.217 0.232<br />
10 0.390 0.444 0.288 0.324 0.229 0.255<br />
12 0.410 0.482 0.308 0.358 0.247 0.285<br />
(iii)<br />
c = 40, p = 70,80,90,100, v = 25, v = 10, l = 2, l = l+∆l *<br />
*<br />
Table 3 β , β vary with p and ∆ l<br />
∆ l<br />
*<br />
β<br />
*<br />
β<br />
*<br />
β<br />
*<br />
β<br />
m r<br />
p = 70<br />
p = 80<br />
p = 90<br />
2 0.221 0.210 0.225 0.214 0.226 0.215<br />
4 0.281 0.285 0.281 0.285 0.282 0.285<br />
6 0.325 0.347 0.329 0.347 0.331 0.347<br />
8 0.357 0.4 0.364 0.4 0.369 0.4<br />
10 0.380 0.444 0.390 0.444 0.398 0.444<br />
*<br />
β<br />
II. CONCLUSIONS<br />
This paper has formulated a supply chain coordination<br />
problem with asymmetric information between one<br />
supplier and one retailer for a single-period product. This<br />
paper assumes that the salvage value <strong>of</strong> unsold products<br />
is higher for the supplier than for the retailer. The<br />
supplier wants to coordinate by proper return contract. In<br />
this return policy we assume that the excess goods<br />
refunded by the retailer and the cost <strong>of</strong> reverse logistics is<br />
asymmetric information. This paper formulates a contract<br />
menu with return price and wholesale price. The<br />
observations are developed and show that this contract<br />
menu enables the retailer report the logistics cost honestly<br />
and can achieve the coordination. The solution <strong>of</strong> this<br />
contract menu is derived, and the numerical examples<br />
illustrate that the greater variation <strong>of</strong> the supplier’s<br />
estimation about the logistics cost is disadvantageous for<br />
the supplier. The greater variation will produce more<br />
harm to the supplier who has the higher cost.<br />
© 2011 ACADEMY PUBLISHER<br />
*<br />
β<br />
*<br />
β<br />
*<br />
β<br />
*<br />
β<br />
*<br />
β<br />
There a number <strong>of</strong> possible extensions <strong>of</strong> our study<br />
that can constitute future research endeavors in this area.<br />
One immediate extension is to consider the cooperating<br />
reverse logistics between the channel members.<br />
Developing better contract menu to deal with the<br />
asymmetric information is another interesting theme.<br />
ACKNOWLEDGMENT<br />
The authors would like to acknowledge the supports <strong>of</strong><br />
research Grants from National Natural Science<br />
Foundation <strong>of</strong> China (Project No. 70901001) and Anhui<br />
Provincial Natural Science Foundation (Project No.<br />
090416244).<br />
REFERENCES<br />
[1] Spengler, J., 1950. Vertical integration and anti-trust<br />
policy. <strong>Journal</strong> <strong>of</strong> Political Economy 58, 347-552.<br />
[2] Tirole, J., 1989. The Theory <strong>of</strong> Industrial Organization.<br />
MIT Press, Cambridge MA.<br />
[3] Pasternack, B.A., 1985. Optimal pricing and return policies<br />
for perishable commodities. Marketing Science, 4 (2), 166-<br />
176.<br />
[4] Cachon, G., Lariviere, M., 2005. Supply chain<br />
coordination with revenue sharing: Strengths and<br />
limitations. Management Science, 51 (1), 30-44.<br />
[5] Weng, Z.K., 1995. Channel coordination and quantity<br />
discounts. Management Science, 41 (9), 1509-1521.<br />
[6] Tsay, A.A., 1999. The quantity flexibility contract and<br />
supplier–customer incentives. Management Science, 45<br />
(10), 1339-1358.<br />
[7] Taylor, T.A., 2002. Supply chain coordination under<br />
channel rebates with sales effort effects. Management<br />
Science, 48 (8), 992-1007.<br />
[8] Cachon, G., 2003. Supply chain coordination with<br />
contracts. In: Graves, Steve, de Kok, Ton (Eds.),<br />
Handbooks in Operations Research and Management<br />
Science: Supply Chain Management. North Holland,<br />
Amsterdam.<br />
[9] Munson, C.L., Rosenblatt, M.J., 2001. Coordinating a<br />
three-level supply chain with quantity discounts. IIE<br />
Transactions 33, 371–384.<br />
[10] Krishnan, H., Kapuscinski, R.K., Butz, D.A., 2004.<br />
Coordinating contracts for decentralized supply chain with<br />
retailer promotional effect. Management science, 50 (1),<br />
48-62.<br />
[11] Choi,T.M., Li, D., Yan,H., 2004. Optimal returns policy<br />
for supply chain with e-marketplace. International <strong>Journal</strong><br />
<strong>of</strong> Production Economics, (88), 205-227.<br />
[12] Tsay, A., 2001. Managing retail channel overstock:<br />
markdown money and return policies. <strong>Journal</strong> <strong>of</strong> Retailing<br />
77, 457–492.<br />
[13] Ferguson, M., Jr, V. G., Souza, G., 2004. Supply chain<br />
coordination for false failure returns.Georgia Institute <strong>of</strong><br />
Technology ,Working paper.<br />
[14] Fudenberg, D., Tirole, J., 1991. Game Theory, The MIT<br />
Press, Cambridge, Massachusetts, London, England.<br />
Ting Long Zhang, Ph.D., Associate Pr<strong>of</strong>essor <strong>of</strong> Institute <strong>of</strong><br />
Economics and Management, Anhui Normal University. Field<br />
<strong>of</strong> Research: Management Science, Supply Chain Management,<br />
Logistics.
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1891<br />
Economic Development and Financial Support<br />
for Coal Resource Cities<br />
— A Panel Data Analysis<br />
Zuhuai Yuan<br />
School <strong>of</strong> Management/China University <strong>of</strong> Mining and Technology, Xuzhou, China<br />
E-mail: yzhhn@tom.com<br />
Li Yang<br />
School <strong>of</strong> Management/Anhui University <strong>of</strong> Science & Technology, Huainan, China<br />
E-mail: yangli081003@163.com<br />
Jing Han<br />
Huainan Vocational & Technical College, Huainan, china<br />
E-mail: hanjing623@163.com<br />
Keliang Wang<br />
School <strong>of</strong> Management/Anhui University <strong>of</strong> Science & Technology, Huainan, China<br />
E-mail: klwang@163.com<br />
Abstract — This paper uses measurement methods and<br />
selects relevant indicators from both quantitative and<br />
structural aspects, empirically analyses the relationship<br />
between financial development and economic development<br />
<strong>of</strong> 2000-2008 in more than 18 coal- resourced cities in China.<br />
The results show that financial development takes a<br />
significant role in the economic development <strong>of</strong><br />
coal-resourced cities. However, the high industry<br />
concentration <strong>of</strong> financial resources leads to a decline in<br />
financial resource allocation efficiency.<br />
Index Terms—Resource-based city; Economic development;<br />
Financial support<br />
I. INTRODUCTION<br />
Coal City is an important component <strong>of</strong> the urban<br />
system in China. According to the survey, there are 68<br />
coal cities nationwide, accounting for 38.2% <strong>of</strong> the total<br />
number <strong>of</strong> mining cities, 10.3% <strong>of</strong> the total number <strong>of</strong><br />
cities; supplying for 93.6% <strong>of</strong> the coal to national<br />
economic construction [1]. Relying on coal resources,<br />
coal cities make huge contributions in national<br />
urbanization process, in promoting national economic<br />
development, and expansion <strong>of</strong> social employment.<br />
Meanwhile,a group <strong>of</strong> highly coal-relied cities formed<br />
accordingly. With the depletion and reduction <strong>of</strong><br />
resources, coal cities, like many other mining cities, will<br />
face with the "mine dry up, city fall" threat. Finance is the<br />
core <strong>of</strong> modern economy. Between financial development<br />
and economic development, there is an inherent<br />
mechanism. Coal-based cities can’t develop sustainably<br />
without effective financial support. Many scholars at<br />
home and abroad used empirical analysis methods,<br />
© 2011 ACADEMY PUBLISHER<br />
doi:10.4304/jcp.6.9.1891-1895<br />
verified the quite significant relationship between<br />
financial development and economic development. For<br />
instance, Goldsmith (1969), who started the earliest<br />
quantitative research on financial and economic relations,<br />
found out simultaneous development <strong>of</strong> finance and<br />
economy, a period <strong>of</strong> rapid economic growth always went<br />
along with the ultra level financial development [2]. After<br />
introducing other factors that affect economic<br />
development, King and Levine studied the relevant data<br />
during 1960-1989 <strong>of</strong> 80 countries. It shows that the<br />
financial development and economic growth in a positive<br />
correlation [3]. Han Tingchun uses causality tests on the<br />
relevant data <strong>of</strong> 1981-2002 in China. It shows that from<br />
1981 to 1991 financial development and economic<br />
growth causality is not obvious, however, during<br />
1992-2002, financial development is the direct cause <strong>of</strong><br />
economic growth [4]. Tan Ruyong, after studied by OLS<br />
on the quater data <strong>of</strong> China in 1993-1998, concluded that<br />
China's financial intermediary development and<br />
economic growth have a significant posotive correlation<br />
between each other [5]. Cao Tingqiu and Wang Xihang<br />
studied the panel data <strong>of</strong> 1995-2001 in each region in<br />
Shandong, among sample area, both finance and<br />
economy growth in an obvious trend, the relations among<br />
areas certain difference[6]. The above studies objects are<br />
mainly nations and provinces. According to the data that<br />
the author has, as now, coal-based city development is<br />
also limited to the capital, labor, technology integration<br />
and other traditional elements [7]. Coal-based urban<br />
development studies which put into the relationship<br />
between financial development and economic<br />
development are mostly qualitative and case studies,<br />
empirical research literature are still few. Therefore, the<br />
empirical study <strong>of</strong> economic development <strong>of</strong> coal
1892 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
resources city's financial support issues, has certain<br />
practical and theoretical significance.<br />
This assay will target on 18 prefecture-level coal-based<br />
cities1, select panel data <strong>of</strong> these cities in 2000-2008, use<br />
Eviews5.0 s<strong>of</strong>tware, empirically analyze quantitative and<br />
structural support issues in the economic development <strong>of</strong><br />
coal cities, to provide some ideas for exploring financial<br />
support for the sustainable development <strong>of</strong> coal resource<br />
cities .<br />
The research data come mainly from China's<br />
economic statistical databases and related city<br />
government statistics website.<br />
II. EMPIRICAL ANALYSIS OF FINANCIAL<br />
SUPPORT TO COAL-BASED URBAN<br />
ECONOMIC DEVELOPMENT<br />
A regional financial development matching with local<br />
economic development includes two aspects. In<br />
quantitative aspect, financial element input adjust to<br />
regional economic development requirement, push the<br />
arise <strong>of</strong> gross regional economy. In structural aspect,<br />
internal economic factor delivery structure is compatible<br />
with regional economic structure, which promotes the<br />
adjustment <strong>of</strong> economic structure optimizing and<br />
upgrading. Therefore, this paper will be about empirical<br />
analysis on the relationship between financial<br />
development and economic development <strong>of</strong> coal resource<br />
cities from quantitative and structural support aspects.<br />
A. Quantitative support analysis (quantity effect)<br />
a. Theoretical model construction and variable<br />
description<br />
Economic development is inseparable from human,<br />
finance and material. Production function is the most<br />
common used model in the quantitative study about<br />
economic development. Traditional production function<br />
mainly inspects on the relationship between production<br />
element labor and material input and output variables<br />
among each other. The above literatures have proved the<br />
close relationship between financial development and<br />
economic growth and the regional financial resource<br />
increase can push regional economic development. This<br />
paper uses Cobb-Douglas production function as the<br />
basic model, through the introduction <strong>of</strong> variable<br />
financial scale which influences economic development,<br />
to study the quantitative match relationship between<br />
financial development and economic expansion:<br />
Q=AKαLβFγ (1)<br />
Take on both sides <strong>of</strong> the model number:<br />
LNQ = LNA+αLNK + βLNL+ γLNF (2)<br />
Here, Q for economic development: as the economic<br />
development indicators, the existing papers have chosen<br />
GDP, GDP growth rate or per capita GDP, this article<br />
chose GDP as economic growth rate indicator variables,<br />
and the unit is ten thousand Yuan.<br />
A for integrated function coefficient.<br />
1 It refers to Chifeng, Datong, Huaibei, Hegang, Hebi, Huainan,<br />
Jincheng, Jixi, Jiaozuo, Pingdingshan, Qitaihe, Shuangyashan,<br />
Shuozhou, Wuhai, Xianyang, Yulin, Yangquan, Zaozhuang.<br />
© 2011 ACADEMY PUBLISHER<br />
K for material input: indicate as fixed assets<br />
investment amount in current year with reference to<br />
previous research, the unit is ten thousand Yuan.<br />
L for labor input: taking into account "the number <strong>of</strong><br />
unit employees" statistical coverage limitations and data<br />
sources availability, this paper replaces it by district<br />
population at the end <strong>of</strong> year, unit is ten thousand people.<br />
F for the financial investment: taking into account<br />
data availability, this paper replaces it by the loan balance<br />
<strong>of</strong> regional financial institutions; the unit is ten thousand<br />
Yuan.<br />
α, β and γ each represent the output elasticity <strong>of</strong><br />
capital, labor output elasticity <strong>of</strong> output and financial<br />
flexibility.<br />
b. Econometric model analysis<br />
For the study <strong>of</strong> the overall characteristics and<br />
differences between cities in coal resources city between<br />
financial development and economic development, this<br />
paper takes use <strong>of</strong> Eviews5.0 to establish following<br />
econometric model.<br />
i. Mixed estimated model<br />
Most <strong>of</strong> the coal resource cities passed a similar<br />
development path, therefore they should have the same<br />
characteristic <strong>of</strong> economic and financial relations. By<br />
ordering s<strong>of</strong>tware outputs from Eviews5.0, draw the<br />
conclusion:<br />
LNQ =0.739+0.396LNK + 0.177LNL+ 0.46LNF (3)<br />
(5.31) (21.20) (6.76) (11.63)<br />
2<br />
R2=0.9998 R =0.9998 F=243201 P=0.000 DW=0.58<br />
See from the related indicators <strong>of</strong> output equation, the<br />
estimated model fit good(R 2), the overall equation is<br />
significant(passed the F test), the T value tests <strong>of</strong> single<br />
parameters are satisfactory, the estimate results are<br />
reliable and reasonable from the economic sense. From<br />
the elasticity <strong>of</strong> each variable, during the sample interval<br />
2000 -2008, financial development played a much more<br />
important role,and contributed more to the economic<br />
development <strong>of</strong> coal resource cities. The overall financial<br />
development adapted to the urban economic<br />
development. As the level <strong>of</strong> the financial scale increased<br />
1%, the total regional economy will increase 0.46%.<br />
ii.Phased mix estimated model<br />
Due to the impact by external policy, economic<br />
environment changes and changes in the city’s own stage,<br />
coal resource cities may show different relationship<br />
between economic and financial development in different<br />
phases. Use Eviews5.0 to analyze phases and organize<br />
available:<br />
2000-2001:<br />
LNQ = 1.838+0.11LNK + 0.35LNL+ 0.46LNF (4)<br />
(4.3) (2.05) (3.93) (6.5)<br />
2<br />
R2= 0.9998 R =0.9998 F=266630 P=0.000 DW=0.56<br />
2002-2003:<br />
LNQ =1.426+0.216LNK + 0.302LNL+ 0.46LNF (5)<br />
(5.23) (6.60) (6.76) (8.19)<br />
2<br />
R2=0.9999 R =0.9999 F=156612 P=0.000 DW=0.54
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1893<br />
2004-2005:<br />
LNQ =1.178+0.357LNK + 0.30LNL+ 0.38LNF (6)<br />
(3.28) (5.72) (5.38) (3.89)<br />
R 2= 0.9998 2<br />
R =0.9998 F=74307 P=0.000<br />
DW=1.00<br />
2006-2008:<br />
LNQ = 1.32+0.42LNK + 0.22LNL+ 0.32LNF (7)<br />
(8.89) (14.3) (9.36) (7.62)<br />
R2=0.9999<br />
2<br />
R =0.9999 F=192508 P=0.000 DW=0.78<br />
In various stages, hybrid estimation model fitness, the<br />
overall equation satisfaction and T value test are both<br />
fine. Comparing the above-mentioned equation elements<br />
output elasticity coefficient, we find out: firstly, coal<br />
resources cities, as investment-driven cities, whose<br />
overall features are increasingly strengthen. Fixed assets<br />
investment output elasticity coefficient α increased in the<br />
sample time interval from 0.11 to 0.42, investment in the<br />
development <strong>of</strong> coal resources cities is getting<br />
increasingly important; secondly, input-output elasticity<br />
coefficient <strong>of</strong> labor β negatively developed, the growing<br />
role <strong>of</strong> regional economic development reduced,it is<br />
more in line with the reality that in recent years, cities<br />
step up modernization <strong>of</strong> coal mine construction, mine<br />
reduced capital investment and the actual labor; thirdly,<br />
financial output elasticity coefficient γ fell to 0.32 from<br />
0.46, indicating that financial development <strong>of</strong> coal<br />
resources based on the contribution <strong>of</strong> urban economic<br />
development is declining, the output effect <strong>of</strong> credit funds<br />
is gradually decreased.<br />
iii.Variable coefficient estimation model<br />
In order to find out the difference between coal<br />
resource cities <strong>of</strong> financial development and economic<br />
development, this paper establishes vary coefficient<br />
estimation model for financial development variable.<br />
Relevant output situation ordered by Eviews5.0 is as<br />
Table 1.<br />
Through comprehensive analysis <strong>of</strong> the financial<br />
output elasticity coefficient <strong>of</strong> each city in Table 1, it can<br />
be drawn that during the sample interval, the coal cities<br />
financial development and economic development are<br />
positively correlated in case other factors not considered,<br />
but the contribution varies a certain from level <strong>of</strong> the<br />
financial scale <strong>of</strong> the regional economic development.<br />
Output elasticity in [0.3239 0.3817], with an average<br />
output elasticity <strong>of</strong> 0.3596, the standard deviation <strong>of</strong><br />
output elasticity is 0.014.<br />
Table1: Coal-based city financial input-output coefficients situation<br />
Variable Coefficient Std. Error t-Statistic Prob.<br />
LNA 0.6227 0.8174 0.7618 0.4475<br />
LNK? 0.4224 0.0179 23.6284 0.0000<br />
LNL? 0.4312 0.4152 1.0385 0.3008<br />
Cf 0.3391 0.0672 5.0440 0.0000<br />
Dt 0.3560 0.0593 6.0032 0.0000<br />
Hb 0.3560 0.0563 6.3246 0.0000<br />
Hb 0.3715 0.0524 7.0916 0.0000<br />
Hg 0.3582 0.0497 7.2111 0.0000<br />
Hn 0.3503 0.0552 6.3440 0.0000<br />
Jc 0.3565 0.0548 6.5011 0.0000<br />
Jx 0.3757 0.0537 6.9994 0.0000<br />
Jz 0.3659 0.0619 5.9159 0.0000<br />
Pds 0.3651 0.0683 5.3459 0.0000<br />
Qth 0.3727 0.0511 7.2983 0.0000<br />
Sys 0.3568 0.0515 6.9303 0.0000<br />
Sz 0.3648 0.0523 6.9708 0.0000<br />
Wh 0.3818 0.0561 6.8048 0.0000<br />
Xy 0.3394 0.0686 4.9500 0.0000<br />
Yl 0.3239 0.0631 5.1334 0.0000<br />
Yq 0.3724 0.0504 7.3911 0.0000<br />
Zz 0.3676 0.0630 5.8376 0.0000<br />
R 2 =0.9997<br />
2<br />
R =0.9997 F=31790 P=0.000 DW=1.01<br />
B. Structural support (structural effect)<br />
The transformation <strong>of</strong> resource-based cities is<br />
essentially a process <strong>of</strong> economic restructuring, namely,<br />
an industrial restructuring process. The study result <strong>of</strong><br />
quantitative effect between financial development and<br />
economic development reflects that financial<br />
development has become an important factor in the<br />
development <strong>of</strong> coal resources cities, but in different<br />
years or different cities there are some differences in this<br />
role, that is, the same financial development can not<br />
produce the same total amount <strong>of</strong> economic success, the<br />
financial input-output affects differently. This difference<br />
should be a structural difference, that is, the fit issues<br />
between financial development and economic<br />
development structure.<br />
a. Two phases comparison and analysis<br />
Compare the financial output elasticity coefficient with<br />
industrial structure changes in two stages (2000-2003,<br />
2004-2008) (see Table 2) in coal-resourced cities, known:<br />
during 2004-2008, among the sample cities, the average<br />
proportion <strong>of</strong> secondary industries in other cities except<br />
Jixi, Wuhai has increased compared with 2000-2003. But<br />
along with the increase in the proportion <strong>of</strong> secondary<br />
industry, the financial outputs are not synchronized grow,<br />
the overall trend <strong>of</strong> both is always on reverse, the<br />
proportion <strong>of</strong> changes in the opposite direction was<br />
94.4%.<br />
Table 2: Financial output coefficient and industrial structure diversification in coal resource cities<br />
Cf Dt Hb Hb Hg Hn Jc Jx Jz<br />
coefficient 2000-2003 0.38 0.37 0.35 0.34 0.31 0.36 0.36 0.35 0.39<br />
2004-2008 0.13 0.18 0.19 0.22 0.25 0.18 0.20 0.23 0.17<br />
variation symbol - - - - - - - - -<br />
Sec.<br />
2000-2004 33.9 53.9 50.1 50.1 39.9 48.4 55.2 40.8 52.7<br />
industry proportion<br />
2004-2008 43.1 54.2 56.3 60.5 42.2 55.2 63.9 32.7 63.7<br />
© 2011 ACADEMY PUBLISHER
1894 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
variation symbol + + + + + + + - +<br />
Pds. Qth. Sys. Sz. Wh. Xy. Yl. Yq. Zz.<br />
coefficient 2000-2003 0.41 0.31 0.33 0.33 0.27 0.38 0.33 0.34 0.40<br />
Sec.<br />
industry<br />
proportion<br />
2004-2008 0.15 0.26 0.22 0.23 0.31 0.13 0.15 0.24 0.17<br />
variation symbol - - - - + - - - -<br />
2000-2004 53.4 48.8 38.8 48.9 66.8 42.6 51.4 58.7 51.8<br />
2004-2008 61.6 52.6 42.2 60.0 66.3 45.6 69.2 59.5 63.0<br />
variation symbol + + + + - + + + +<br />
Remarks: Each city's financial output coefficient in the second phase is the result from Eviews5.0. Industrial structure is the<br />
average which secondary industry added-value in various stages accounted for GDP.<br />
b. Allocation efficiency <strong>of</strong> financial industry<br />
For further analysis <strong>of</strong> reverse problems between<br />
financial development and economic growth in coal<br />
resource cities, and also take the data availability into<br />
consideration, this paper selects Huainan2 City as the<br />
typical case to study financial structure and economic<br />
structure relationship, and establishes panel analysis<br />
model based on regional industrial structure and credit<br />
structure in 2000-2008 as a substitute for economic and<br />
financial structure.<br />
i. Development <strong>of</strong> industrial and credit structure<br />
3000000<br />
2500000<br />
2000000<br />
1500000<br />
1000000<br />
500000<br />
0<br />
2000 2001 2002 2003 2004 2005 2006 2007 2008<br />
3500000<br />
3000000<br />
2500000<br />
2000000<br />
1500000<br />
1000000<br />
500000<br />
增加值 第一产业 增加值 第二产业 增加值 第三产业<br />
贷款余额 第一产业 贷款余额 第二产业 贷款余额 第三产业<br />
Figure 1: The added value <strong>of</strong> industries and the loan balance in recent<br />
years in Huainan City<br />
Unit: Ten thousand yuan<br />
In recent years, Huainan City, increasing credit funds<br />
from financial institutions increasingly focused on<br />
coal-based secondary industry whose proportion <strong>of</strong> loans<br />
increased from 70.7% in 2000 to 79.2% in 2008.<br />
Industrial and customer concentration <strong>of</strong> financial<br />
institution loans is quite high3; meanwhile, the status <strong>of</strong><br />
the second pillar industry is consolidated continuously,<br />
and the proportion <strong>of</strong> the added value <strong>of</strong> the secondary<br />
2 As a coal resource-based city built by coal mines, Huainan has a very<br />
rich coal resources <strong>of</strong> 44.4 billion tons vision, proven reserves <strong>of</strong> 15.3<br />
billion tons, accounting for 32% <strong>of</strong> east China.In 2008, coal and power<br />
industries achieved added value <strong>of</strong> 20.9 billion yuan, accounting for the<br />
city's industrial added value, GDP 84.6%, 46.13% respectively.<br />
3 By the end <strong>of</strong> 2008, the year-end value <strong>of</strong> loans was 24.482 billions <strong>of</strong><br />
9 coal mining and power enterprises, including Huainan Mining (Group)<br />
Co., Ltd and Huainan-Zhejiang coal and power Co., Ltd, accounting for<br />
64.38 percent <strong>of</strong> the aggregate value <strong>of</strong> loans, 42.3 percentage points<br />
higher than that in 2001.<br />
© 2011 ACADEMY PUBLISHER<br />
0<br />
industry accounting for the regional GDP rose to 61.1%4<br />
in 2008 from 46.6% in 2000. Regional industrial structure<br />
adjustment pressure increased.<br />
ii. Panel data estimation<br />
The results <strong>of</strong> the Eviews5.0 output <strong>of</strong> coal cities<br />
reached the relationship between credit structure and<br />
industrial structure as the following (Table 3).<br />
Table 3: The relationship between financial structure and industrial<br />
structure in coal resource cities<br />
2000-<br />
2003<br />
2004-<br />
2008<br />
Variable Coefficient Std.Error t-Statistic Prob. Statistics<br />
Y--F 0.1296 0.1938 0.6688 0.5285 R 2 =0.999<br />
E--F 1.1978 0.4170 2.8728 0.0283 F=174815<br />
S--F 0.6296 0.0746 8.4362 0.0002 P=0.000<br />
Fixed<br />
Effects<br />
Y-C=4.750 E-C=-1.209 S-C=2.265 DW=2.83<br />
8<br />
Y--F 0.0920 0.0314 2.9301 0.0168 R 2 =0.999<br />
E--F 0.6823 0.1616 4.2232 0.0022 F=4753<br />
S--F 1.1078 0.4682 2.3659 0.0422 P=0.000<br />
Fixed<br />
Effects<br />
Y-C=5.06 E-C=1.947 S-C=-0.330 DW=2.39<br />
6<br />
Remarks: Y, E and S represent primary, secondary, tertiary<br />
industrial added value logarithm respectively, F Y, FE, FS, each<br />
represents primary, secondary, tertiary industrial credit balance<br />
logarithm at end <strong>of</strong> year in same period. Y-C, E-C, and S-C are<br />
the permanent effects <strong>of</strong> each industrial development.<br />
The relevant data in Table 3 show that with the<br />
financial development, the credit loan output effect in<br />
each industry changed: firstly, the credit loan elasticity<br />
coefficient in primary and secondary industries<br />
respectively declined 0.038 and 0.515, in which the<br />
decline degree <strong>of</strong> credit loan elasticity coefficient <strong>of</strong><br />
tertiary industry obviously exceeds the growth rate <strong>of</strong><br />
tertiary industry. In three industrial credit loan<br />
configuration efficiency, secondary funds configuration<br />
efficiency has been the highest in 2000-2003 dropped to<br />
the second highest in 2004-2008. The decline in the fund<br />
allocation efficiency and the increasing concentration <strong>of</strong><br />
credit loans to coal-based secondary industry, co-led the<br />
4 In 2008, the added value <strong>of</strong> coal and power industries was up to 20.9<br />
billions in Huainan, accounting for 84.6 percent <strong>of</strong> the aggregate<br />
industrial added value and 46.13 percent <strong>of</strong> the regional GDP.
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1895<br />
overall decreasing efficiency <strong>of</strong> the allocation <strong>of</strong> credit<br />
funds.<br />
C. Overall conclusion<br />
The research which analyzes the coal city’s financial<br />
support for economic development from two<br />
perspective--"quantity" and "structure" shows: Under the<br />
circumstance that investment boosts growth is still the<br />
growth mode <strong>of</strong> coal resource cities, financial<br />
development has clear effect and high contribution to<br />
economic development, generally adapted to the<br />
economic development need, but the contribution to<br />
economic presents falling trend. Coal resource city's<br />
economic structure and credit structure show a trend <strong>of</strong><br />
mutual strengthening. Credit funds are increasingly<br />
concentrating to coal-based secondary industry; financial<br />
support for primary and tertiary industries is weakening.<br />
The high industry-concentration and high<br />
customer-concentration features <strong>of</strong> credit loans finally<br />
lead to the financial assets allocation efficiency decline in<br />
coal resources cities.<br />
III. STRATEGIES FOR PROMOTING FINANCIAL<br />
SUPPOTT FOR COAL-BASED CITTIES<br />
SUSTAINABLE DEVELOPMENT<br />
A. Promoting financial development<br />
Establish a sound financing system for city<br />
transformation, vigorously develop local financial<br />
institutions, and actively introduce joint-stock<br />
commercial banks, strengthen the financial viability <strong>of</strong><br />
coal resources city; develop direct financing and give full<br />
play to support resource-based capital market in the role<br />
<strong>of</strong> urban transformation, to encourage support <strong>of</strong> large<br />
coal companies using short-term financing bills, bonds<br />
and notes, reducing large-scale enterprises, enterprises<br />
with financial dependence on bank credit and small and<br />
medium private enterprises <strong>of</strong> the credit squeeze effect;<br />
strengthen the economic and financial information<br />
exchange, and promote political bank-enterprise<br />
communication and collaboration; strengthen the credit<br />
system building, create a favorable sustainable<br />
development <strong>of</strong> coal resources city a good financial<br />
environment.<br />
B. Optimize credit loan structure<br />
In the meantime <strong>of</strong> actively supporting local<br />
characteristic industries and pillar industries, financial<br />
institutions should also increase effective credit loan<br />
release focusing on funds need <strong>of</strong> resource-based<br />
economy transformation follow-up industries. Financial<br />
institutions should reasonably allocate credit resources,<br />
improve the fund input intensity for tertiary industry and<br />
other high credit fund allocation efficiency industries,<br />
improve credit fund application efficiency; increase the<br />
support intensity for recycling economy, bio-medical and<br />
other high-tech industries, avoid further solidity <strong>of</strong><br />
coal-resource cities economy and financial structure<br />
result from excessive credit concentration, and more<br />
difficulties in resource cities transformation; enhance risk<br />
assessment and monitoring, strengthen and improve the<br />
© 2011 ACADEMY PUBLISHER<br />
awareness <strong>of</strong> energy financial risks, to effectively protect<br />
credit fund security in urban transformation.<br />
C. Adjust investment structure<br />
Single industrial structure is a major issue which<br />
resource cities will face in its development, for which<br />
resource cities are inevitably face "resource-curse".<br />
"Today's investment structure is tomorrow's industrial<br />
structure." An effective program for the curse is to reduce<br />
dependence on resource sectors, that is, the<br />
implementation <strong>of</strong> industrial diversification. Relevant<br />
government departments should encourage coal<br />
companies "based on coal," as well as make "extended<br />
coal" article, stimulate large-scale coal enterprises<br />
achieve diversification business and development<br />
methods transfer; actively guide the private capital to<br />
increase investment to non-coal tertiary industries, do a<br />
good "surpass coal" article, particularly increase<br />
investment to high-tech industries which have major<br />
breakthrough and stimulates effect to economic growth,<br />
provide new credit carrier for financial sectors, eventually<br />
through "incremental" tune "deposit quantity "approach<br />
to achieve industrial structure optimization and upgrading<br />
<strong>of</strong> coal resource cities .<br />
ACKNOWLEDGMENT<br />
This work is supported by the National Natural<br />
Science Foundation <strong>of</strong> China(71071003), the MOE<br />
Project <strong>of</strong> Youth Foundation <strong>of</strong> Humanities and Social<br />
Science(09YJC630004), and the Anhui Philosophy<br />
Society and Science Projects (AHSK05-06D55).<br />
REFERENCES<br />
[1] Sheng Kerong, Sun Wei. “Discussing the factors <strong>of</strong><br />
economical development <strong>of</strong> the coal-based cities in China,”<br />
Mining Research and Development, vol.24, mar.2004,pp.<br />
[2] Goldsmith. 1969, Financial Structure and Development,<br />
New.haven, CT: Yale. University press.<br />
[3] King and Levine. 1993, Finance and Growth: Schunpeter<br />
might be Right, Quarterly <strong>Journal</strong> <strong>of</strong> Economics, vol.108,<br />
pp.717-738.<br />
[4] Han Tingchun. “Financial development and economical<br />
growing: a positive analysis <strong>of</strong> China,” Economical<br />
technology, Vol.3, 2001, pp<br />
[5] Tan Ruyong. “A positive analysis <strong>of</strong> the relationship<br />
between finacial development and economical growing in<br />
china,” Economical Research, vol. 10, 1999, pp<br />
[6] Cao Tingqiu, Wang Xihang. “Financial development and<br />
economical growing: a positive analysis on the cities in<br />
Shandong province,” Shandong society technology, vol.<br />
1, 2006, pp.<br />
[7] Tang Jianying, Zhou Dequn etc. “A positive analysis on<br />
the variation <strong>of</strong> the total factor productivity <strong>of</strong> the coal<br />
cities in China,” <strong>Journal</strong> <strong>of</strong> China University <strong>of</strong> Mining &<br />
Techonology, vol. 11, 2007, pp.<br />
Zuhuai Yuan is currently a doctor candidate in the School <strong>of</strong><br />
Management at the University <strong>of</strong> Mining and Technology <strong>of</strong><br />
China, Xuzhou, Jiangsu, China. As well, he is the director <strong>of</strong> the<br />
municipal research institution, Huainan, Anhui, China. (E-mail:<br />
yzhhn@tom.com).
1896 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
Solving the Sparsity Problem in Recommender<br />
Systems Using Association Retrieval<br />
YiBo Chen<br />
Computer school <strong>of</strong> Wuhan University, Wuhan, Hubei, China<br />
chenyibo8224@yahoo.com.cn<br />
ChanLe Wu, Ming Xie and Xiaojun Guo<br />
Computer school <strong>of</strong> Wuhan University, Wuhan, Hubei, China<br />
National Engineering Research Center for Multimedia S<strong>of</strong>tware, Wuhan, China<br />
Email:{chanle.wu, Guoxiaojun}@gmail.com<br />
Abstract—Recommender systems are being widely applied<br />
in many fields, such as e-commerce etc, to provide products,<br />
services and information to potential customers.<br />
Collaborative filtering as the most successful approach,<br />
which recommends contents to the current customers<br />
mainly is based on the past transactions and feedback <strong>of</strong> the<br />
similar customer. However, it is difficult to distinguish the<br />
similar interests between customers because the sparsity<br />
problem is caused by the insufficient number <strong>of</strong> the<br />
transactions and feedback data, which confined the usability<br />
<strong>of</strong> the collaborative filtering. This paper proposed the direct<br />
similarity and the indirect similarity between users, and<br />
computed the similarity matrix through the relative distance<br />
between the user’s rating; using association retrieval<br />
technology to explore the transitive associations based on<br />
the user’s feedback data, realized a new collaborative<br />
filtering approach to alleviate the sparsity problem and<br />
improved the quality <strong>of</strong> the recommendation. In the end, we<br />
implemented experiment based on Movielens data set, the<br />
experiment results indicated that the proposed approach<br />
can effectively alleviate the sparsity problem, have good<br />
coverage rate and recommendation quality.<br />
Index Terms—collaborative filtering; association retrieval;<br />
sparsity problem; recommendation quality<br />
I. INTRODUCTION<br />
Along with the rapidly development <strong>of</strong> the Internet, the<br />
number <strong>of</strong> the servers connected to Internet and the Webs<br />
on WWW show a trend <strong>of</strong> exponential growth. The<br />
rapidly development <strong>of</strong> the Internet present a mass <strong>of</strong><br />
information to us at the same time, for example, there are<br />
tens <strong>of</strong> thousands movies in Netflix, millions <strong>of</strong> books in<br />
Amazon, more than 10 billion page collection in<br />
Del.icio.us, so much information, not to mention find<br />
some interesting contents, it is impossible that to gave all<br />
<strong>of</strong> information the once-over. The traditional search<br />
Manuscript received January 1, 2010; revised June 1, 2010; accepted<br />
July 1, 2010.<br />
Copyright credit, project number, corresponding author, etc.<br />
© 2011 ACADEMY PUBLISHER<br />
doi:10.4304/jcp.6.9.1896-1902<br />
algorithm only presents the same ordered results to all <strong>of</strong><br />
users; can not to provide different service to different<br />
users according to their different interests The<br />
information explosion reduced the use ratio <strong>of</strong> the<br />
information, this phenomenon is called information<br />
overload. Personalized recommendation, included<br />
personalized search, has been thought as one <strong>of</strong> the most<br />
effective tools to resolve the problem <strong>of</strong> information<br />
overload. Radically, the recommendation problem is to<br />
substitute user to evaluate the strange products, which<br />
include books, movies, CD, web and so on, it is a process<br />
from know to unknown [1]。.<br />
Recommendation as a social process plays an<br />
important role in many applications for consumers,<br />
because it is overly expensive for every consumer to learn<br />
about all possible alternatives independently. Depending<br />
on the specific application setting, a consumer might be a<br />
buyer, an information seeker, or an organization<br />
searching for certain expertise [2].<br />
Until 1990s, personalized recommendation research as<br />
an independent concept be advanced. It rapidly<br />
development origin from the web2.0‘s maturity, which<br />
make the user become a participant from browser. In an<br />
actually recommender system, there are tens <strong>of</strong><br />
thousands, or even more than one millions products need<br />
to be recommended, for instance, Amazon, eBay,<br />
Youtube, etc, also there are huge users. Accurate and<br />
high-performance recommender system can mine the<br />
potential propensity to consume <strong>of</strong> the user and provide<br />
personalized services for users. In the increasingly fierce<br />
competitive environment, personalized recommendation<br />
system is not just business marketing means, more<br />
importantly, it can improve the user‘s loyalty and prevent<br />
the loss <strong>of</strong> users.<br />
A recommender system is compose <strong>of</strong> three parts:<br />
action recorder module collect the user‘s information,<br />
model analysis module analyze the user‘s preference and<br />
recommendation algorithm module, thereinto, the<br />
recommendation algorithm module is the most core part<br />
<strong>of</strong> the recommendation system [3]. At present,<br />
recommendation algorithm mainly includes collaborative<br />
filtering algorithm, content-based algorithm, the bipartite
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1897<br />
relationship graph recommendation algorithm based on<br />
user-product and hybrid recommendation algorithm. This<br />
paper focus on the sparsity and precision problem,<br />
compute the similarly matrix through the relatively<br />
distance between the user‘s rating and use the association<br />
retrieval technology to realize a new collaborative<br />
filtering approach.<br />
The remainder <strong>of</strong> the paper is organized as follows.<br />
Section 2 surveys existing work on collaborative filtering<br />
and discusses the sparsity problem in detail. Section 3<br />
introduces associative retrieval and summarizes our<br />
associative retrieval-based approach to dealing with the<br />
sparsity problem and improve the quality <strong>of</strong> the<br />
recommendation. Section 4 presents an experimental<br />
study and the experimental data analysis. Section 5<br />
concludes the article by summarizing our research<br />
contributions and pointing out future directions.<br />
II. COLLABORATIVE FILTERING AND THE SPARSITY<br />
PROBLEM<br />
A. Collaborative filtering<br />
Collaborative filtering aggregates the experiences <strong>of</strong><br />
similar users in the system to generate personalized<br />
recommendations. One key aspect <strong>of</strong> collaborative<br />
filtering is the identification <strong>of</strong> users similar to the one<br />
who needs a recommendation depends on the preference<br />
patterns <strong>of</strong> users makes it more general than other tasks<br />
such as ad hoc information retrieval and content-based<br />
filtering [4].<br />
Collaborative filtering has been the most successful<br />
recommendation system approach to date and has been<br />
widely applied in various applications, thereinto, Grundy<br />
have been considered the first collaborative filtering<br />
system [5]. Grundy system can build user‘s preference<br />
model to recommend relevant books to every users.<br />
Tapestry mail processing system, manpower deal with the<br />
similarity between users. The more users, the lower<br />
precision [6]. GroupLens build the user‘s information<br />
group, within group <strong>of</strong> users can publish their own<br />
information, and with other users make collaborative<br />
recommendation [7]. Ringo make use <strong>of</strong> the same social<br />
information filtering method to recommend music to<br />
users [8]. There are some other typically collaborative<br />
recommendation system, such as Amazon.Com [9], Jester<br />
[10], Phoaks [11], and so on.<br />
Many algorithms have been proposed to deal with the<br />
collaborative filtering problem. Most collaborative<br />
filtering algorithms can be categorized into two classes<br />
[12]: Memory-based algorithms and model-based<br />
algorithms.<br />
The memory -based algorithms first find the users from<br />
the training database that are most similar to the current<br />
test user in terms <strong>of</strong> the rating pattern, and then combine<br />
the ratings given by those similar users to obtain the<br />
prediction for the test user. The two most commonly<br />
methods is Pearson correlation and cosine <strong>of</strong> the angle.<br />
Many enhanced method have been applied into the<br />
Pearson correlation and cosine <strong>of</strong> the angle. For example,<br />
absentee voting, case extended, weighted advantage<br />
predication, etc. Otherwise, Chen and Cheng make use <strong>of</strong><br />
© 2011 ACADEMY PUBLISHER<br />
the order within product list to compute the similar<br />
degree between users; the high-order products have<br />
higher weight when computing the user‘s comparability<br />
[13]. But Yang and Gu proposed that using user‘s<br />
behavior information to construct the user‘s interest<br />
point, make use <strong>of</strong> the interest point to compute the<br />
comparability [3][14].<br />
Model-based algorithm collects rating data to study,<br />
infer the user‘s action model, and predicate rating for a<br />
product. The difference between model-based<br />
collaborative filtering and memory-based collaborative<br />
filtering is that model-based approach not based on some<br />
<strong>of</strong> heuristic rule to predicate, but based on data<br />
application statistics and machine learning to get model<br />
to predicate. Breese et al. proposed two selection<br />
probability models: Clustering model and Bayes network<br />
[15]. In first model, suppose the user‘s rating<br />
independently, the similarly user cluster into a class, give<br />
the user class a mark number. In Bayes network, the<br />
number <strong>of</strong> class and model parameter can obtain from<br />
existing data through learning. Other model-based<br />
collaborative filtering system have probability correlation<br />
model [16], maximum entropy model, linear regression<br />
model, and so on.<br />
Despite its success in many application settings, the<br />
collaborative filtering approach nevertheless has been<br />
reported to have several major limitations including the<br />
sparsity, scalability, and synonymy problems. The<br />
sparsity problem occurs when transactional or feedback<br />
data is sparse and insufficient for identifying neighbors<br />
and it is a major issue limiting the quality <strong>of</strong><br />
recommendations and the applicability <strong>of</strong> collaborative<br />
filtering in general. Our study focused on developing an<br />
effective approach to making high-quality<br />
recommendations even when sufficient data is<br />
unavailable.<br />
B. The sparsity problem<br />
In collaborative filtering systems, users or consumers<br />
are typically represented by the items they have<br />
purchased or rated. For instance, in an online cinema<br />
have 3 million movies; each consumer is represented by a<br />
Boolean feature vector <strong>of</strong> 3 million elements. The value<br />
for each element is determined by whether this consumer<br />
has viewed the corresponding movie in the past time.<br />
Typically the value <strong>of</strong> 1 to 5 indicates that such a view<br />
had occurred and 0 indicates that no such event has<br />
occurred. When multiple consumers are concerned, a<br />
matrix composed <strong>of</strong> all vectors representing these<br />
consumers can be used to capture past view events. We<br />
call this matrix the consumer–product interaction matrix.<br />
In this article, we use C to denote the set <strong>of</strong> consumers<br />
and I to represent the set <strong>of</strong> items. We represent the<br />
consumer–product interaction matrix by a |C|×|I| matrix R<br />
= (rij), such that<br />
In many large-scale applications, both the number <strong>of</strong><br />
items and the number <strong>of</strong> consumers are large. In such<br />
cases, even when many events have been recorded, the<br />
consumer–product interaction matrix can still be<br />
(1)
1898 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
extremely sparse, that is, there are very few elements in R<br />
whose value is not 0. This problem, commonly referred to<br />
as the sparsity problem, has a major negative impact on<br />
the effectiveness <strong>of</strong> a collaborative filtering approach.<br />
Because <strong>of</strong> sparsity, it is highly probable that the<br />
similarity (or correlation) between two given users is<br />
zero, rendering collaborative filtering useless [17]. Even<br />
for pairs <strong>of</strong> users that are positively correlated, such<br />
correlation measures may not be reliable.<br />
The cold-start problem further illustrates the<br />
importance <strong>of</strong> addressing the sparsity problem. The coldstart<br />
problem refers to the situation in which a new user<br />
or item has just entered the system [18]. Collaborative<br />
filtering cannot generate useful recommendations for the<br />
new user because <strong>of</strong> the lack <strong>of</strong> sufficient previous ratings<br />
or purchases. Similarly, when a new item enters the<br />
system, it is unlikely that collaborative filtering systems<br />
will recommend it to many users because very few users<br />
have yet rated or purchased this item. Conceptually, the<br />
cold-start problem can be viewed as a special instance <strong>of</strong><br />
the sparsity problem, where most elements in certain<br />
rows or columns <strong>of</strong> the consumer–product interaction<br />
matrix A are 0 [2].<br />
Many researchers have attempted to alleviate the<br />
sparsity problem. In [19], the author proposed an itembased<br />
approach to addressing both the scalability and<br />
sparsity problems. Another proposed approach,<br />
dimensionality reduction, aims to reduce the<br />
dimensionality <strong>of</strong> the consumer–product interaction<br />
matrix directly. A simple strategy to reduce the<br />
dimensionality is to form clusters <strong>of</strong> items or users and<br />
then use these clusters as basic units in the prediction.<br />
More advanced techniques can be applied to achieve<br />
dimensionality reduction. Examples are statistical<br />
techniques such as Principle Component Analysis (PCA)<br />
[10] and information retrieval techniques such as Latent<br />
Semantic Indexing (LSI). Essentially, dimensionality<br />
reduction approaches deal with the sparsity problem by<br />
generating a denser user-item interaction matrix that<br />
considers only the most relevant users and items.<br />
Predictions are then made using this reduced matrix.<br />
Empirical studies indicate that dimensionality reduction<br />
can improve recommendation quality significantly in<br />
some applications, but performs poorly in others, the<br />
potentially useful information might be lost during this<br />
reduction process [20].<br />
Researchers have also attempted to combine<br />
collaborative filtering with content-based<br />
recommendation approaches to alleviate the sparsity<br />
problem [21][22]. In addition to user-item interactions,<br />
such techniques also consider similarities between items<br />
derived from their content, which allow them to make<br />
more accurate predictions. However, the hybrid approach<br />
requires additional information regarding the products<br />
and a metric to compute meaningful similarities among<br />
them. In practice, such product information may be<br />
difficult or expensive to acquire and a related similarity<br />
metric may not be readily available.<br />
Another category <strong>of</strong> methods consider the data as a<br />
bipartite graph where nodes represent the users and items,<br />
© 2011 ACADEMY PUBLISHER<br />
and an edge (i, j) exists between a user i and an item j if i<br />
has rated j. Moreover, edge (i, j) is given a weight<br />
corresponding to the rating given by i to j. These methods<br />
then derive global similarities between users or items<br />
using graph theoretic measures. For instance, one such<br />
method computes similarities between two users as the<br />
average commute time between their respective nodes in<br />
a random-walk <strong>of</strong> the graph. Other graph theoretic<br />
measures were also investigated, such as the minimal hop<br />
distance between nodes <strong>of</strong> the graph, and the spread<br />
activation <strong>of</strong> the nodes in the graph. The main drawback<br />
<strong>of</strong> these approaches is that there is <strong>of</strong>ten no good<br />
interpretation <strong>of</strong> the similarity measures in the context <strong>of</strong><br />
the prediction problem [23].<br />
Our research focuses on developing a computational<br />
approach to exploring transitive between users to address<br />
the sparsity problem and improving the accurate in the<br />
context <strong>of</strong> collaborative filtering.<br />
III. COLLABORATIVE FILTERING BASED ON ASSOCIATION<br />
RETRIEVAL<br />
A. Association retrieval<br />
Associative retrieval has its origin in statistical studies<br />
<strong>of</strong> associations among terms and documents in a text<br />
collection. The basic idea behind associative retrieval is<br />
to build a graph or network model <strong>of</strong> documents and<br />
index terms and queries, and then to explore the transitive<br />
associations among terms and documents using this graph<br />
model to improve the quality <strong>of</strong> information retrieval.<br />
This relationship is also reflected in people's daily life,<br />
for instance, Lisi is Wanwu‘s friend, Zhanshan is Lisi‘s<br />
friend, Wanwu can recommend movie A to Zhanshan, so<br />
there is an association relationship between Zhanshan and<br />
Wanwu. We found that recommender system can make<br />
use <strong>of</strong> this relationship between users to address the<br />
sparsity by studying.<br />
B. Finding the relationship between users by association<br />
retrieval<br />
Firstly, we supposed that represent a<br />
user set which includes 3 users,<br />
represents a movie set which includes 4 movies,<br />
represent a user‘s rating matrix which<br />
includes elements.<br />
The rows represent the user, the columns represents the<br />
movie, for example, the first row represents the user c1<br />
viewed the movies i2 and i4, the rating is 3 and 4<br />
respectively.<br />
From the second line in the matrix B, we can know that<br />
the user c2 viewed the movie i2, i3 and i4. It is easy to find<br />
that the user c1 and c2 viewed the movie i2 and i4 from<br />
matrix R and B. According to similarity theory, we can<br />
ascertain that the user c1 is similarity with the user c2, so<br />
the movie i3 can be recommended to the user c1 through
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1899<br />
the user c2, but the movie i1 cannot be recommended to c1<br />
forever. However, the above example only has 4 movies.<br />
At present, the online movie provider more than millions<br />
movies, the ―dark information‖ will appear if only<br />
through the direct similarity users to recommend, some <strong>of</strong><br />
movies will cannot be recommended to some <strong>of</strong> users, the<br />
requirements <strong>of</strong> the user cannot be satisfied.<br />
According to the association retrieval theory, users as a<br />
set <strong>of</strong> nodes, the products as a set <strong>of</strong> nodes, we use the<br />
bipartite graph to express the matrix B, as shown in Fig1.<br />
Customer Nodes<br />
C1<br />
C2<br />
C3<br />
Product Nodes<br />
i1<br />
i2<br />
i3<br />
i4<br />
Figure 1. Transitive associations in collaborative filtering.<br />
Accordering to Fig 1, the length <strong>of</strong> the association path<br />
is assumed to be 3, there are c1-i2-c2-i3 and c1-i4-c2-i3 two<br />
paths, the movie i3 is recommended to the user c1, but<br />
there is not a path whose length is 3 between i1 and c1, so<br />
i1 will not be recommended to the user c1. If the length <strong>of</strong><br />
the path is extended to 5, we can find that the movie i1<br />
can be recommended to the user c1 through the path c1-i2c2-i3-c3-i1<br />
and c1-i4-c2-i3-c3-i1.<br />
Accorder to the above analysis, this paper makes some<br />
<strong>of</strong> define are as follows:<br />
Definition 1: direct recommendation path represent a<br />
user recommend item to a target user directly.<br />
Definition 2: indirect recommendation path represent a<br />
user recommend item to a target user through one or<br />
more than one user.<br />
Definition 3: user direct similarity degree represents<br />
the similarity degree between users in direct<br />
recommendation path.<br />
Definition 4: user indirect similarity degree represents<br />
the similarity degree between recommendation user and<br />
target user in indirect recommendation path.<br />
From the above analysis, we know that the association<br />
retrieval method can explore the transitive between users<br />
to get a set <strong>of</strong> paths and the direct or indirect similarity<br />
degree. Through formula (2) to compute the value <strong>of</strong><br />
in the sparsity matrix to address the sparsity problem.<br />
(2)<br />
Note that i represents user, j is item,<br />
is the set <strong>of</strong> recommendation path,<br />
represents an ordered set <strong>of</strong> a recommendation path the<br />
user passed, is similarity degree between and .<br />
C. Computing the direct similarly matrix<br />
In the computing <strong>of</strong> the direct similarity matrix, we do<br />
not use the Pearson-correlation and cosine <strong>of</strong> the angle.<br />
Through the research we find that whatever the user<br />
© 2011 ACADEMY PUBLISHER<br />
rating is high or low after the user viewed the movie, to<br />
some extent, which express some <strong>of</strong> similarity between<br />
users both on the personal preferences and the preference<br />
<strong>of</strong> ratings. For example, in the matrix R, the user c1 and c2<br />
rated i2 and i4, the rating value <strong>of</strong> the c1 is 3 and 4, the<br />
rating value <strong>of</strong> the c2 is 2 and 5, we can use formula (3) to<br />
compute the rating similarity degree between c1 and c2 for<br />
the same movie and ,<br />
max is the maximum value function; abs is the<br />
absolute value function; R represents the value set <strong>of</strong> the<br />
rating, such as R={0,1,2,3,4,5}; , the value <strong>of</strong> the user<br />
i rate product k. Formula (4) was used to compute the<br />
user similarity between i and j after get the rating<br />
similarity degree.<br />
Note that m, the number <strong>of</strong> the products. We use the<br />
rating matrix R as an example, the user similarity<br />
, according to this method,<br />
we can get the user similarity matrix as follows:<br />
Next, we combine the association retrieval and direct<br />
similarity matrix to compute in order to get the<br />
recommendation matrix after getting the user similarity<br />
matrix.<br />
D. Computing the recommender matrix<br />
We use the data the section 3.2 provided to recommend<br />
for the user c1. When M=3, we can find that c1 has two<br />
recommendation path c1-i2-c2-i3 and c1-i4-c2-i3 from the<br />
data; the similarity between c1 and c2 is 0.4 from the<br />
similarity matrix in section 3.3, the weight <strong>of</strong> the path is<br />
0.4; so we get the correlation degree <strong>of</strong> the i3 is<br />
; Because c1 and c2 have the highest<br />
similarity, the rating value <strong>of</strong> the c2 for i3 is 3, so the<br />
recommendation value is . When M=5,<br />
there are two recommendation path c1-i2-c2-i3-c3-i1 and c1i4-c2-i3-c3-i1,<br />
the weight is<br />
, the value <strong>of</strong> the correlation degree is ,<br />
the rating value <strong>of</strong> the c3 for i1 is 4, so the<br />
recommendation value is .<br />
The recommendation matrix was defined in<br />
(5)<br />
(3)<br />
(4)<br />
(5)<br />
Note that R, the rating matrix, is the similarity<br />
matrix, B is the marked matrix. Using the data in section<br />
3.2, we get the recommendation matrix and<br />
through formula (4) when M=3 and M=5.
1900 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
From the above recommendation matrix, we can know<br />
that ,<br />
, which consistent with the above computing<br />
outcome.<br />
IV. ALGORITHM<br />
The algorithm is as follows:<br />
Algorithm1. Collaborative algorithm based on<br />
association retrieval<br />
Input:user rating matrix R,the length <strong>of</strong> path M<br />
Output:Recommendation matrix<br />
Step1. Matrix B = Matrix R, If not equal 0 then<br />
for each .<br />
Step2. Set the iteration variable N=1.<br />
Step3. Original recommendation matrix .<br />
Step4. Compute the direct similarity matrix<br />
according to formula (3) and (4).<br />
Step5. Compute the transpose .<br />
Step6. Compute the matrix according to<br />
formula (5).<br />
Step7. If N+2 less than M then N=N+2, goto Step 3 until<br />
N larger than M.<br />
A. Experiment data<br />
V. EXPERIMENT AND ANALYSIS<br />
The datasets were collected by the GroupLens<br />
Research Project at the University <strong>of</strong> Minnesota.<br />
The data set consists <strong>of</strong> 100,000 ratings (1-5) from 943<br />
users on 1682 movies. Each user has rated at least 20<br />
movies; the sparsity degree is 99.937%.<br />
R<br />
C. Experiment results<br />
In our experiment, we called our approach ARC. We<br />
respectively compute the precision, recall and F-measure<br />
based on Movielens data set for the ARC, PC and COS<br />
algorithms. In the ARC, the value <strong>of</strong> the M is 3.<br />
Summarized bar charts are shown in Figs. 2–5. Table1 is<br />
the comprehensive comparison about the precision, recall<br />
F-measure and coverage between ARC, PC and COS<br />
algorithms.<br />
In the aspect <strong>of</strong> the precision, the ARC increased by<br />
18.40% compared with PC and 33.58% compared with<br />
COS. In the aspect <strong>of</strong> the recall, the ARC increased by<br />
17.65% compared with PC and 66.68% compared with<br />
COS. In the aspect <strong>of</strong> the F-measure, the ARC increased<br />
by 18.39% compared with PC and 34.13% compared<br />
with COS. In the aspect <strong>of</strong> coverage, the ARC increased<br />
© 2011 ACADEMY PUBLISHER<br />
B. Experiment procedure<br />
For each target consumer, we retrieved the entire set <strong>of</strong><br />
previously viewed items and sorted them into<br />
chronological order by view date. The first 90% <strong>of</strong> these<br />
items was treated as ―past‖ views to serve as input to be<br />
fed into different methods to generate recommendations.<br />
For comparison purposes, the second 10% <strong>of</strong> these items<br />
were treated as ―future‖ views <strong>of</strong> the customer and hidden<br />
from the recommender system.<br />
In the experiment, we compared the outcome <strong>of</strong> the<br />
Pearson-correlation, Vector similarly, Item-based and our<br />
approach. We use precision, recall, coverage and Fmeasure<br />
to measure the effectiveness <strong>of</strong> a given<br />
recommendation approach. These measures are widely<br />
accepted in information retrieval and recommender<br />
system research [24].<br />
The baseline methods are described below.<br />
Pearson Correlation Coefficient (PCC)<br />
Pearson Correlation Coefficient method predicts the<br />
rating <strong>of</strong> a test user x on item i as:<br />
(5)<br />
Where the coefficient is computed as<br />
Vector Similarity (VS)<br />
This method is very similar to the previous method<br />
except that the correlation coefficient is<br />
computed as:<br />
The definition <strong>of</strong> the precision, recall, coverage and Fmeasure<br />
are as follows.<br />
(8)<br />
(9)<br />
(10)<br />
(11)<br />
by 4.66% compared with PC and 24.78% compared with<br />
COS. From the results, we can see that there are greatly<br />
improved in the aspect <strong>of</strong> the precision, recall, F-measure<br />
and coverage. But from the above data, we find that the<br />
COS is worst in the situation <strong>of</strong> the sparsity. Otherwise,<br />
in the aspect <strong>of</strong> coverage, the ARC increased by only<br />
4.66% compared with PC. We also make another<br />
experiment, the results show that the coverage can<br />
increase more than 10% when the M equals 5, the<br />
overhead <strong>of</strong> the computing have great increased, but the<br />
increase was very little in the recommendation precision.<br />
This paper considers that a low coverage rate increase for<br />
two reasons, on the one hand, it is because the value <strong>of</strong><br />
the M is 3; on the other hand, maybe the sparse degree <strong>of</strong><br />
the experiment data set is not enough.<br />
(6)<br />
(7)
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1901<br />
ASS<br />
Value <strong>of</strong> the precision<br />
Value <strong>of</strong> the recall<br />
0.018<br />
0.016<br />
0.014<br />
0.012<br />
0.01<br />
0.008<br />
0.006<br />
0.004<br />
0.002<br />
0<br />
1<br />
0.9<br />
0.8<br />
0.7<br />
0.6<br />
0.5<br />
0.4<br />
0.3<br />
0.2<br />
0.1<br />
0<br />
TABLE1.<br />
COMPREHENSIVE COMPARISON TABLE<br />
PC COS<br />
Precision Recall F-measure Coverage Precision Recall F-measure Coverage<br />
D-value 0.00256 0.1429 0.0503 0.0378 0.00414 0.381 0.0824 0.1685<br />
The percent <strong>of</strong> the<br />
improving<br />
0.01647<br />
18.40% 17.65% 18.39% 4.66% 33.58% 66.68% 34.13% 24.78%<br />
Precision<br />
0.01391<br />
Figure 2. The comparison <strong>of</strong> the predictive precision<br />
Figure 3. The comparison <strong>of</strong> the recall<br />
VI. CONCLUSION<br />
0.01233<br />
ARC PC COS<br />
Name <strong>of</strong> the method<br />
0.9524<br />
0.8095<br />
0.5714<br />
ARC PC COS<br />
Name <strong>of</strong> the method<br />
In this paper, we aimed to alleviate the sparsity<br />
problem and improve the recommendation precision in<br />
collaborative filtering systems. We use the association<br />
retrieval technology to alleviate the sparsity problem and<br />
proposed a new collaborative filtering algorithm to<br />
increase the recommendation precision. The effectiveness<br />
<strong>of</strong> the approach was evaluated experimentally using data<br />
from the movielens data set. The experiment indicated<br />
that our approach alleviated the sparsity problem and<br />
achieved significantly better recommendation quality<br />
than the standard collaborative filtering approaches<br />
Meanwhile, there is a great problem for the proposed<br />
approach in this paper. The volume <strong>of</strong> data these s ystems<br />
utilize will continue increasing over time. In this<br />
situation, our approach will cause the data overload<br />
problem. As a result, it will present a significant<br />
© 2011 ACADEMY PUBLISHER<br />
Recall<br />
0.4<br />
0.35<br />
0.3<br />
0.25<br />
0.2<br />
0.15<br />
0.1<br />
0.05<br />
0<br />
Value <strong>of</strong> the F-measure<br />
Value <strong>of</strong> the coverage<br />
0.9<br />
0.8<br />
0.7<br />
0.6<br />
0.5<br />
0.4<br />
0.3<br />
0.2<br />
0.1<br />
0<br />
0.3238<br />
F-measure<br />
0.2735<br />
Figure 4. The comparison <strong>of</strong> the F-measure<br />
Figure 5. The comparison <strong>of</strong> the coverage<br />
challenge for the scalability <strong>of</strong> collaborative filtering<br />
recommenders. So, the next research, we will consider<br />
the scalability problem <strong>of</strong> collaborative filtering<br />
recommenders.<br />
ACKNOWLEDGMENT<br />
This work was supported in the National Natural Science<br />
Foundation <strong>of</strong> China under Grant No. 60672051. The<br />
Fundamental Research Funds for the Central Universities<br />
(3105005). Wuhan science and technology plan projects<br />
(201010621209)<br />
REFERENCES<br />
0.2414<br />
ARC PC COS<br />
Name <strong>of</strong> the method<br />
0.8484<br />
Coverage<br />
0.8106<br />
0.6799<br />
ARC<br />
Name <strong>of</strong> the<br />
PC<br />
method<br />
COS<br />
[1] Liu Jianguo, Zhou Tao, et al. Progress <strong>of</strong> the personalized<br />
recommendation systems. Progress <strong>of</strong> Nature and Science,<br />
200919(1):1-15<br />
[2] Zan Huang, Hsinchun Chen, et al. Applying Associative<br />
Retrieval Techniques to Alleviate the Sparsity Problem in
1902 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
Collaborative Filtering. ACM Transactions on Information<br />
Systems, Vol. 22, No. 1, January 2004, 116–142.<br />
[3] Liu Jianguo, Zhou Tao, et al. Overview <strong>of</strong> the Evaluated<br />
Algorithms for the Personal Recommendation Systems.<br />
Complex System and Complexity Science. 2009, Vol.6,<br />
No.3, 1-10.<br />
[4] Rong Jin, Luo Si, et al. Collaborative Filtering with<br />
Decoupled Models for Preferences and Ratings. CIKM ’03,<br />
New Orleans, Louisiana, USA, November 3-8, 2003.<br />
[5] Rich E. User modeling via stereotypes. Cognitive Science,<br />
1979, 3(4):329-354.<br />
[6] Goldberg D, Nichols D, et al. Using collaborative filtering<br />
to weave an information tapestry. Comm ACM, 1992,<br />
35(12):61-70.<br />
[7] Konstan JA, Miller BN, et al. GroupLens: Applying<br />
collaborative filtering to usenet news. Comm ACM, 1997,<br />
40(3):77-87<br />
[8] Shardanand U, Maes P. Social information filtering:<br />
Algorithms for automating ‗Word <strong>of</strong> Mouth‘. Proc Conf<br />
Human Factors in Computing Systems Denver, 1995, 210-<br />
217.<br />
[9] Linden G, Smith B, et al. Amazon.com recommendations:<br />
Item-to-item collaborative filtering. IEEE Internet<br />
Computing. 2003, 7(1):76-80.<br />
[10] Goldberg K, Roeder T, et al. Eigentaste: A constant time<br />
collaborative filtering algorithm. Information Retrieval.<br />
2001, 4(2):133-151.<br />
[11] Terveen L, Hill W, et al. PHOAKS: A system for sharing<br />
recommendations. Comm ACM, 1997, 40(3):59-62.<br />
[12] J. S. Breese, D. Heckerman, et al. Empirical Analysis <strong>of</strong><br />
Predictive Algorithms for Collaborative Filtering,<br />
Proceeding <strong>of</strong> the Fourteenth Conference on Uncertainty<br />
in Artificial Intelligence (UAI). 1998.<br />
[13] Chen YL, Cheng LC. A novel collaborative filtering<br />
approach for recommending ranked items. Expert Systems<br />
with Applications, 2008, 34(4):2396-2405.<br />
[14] Yang MH, Gu ZM. Personalized recommendation based<br />
on partial similarity <strong>of</strong> interests. Advanced Data Mining<br />
and Applications Proceedings, 2006, 4093:509-516.<br />
[15] Breese JS, Heckerman D, et al. Empirical analysis <strong>of</strong><br />
predictive algorithms for collaborative filtering. Proc 14th<br />
Conf Uncertainty in Artificial Intelligence Madison, 1998,<br />
43-52<br />
[16] Getoor L, Sahami M. Using probabilistic relational models<br />
for collaborative filtering. Proc Workshop Web Usage<br />
Analysis and User Pr<strong>of</strong>iling, San Diego. 1999.<br />
[17] Billsus, D., Pazzani, M. J. Learning collaborative<br />
information filters. In Proceedings <strong>of</strong> the 15th<br />
International Conference on Machine Learning, 1998, 46–<br />
54.<br />
[18] Schein, A. I., Popescul, A., et al. Methods and metrics for<br />
coldstart recommendations. In Proceedings <strong>of</strong> the 25th<br />
Annual International ACM SIGIR Conference on Research<br />
and Development in Information Retrieval (SIGIR 2002).<br />
(Tampere, Finland), 2002, 253–260.<br />
[19] Sarwar, B., Karypis, G., et al. Item-based collaborative<br />
filtering recommendation algorithms. In Proceedings <strong>of</strong> the<br />
10th International World Wide Web Conference. 2001,<br />
285–295.<br />
[20] Sarwar, B., Karypis, G., et al. Application <strong>of</strong><br />
dimensionality reduction in recommender systems: A case<br />
study. In Proceedings <strong>of</strong> the WebKDD Workshop at the<br />
ACM SIGKKD. ACM, New York.2000.<br />
[21] Good, N., Schafer, J., et al. Combining collaborative<br />
filtering with personal agents for better recommendations.<br />
In Proceedings <strong>of</strong> the 16th National Conference on<br />
Artificial Intelligence, 1999, 439–446.<br />
© 2011 ACADEMY PUBLISHER<br />
[22] Huang, Z., Chung, W., et al. A graph-based recommender<br />
system for digital library. In Proceedings <strong>of</strong> the 2nd<br />
ACM/IEEE-CS Joint Conference on Digital Libraries<br />
(Portland, Ore.). ACM, New York, 2002, 65–73.<br />
[23] Chrsistian Desrosiers, George Karypis. Solving the<br />
Sparsity Problem: Collaborative Filtering via Indirect<br />
Similarities. Technical Report. Department <strong>of</strong> Computer<br />
Science and Engineering University <strong>of</strong> Minnesota 4-192<br />
EECS Building 200 Union Street SE Minneapolis, MN<br />
55455-0159 USA. 2008.<br />
[24] Sarwar, B., Karypis, G., et al. Analysis <strong>of</strong> recommendation<br />
algorithms for e-commerce. In Proceedings <strong>of</strong> the ACM<br />
Conference on Electronic Commerce. ACM, New York,<br />
2000, 158–167.<br />
Yibo Chen, born in 1982, Ph.D. candidate. The research<br />
interests include personalization recommendation and semantic<br />
web.<br />
Chanle Wu, born in 1945, pr<strong>of</strong>essor, The interests include<br />
computer networks, e-Learning, grid computing and semantic<br />
web.<br />
Ming Xie, born in 1978, Ph.D. candidate. The research interests<br />
include data mining and semantic web.<br />
Xiaojun Guo, born in 1984, Ph.D. candidate. The research<br />
interests include semantic web and e-Learning.
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1903<br />
Integrated Structure and Control Design<br />
for Servo System Based on Genetic<br />
Algorithm and Matlab<br />
Dingzhen Li<br />
Department <strong>of</strong> Electronics and Electrical Engineering, Nanyang Institute <strong>of</strong> Technology, Henan 473004, China<br />
Email: lidingzhen.student@sina.com<br />
Ruimin Jin<br />
Department <strong>of</strong> Electronics and Electrical Engineering, Nanyang Institute <strong>of</strong> Technology, Henan 473004, China<br />
Email: jinruimin2004@163.com<br />
Abstract—The integrated design was introduced about the<br />
pitching part <strong>of</strong> airborne radar servo system. The paper<br />
analyzed both the servo system model and the dynamic<br />
character <strong>of</strong> this system. Then the research about parameter<br />
optimization and simulation had been done by using the<br />
Matlab. The electromechanical coupling model and the<br />
optimization model were built up based on model <strong>of</strong><br />
mechanism transmission system and electricity control<br />
system. The optimization model includes the integrated<br />
design <strong>of</strong> structure parameters and control parameters. The<br />
dynamics model <strong>of</strong> mechanism transmission system includes<br />
the nonlinearity <strong>of</strong> backlash. It considered the influence <strong>of</strong><br />
parameters for dynamics property in structure <strong>of</strong> the<br />
mechanism transmission system. Furthermore, the model <strong>of</strong><br />
electricity control system <strong>of</strong> the airborne radar servo system<br />
has three feedback-control loops. The method <strong>of</strong> integrated<br />
structure and control design was applied on the<br />
optimization model using Genetic Algorithm (GA).<br />
Simulation had been done based on Matlab/Simulik.<br />
Simulation results showed that the method <strong>of</strong> integrated<br />
structure and control design is very good. It is feasible and<br />
effective for airborne radar servo system. It proved the<br />
method used in the task is right and the practicability <strong>of</strong><br />
Genetic Algorithm.<br />
Index Terms—Radar Servo System, Integrated Design,<br />
Electromechanical Coupling Model, Genetic Algorithm,<br />
Matlab Simulation<br />
I. INTRODUCTION<br />
Modern mechanical and electrical systems have higher<br />
demanding for the system accuracy and steady-state<br />
dynamic performance in a variety <strong>of</strong> extreme conditions.<br />
The traditional design method is that the mechanical part<br />
is designed first, then the control part is designed. This<br />
method ignores the dynamics <strong>of</strong> the mechanical system,<br />
the interactions and the mutual coupling <strong>of</strong> the control<br />
system. The repeated design <strong>of</strong> the structural parameters<br />
and control parameters also causes the design cycle<br />
© 2011 ACADEMY PUBLISHER<br />
doi:10.4304/jcp.6.9.1903-1912<br />
getting long and the cost increasing. It is also difficult to<br />
achieve the best performance <strong>of</strong> the electromechanical<br />
systems.<br />
In order to improve the overall performance <strong>of</strong> the<br />
system, we should unify the model to the mechanical and<br />
electrical systems and have an integrated design on the<br />
base <strong>of</strong> this. Namely, the integrated design <strong>of</strong> structure<br />
parameters and control parameters will be done.<br />
This paper studied the airborne radar pitch servo<br />
system. The first unit is a single stage gear meshing. We<br />
research the dynamics model <strong>of</strong> the pitch<br />
servo-mechanical drive system. The influence <strong>of</strong> the<br />
system dynamic characteristics is analyzed on the<br />
structural parameters <strong>of</strong> the mechanical transmission<br />
system. Secondly, the design <strong>of</strong> the loop circurt <strong>of</strong> servo<br />
system is introduced. Then we establish a servo-electric<br />
control system <strong>of</strong> this model and study the simulation<br />
model. Finally, the servo control system is coupled with<br />
mechanical system and electrical system. The mechanical<br />
structure is established including the structure parameters<br />
and control parameters <strong>of</strong> the electromechanical coupling<br />
model. Then we create an integrated optimization model<br />
for the structure/control integrated design.<br />
Airborne radar servo system includes machine system<br />
and electrical control system [1]. Traditional design<br />
methods all make the structure and control <strong>of</strong> radar servo<br />
system as a separate module, later the structure design is<br />
optimized, and then is added the optimal controller to the<br />
structure. It is sequential, serial. All parts are independent<br />
design, which ignore the strong coupling between the<br />
structural parameters and control parameters. It is very<br />
difficult to achieve the global optimum. To overcome this<br />
shortcoming, the design uses a method that makes the<br />
structure and control integrated. During the process <strong>of</strong> the<br />
design, we synchronize the optimization <strong>of</strong> structure<br />
parameters and control parameters. The simulation results<br />
show that the method can improve the synthetic<br />
performance <strong>of</strong> radar servo system effectively.
1904 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
II. INTRODUCTION OF THE OPTIMIZATION<br />
ALGORITHM<br />
The model <strong>of</strong> radar pitch servo system considers the<br />
influence <strong>of</strong> backlash nonlinear factors. Because <strong>of</strong> the<br />
multiple constraints <strong>of</strong> structure and control, the<br />
optimization design is more difficult, so that the<br />
optimization problem is finally reduced to the solution <strong>of</strong><br />
a nonlinear function which is constrained. Using<br />
quasi-Newton method <strong>of</strong> nonlinear optimization<br />
algorithm is not only large, but also falling into local<br />
optimal solution frequently, resulting in the failure <strong>of</strong><br />
optimization. The genetic algorithm is based on the<br />
fitness function, by manipulating implementation for the<br />
genetic <strong>of</strong> all individuals <strong>of</strong> the population to achieve<br />
restructuring within the group iterative process <strong>of</strong><br />
individual search method. And the search does not<br />
depend on gradient information, especially for dealing<br />
with complex problems and nonlinear problems that<br />
traditional search methods can not solve [2], and get<br />
optimal solution <strong>of</strong> the global system, so it has been<br />
widely used in integrated design <strong>of</strong> the structure and<br />
control. Therefore, the design uses a genetic algorithm<br />
theory.<br />
Genetic Algorithm(GA) is random optimization search<br />
method by simulating the natural selection and heredity<br />
mechanism in the natural biology evolvement.In the<br />
engineering applications,there are many problems in<br />
dealing with the multi-parameter and multi-objective,<br />
such as optimizing the parameter <strong>of</strong> the servo system<br />
adjustor. It can be worked by using the GA which is used<br />
to design the program by different requirements <strong>of</strong> the<br />
system design. With the Matlab/Simulink, the best result<br />
can be searched in whole space and be given at last. The<br />
paper used Genetic Algorithm to optimize the parameter<br />
<strong>of</strong> P and PI, and received fine impression. It proved the<br />
method used in the task is right and the practicability <strong>of</strong><br />
genetic algorithm.<br />
x , endPop , bPop , traceInfo ] = ga(<br />
bounds ,<br />
selectFN<br />
x is the optimal solution obtained. endPop is the final<br />
population we get. bPop is the search trajectory <strong>of</strong> the<br />
optimal population. traceInfo is the information<br />
optimized. bounds is the matrix which represents the<br />
upper and the lower bounds <strong>of</strong> the variable input<br />
parameters. evalFN is the fitness function whose format<br />
is: function [val, sol] = evalFN (sol, options), <strong>of</strong> which<br />
val indicate fitness defined in the fitness function, sol is<br />
the design bariables in the process <strong>of</strong> the optimication.<br />
Namely the genetic algorithm is individual. startPop is<br />
the initial population function, which is used to initialize<br />
the genetic algorithm. This is the format: startPop = in<br />
itializega (PopulationSize, bounds, evalFN, evalOps,<br />
options ), in which, populationSize is used to specify the<br />
size <strong>of</strong> the initial individuals <strong>of</strong> each generation, bounds,<br />
evalFN is the same as the previous definition. For other<br />
parameters, you may see the help.<br />
With these prepared knowledge above, you can use the<br />
GAOT toolbox directly to design the system <strong>of</strong> this paper<br />
in the following integrated design.<br />
Genetic algorithm (GA) is a searching for optimization<br />
algorithm that is based on the principle <strong>of</strong> natural<br />
selection and genetic mechanism. The major steps<br />
include coding, initial generation <strong>of</strong> population,<br />
adaptation detection and evaluation, selection, crossover<br />
and mutation. The flow chart is shown in Fig.1. Optimal<br />
value is finally output.<br />
Figure 1. GA flow chart<br />
There are three toolboxes <strong>of</strong> genetic algorithm, say<br />
GAOT, GATBX and GADS. GAOT is a free toolbox<br />
circulated on the Internet. It is not the s<strong>of</strong>tware that<br />
comes from MATLAB. But it can be easily configured to<br />
use. The default toolbox is the objective function for<br />
solving the maximum, while the structure/control<br />
integrated design goal is to solve the objective function <strong>of</strong><br />
the maximum. So it is optimized GAOT toolbox and<br />
more convenient.<br />
GAOT toolbox includes many useful functions [3].<br />
The main program provides a genetic algorithm toolbox<br />
and the external interface. Its function format is as<br />
follows:<br />
evalFN , evalOps , startPop , opts,<br />
termFN , ,<br />
, selectOps,<br />
xOverFNs,<br />
xOverOps,<br />
mutFNs,<br />
mutOps)<br />
[ termOps<br />
© 2011 ACADEMY PUBLISHER<br />
III. ESTABLISH OF ELECTROMECHANICAL<br />
COUPLING MODEL<br />
In order to do the structure/control integration design,<br />
we must establish a servo system electromechanical<br />
coupling model which combines the mechanical system<br />
and control system [4]. In order to show the coupling<br />
model clearly, we can use Matlab/Simulik technology-<br />
related to group the function-related modules as<br />
subsystems and establish multi-level hierarchical model.<br />
Fig. 2 is the pitch servo system electromechanical<br />
coupling model established. The internal structure <strong>of</strong> each<br />
module is shown in Fig. 2(a) ~ Fig. 2(e) below. We will<br />
introduce them in detail.<br />
In the basic design <strong>of</strong> the parameters <strong>of</strong> PWM<br />
controller, time constant and the loop filter constants are<br />
very small, which has very little effect on the results <strong>of</strong><br />
the simulation system, so we ignore these factors in the<br />
coupled model [5].
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1905<br />
In the three-ring control structure, current loop and<br />
speed loop are inner ring, while position loop is outer<br />
ring. This structure can get better dynamic following<br />
performance and anti-jamming performance. Among<br />
them, the function <strong>of</strong> current loop is to change the<br />
transfer function and improve speed <strong>of</strong> the system. It<br />
timely inhibits internal interference <strong>of</strong> current loop and<br />
limits maximum current. It makes the system with<br />
enough accelerate torque and ensures safety operation <strong>of</strong><br />
the system. The role <strong>of</strong> the speed loop is to enhance the<br />
ability <strong>of</strong> system disturbances and inhibit speed<br />
fluctuation. Position loop is to guarantee system static<br />
precision and dynamic tracking performance, making the<br />
whole servo system stabilize, with high-performance<br />
operation. Three-ring controller fits quality or not<br />
directly relates to the servo drive system stability,<br />
accuracy and quickness. For control system has<br />
multi-ring structures, its controller parameters were set<br />
as follows: we first design controller parameters <strong>of</strong> the<br />
inner loop, then design the outer controller parameters<br />
regarding inner loop as a link, and ultimately design<br />
parameters <strong>of</strong> all the control loops in this manner.<br />
A. Electrical Machine Model<br />
Fig. 2 (a) is the module <strong>of</strong> electrical machine model.<br />
In the armature current consecutive cases, the armature<br />
voltage balance equation will be<br />
di di<br />
u− E = iR+ L = R( i+ Ti<br />
) . (1)<br />
dt dt<br />
In equation(1), u is the voltage added to the two<br />
terminals <strong>of</strong> the motor. E is back emf. i is armature<br />
current. R is total resistance. L is all the armature loop<br />
© 2011 ACADEMY PUBLISHER<br />
Figure 2. Diagram <strong>of</strong> the electromechanical coupling model<br />
Figure 2 (a). Module <strong>of</strong> the electrical machine model<br />
inductance. Ti L R = is called electromagnetic time<br />
constant in the armature loop. Then there is<br />
E = K θ . (2)<br />
e m<br />
In equation(2), Ke is called back emf coefficient.<br />
θm is the motor shaft corner.<br />
Concerning motor, its moment balance equation <strong>of</strong><br />
rotation axis are as follows<br />
⎧ ⎪Tm<br />
= I <br />
mθm + Bmθ m + M l<br />
⎨<br />
. (3)<br />
⎪⎩ Tm = K ti<br />
In equation (3), T representes the output torque.<br />
m<br />
K is the moment coefficient. t<br />
l M is the motor<br />
disturbance moment. I is inertia moment <strong>of</strong> motor<br />
m<br />
armature. B is damper <strong>of</strong> motor rotor.<br />
m<br />
Take laplace transformation to the above equation (1)<br />
~ equation (3) at the same time. Then we get equation<br />
(4).<br />
⎧U<br />
( s) − E( s) = R( I( s) + TiI( s) s)<br />
⎪<br />
⎨ E( s) = K eθ<br />
( s)<br />
⎪<br />
I( s) K t = M l + Imθ ⎩<br />
( s) s<br />
. (4)<br />
Simulation square diagram <strong>of</strong> motor mathematical<br />
model by equation (4) can be shown in Fig. 2(a).<br />
Each symbol in the diagram have been definited.<br />
Related parameters with pitch servo system drive motor<br />
in this paper can be seen in table .
1906 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
TABLE Ⅰ.BASIC PARAMETERS OF PITCH SERVO MACHINE<br />
Rated current Ic 5.42 amps<br />
Peak current Ip 25.3 amps<br />
Resistance R 1.81 ohms<br />
Induction L 5.1 mH<br />
Torque constant Kt 0.462 N . m/amp<br />
Constant <strong>of</strong> emf Ke 48.4 V/krpm<br />
Inertia moment Im 1.96E-04 kg . m 2<br />
Damper Bm 4.653E-04 N . m . s/rad<br />
B. Design <strong>of</strong> the Current Loop<br />
In practical engineering applications, we <strong>of</strong>ten add the<br />
current loop circuit in the speed loop circuit to ensure<br />
rapid start-up performance <strong>of</strong> the system. PI controller<br />
makes the current loop have the steady-state tracking<br />
performance <strong>of</strong> the step signal without static error. It can<br />
also effectively reduce the time constant <strong>of</strong> the motor<br />
circuit. This provides the design basis <strong>of</strong> the speed loop<br />
controller with rapid response.<br />
Fig. 2 (b) is the module <strong>of</strong> electrical current loop. In<br />
the current loop diagram, transfer function <strong>of</strong> the<br />
Pulse-Width Modulator is following<br />
K pwm<br />
Gpwm<br />
=<br />
T s+<br />
1<br />
. (5)<br />
pwm<br />
Figure 2 (b). Module <strong>of</strong> the electrical current loop<br />
In equation (5), Kpwm is the voltage amplification<br />
factor <strong>of</strong> PWM controller. TPWM is the time constant <strong>of</strong><br />
the PWM controller.<br />
Transfer function <strong>of</strong> PI regulator in current loop is<br />
following<br />
Tao _i× s + 1<br />
GIT = K _ i×<br />
. (6)<br />
Tao _ i × s<br />
In equation (6), Tao-i is integral time constant <strong>of</strong><br />
current loop. K_i is the scale factor <strong>of</strong> current loop. K_A<br />
is the current amplification factor <strong>of</strong> feedback loop.<br />
C. Design <strong>of</strong> the Speed Loop<br />
Fig. 2(c) is the module <strong>of</strong> speed loop. In practical<br />
system debugging, the speed loop has good or bad effect<br />
on system performance considerably. On the one hand,<br />
increasing the speed loop gain can increase the stiffness<br />
<strong>of</strong> the speed loop, reduce the sensitivity <strong>of</strong> the system for<br />
dynamic and static friction, overcome the dead zone, and<br />
reduce the torque fluctuations. On the other hand, it can<br />
effectively expand the system bandwidth and prevent the<br />
mechanical resonance <strong>of</strong> turntable. During the debugging,<br />
we found that we wanted to realize the system load and<br />
© 2011 ACADEMY PUBLISHER<br />
other disturbances on the robustness, the speed must be<br />
well designed to ensure three-ring steady. This paper<br />
used the speed loop with PI controller still. In the system<br />
block diagram <strong>of</strong> the speed loop, transfer function <strong>of</strong> the<br />
speed loop with PI controller is<br />
T _ v× s+<br />
1<br />
GST = K _ v×<br />
T _ v× s<br />
. (7)<br />
In equation (7), T_v is integral time constant <strong>of</strong> the<br />
speed loop. K_v is proportional coefficient <strong>of</strong> the speed<br />
loop. Transfer function <strong>of</strong> the speed loop filter is<br />
GSL<br />
1<br />
=<br />
. (8)<br />
K _ b× s+<br />
1<br />
In equation (8), K_b is the filter constant <strong>of</strong> the speed<br />
loop. In addition, K_B is the current amplification factor<br />
<strong>of</strong> the feedback loop.<br />
Figure 2 (c). Module <strong>of</strong> the speed loop<br />
D. Design <strong>of</strong> the Position Loop<br />
Fig. 2(d) is the module <strong>of</strong> the position loop. Upon<br />
completion <strong>of</strong> the current loop and speed loop, the<br />
system block diagram is established. The position loop<br />
can be designed as closed-loop. Finally, the performance<br />
indicators <strong>of</strong> the system need be achieved in the last<br />
position loop. Therefore, the position loop is a very<br />
important part <strong>of</strong> servo control system in the whole<br />
design. This position loop controller uses PI controller.<br />
Transfer function <strong>of</strong> the position loop with PI controller<br />
is<br />
Ti × s + 1<br />
GPT = Kp×<br />
. (9)<br />
Ti × s<br />
In equation (9), Ti is integral time constant and Kp is a<br />
proportion coefficient in the position loop. Transfer<br />
function <strong>of</strong> the position loop filter is<br />
GPL<br />
1<br />
=<br />
. (10)<br />
K _c× s+<br />
1<br />
In equation (10), K_c is filter constant <strong>of</strong> the position<br />
loop.<br />
Figure 2 (d). Module <strong>of</strong> the position loop<br />
Fig. 2(e) is the module <strong>of</strong> the load model. Load<br />
module is very complex. We did not introduce in detail.
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1907<br />
In the end, we established the three closed-loop<br />
simulation model <strong>of</strong> control system and got its<br />
simulation. The corresponding initial parameters were<br />
Jeq = 0.018, Beq<br />
= 0.0004653,<br />
K pwm = 4.6, Tpwm= 5e−5, K _ A= 0.3333, K _ a = 5e−5, K _ B = 0.5, K _ b = 0.0002,<br />
K _ c=0.002.<br />
Using the original parameters, we debuged the control<br />
system with PI controller repeatly. Then we adjusted the<br />
controller parameters in the ring. As follows:<br />
© 2011 ACADEMY PUBLISHER<br />
Figure 2 (e). Module <strong>of</strong> the load model<br />
K _ i= 5, Tao_i= 0.0028, K _ v= 47.876, T _ v=<br />
0.32265,<br />
Kp = 59.178, Ti = 789.62.<br />
For example, the position loop simulation curve was<br />
shown in Fig. 3.<br />
The system established under ideal conditions is stable<br />
in Fig. 3(a) and Fig. 3(b). At the same time, the system<br />
transient response <strong>of</strong> step signal is essential to meet the<br />
performance requirements <strong>of</strong> the radar servo system. We<br />
will adjust the controller parameters and system tricyclic<br />
initial design parameters <strong>of</strong> the other. These parameters<br />
are the choice <strong>of</strong> parameter ranges <strong>of</strong> the structure and<br />
control integration design.
1908 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
Position / rad<br />
1.4<br />
1.2<br />
1<br />
0.8<br />
0.6<br />
0.4<br />
0.2<br />
Reference Signal and Response Signal<br />
Reference Signal<br />
Response Signal<br />
0<br />
0 1 2 3 4 5 6 7 8 9 10<br />
10<br />
Time(s)<br />
-2<br />
10 0<br />
10 2<br />
-720<br />
Frequency (rad/sec)<br />
(a) Response curve <strong>of</strong> step signal (b) bode diagram<br />
IV. ESTABLISHMENT OF OPTIMIZATION MODEL<br />
Structure/control integrated design methods can<br />
overcome the internal contradictions and the waste <strong>of</strong><br />
source brought about by the independence <strong>of</strong> structure<br />
and control. It can also promote each other by<br />
coordinating the relationship between them, so that we<br />
can achieve a optimum control efficiency. So, it is <strong>of</strong><br />
great importance. Here we take an example <strong>of</strong> the stage<br />
signal to do the integrated design <strong>of</strong> the radar pitch servo<br />
system.<br />
A. Determination <strong>of</strong> the Objective Function<br />
Since we mainly study the influence on the effect <strong>of</strong><br />
the servo system brought about by nonlinear factors <strong>of</strong><br />
the gear gap, we take gear gap as the objective function<br />
in the optimization model. Under the premise <strong>of</strong> meeting<br />
the demand <strong>of</strong> the system functions, structure/control<br />
integrated design optimized the maximum gear gap value<br />
that the system allowed [6].<br />
There are two levels gear gap in the designed servo<br />
system model: high speed gear gap and low speed gear<br />
gap [7]. To take gear gap as the objective function, we<br />
should consider the weight coefficient <strong>of</strong> the two gear<br />
gaps. So the objective function is taken as in<br />
F = λ b + λ b . (11)<br />
1 1 2 2<br />
The impact <strong>of</strong> the low-level backlash is larger than<br />
that <strong>of</strong> the high level in the second gear transmission<br />
mechanism. So the weight factor is taken as in<br />
λ = 0.2, λ = 0.8 . (12)<br />
1 2<br />
B. Design Variables and Constraint Conditions<br />
1) Selection <strong>of</strong> design variables<br />
The selection <strong>of</strong> design variables on structure:the teeth<br />
number <strong>of</strong> each gear is Z1, Z2, Z3 and Z4 respectively.<br />
Semi-backlash are b1 and b2. We take tricyclic PI<br />
© 2011 ACADEMY PUBLISHER<br />
Magnitude (dB)<br />
Phase (deg)<br />
100<br />
0<br />
-100<br />
-200<br />
-300<br />
-400<br />
0<br />
-180<br />
-360<br />
-540<br />
Figure 3. Location-loop simulation curve<br />
Bode Diagram<br />
controller parameters K_i, Tao_i, K_v, T_v, Kp and the<br />
corresponding drive parameter K_A, PW, K_B as the<br />
design variables in controlling.<br />
2) Specific constraints <strong>of</strong> the needed design variables<br />
After determining the variables <strong>of</strong> the integrated<br />
design, we introduce the specific constraints for the<br />
variables:<br />
a) Specific constraints <strong>of</strong> variables<br />
The mechanical drive mode <strong>of</strong> the airborne radar pitch<br />
servo system is two-level gear transmission mode. This<br />
is shown in Fig. 4.<br />
10 4<br />
Figure 4. Two-level gear transmission mode<br />
The system has very strict limitations <strong>of</strong> space, so<br />
space constraints <strong>of</strong> the model is an very important<br />
constraint. We assumed that the two-level gear<br />
transmission system is limited in a space <strong>of</strong> an outside<br />
diameter W. The constraints are<br />
⎧2(<br />
r1 + r2) ≤ W<br />
⎨<br />
. (13)<br />
⎩r2<br />
+ r3 + 2r4<br />
≤ W<br />
b) Constraints <strong>of</strong> reduction ratio<br />
The constraints <strong>of</strong> the servo system reduction ratio<br />
take the given initial value as reference to define its<br />
boundary <strong>of</strong> reduction ratio. The constraint conditions <strong>of</strong><br />
the servo system reduction ratio take the given initial<br />
value as reference to define its boundary, namely<br />
⎧ a<br />
⎪<br />
⎨b<br />
⎪<br />
⎩c<br />
−<br />
−<br />
−<br />
≤ i<br />
≤ i<br />
≤ i<br />
12<br />
34<br />
总<br />
= r<br />
= r<br />
= i<br />
2<br />
4<br />
12<br />
/ r ≤ a<br />
/ r<br />
i<br />
34<br />
1<br />
3<br />
≤ b<br />
≤ c<br />
+<br />
+<br />
+<br />
10 6<br />
. (14)
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1909<br />
The boundary <strong>of</strong> the high-level gear transmission ratio<br />
a , .<br />
and c , c refer to the boundary <strong>of</strong> two-gear<br />
and the low-level gear transmission ratio are − a+<br />
b− , b+<br />
− +<br />
transmission ratio respectively.<br />
c) Performance constraints<br />
The steady-state error <strong>of</strong> e in the system should satisfy<br />
the system precision. That is e
1910 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
We can see the results from the table and figures.<br />
Mechanical structure parameters and control parameters<br />
<strong>of</strong> the servo system are matched with each other after the<br />
system is integrated design [9]. The gear <strong>of</strong> the meshing<br />
has changed at the premise <strong>of</strong> meeting the performance<br />
index <strong>of</strong> the system. And the transmission ratio is<br />
re-assigned. The space the gear box used was reduced by<br />
2.46% after the integrated design, meeting requirements<br />
Fittness<br />
E rror/rad<br />
-100<br />
0 10 20 30 40 50<br />
Generation<br />
60 70 80 90 100<br />
for the constraints on space <strong>of</strong> the airborne radar servo<br />
system. In addition, the high-speed level and the<br />
low-speed level backlash in this paper had an increase by<br />
22.8% and 76.7%. The system can contain a bigger<br />
backlash after the integrated design, so that the life <strong>of</strong> the<br />
gear is prolonged, the waste <strong>of</strong> resource and the cost <strong>of</strong><br />
production are reduced. As a result, the combination<br />
property <strong>of</strong> the radar servo system is finally promoted.<br />
Position / rad<br />
1.4<br />
1.2<br />
1<br />
0.8<br />
0.6<br />
0.4<br />
0.2<br />
Reference Signal and Response Signal<br />
Reference Signal<br />
Response Signal<br />
0<br />
0 0.5 1 1.5 2 2.5<br />
Time(s)<br />
3 3.5 4 4.5 5<br />
Figure 6.The best solution and the average solution <strong>of</strong> each generation Figure 7 (a). Optimization response curve<br />
1.2<br />
1<br />
0.8<br />
0.6<br />
0.4<br />
0.2<br />
20<br />
-20<br />
-40<br />
-60<br />
-80<br />
0<br />
0<br />
Fittness and Generation<br />
Tracking Error<br />
the best solution<br />
the average solution<br />
-0.2<br />
0 0.5 1 1.5 2 2.5<br />
time(s)<br />
3 3.5 4 4.5 5<br />
Figure 7(b). Optimization tracking error curve Figure 7(c). Optimization bode diagram<br />
Through the analysis and simulation above, we can<br />
make the conclusions:<br />
a) In order to build the model, this model can reflect<br />
the actual system. When establishing the dynamic model<br />
<strong>of</strong> pitch servo-mechanical drive system in this paper, the<br />
nonlinear factors is considered into the backlash and<br />
backlash model uses a non-linear dead zone model. It is<br />
in line with this system characteristics.<br />
b) For multi-gear transmission characteristics <strong>of</strong> the<br />
servo-mechanical systems, we built the basic unit <strong>of</strong> gear<br />
meshing. Thereby, a mechanical drive system dynamic<br />
model was established. The basic unit used this method<br />
© 2011 ACADEMY PUBLISHER<br />
Magnitude (dB)<br />
Phase (deg)<br />
200<br />
0<br />
-200<br />
-400<br />
-600<br />
-800<br />
0<br />
-360<br />
-720<br />
-1080<br />
-1440<br />
10 -2<br />
10 0<br />
Bode Diagram<br />
10 2<br />
10 4<br />
Frequency (rad/sec)<br />
could easily create multi-stage gear transmission system<br />
model.<br />
c) The former research results in most <strong>of</strong> the studies<br />
only considered the motor control. It nearly regarded the<br />
matching mechanical and electrical parameters and<br />
institutional dynamics parameters on the dynamic<br />
performance <strong>of</strong> servo system. Pitching in the closed loop<br />
servo system <strong>of</strong> radar, the mechanical transmission<br />
system was included in the position loop. The electrical<br />
servo control system was not completely separated with<br />
the subsystems. However a new integrated mechanical<br />
and electrical coupling systems was formed by a<br />
10 6<br />
10 8
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1911<br />
feedback loop. Aimed at the feature <strong>of</strong> this system, the<br />
mechanical structure and control parameters was<br />
considered in the structure/control integrated design<br />
Design<br />
variable<br />
Performance parameter<br />
VI. CONCLUSION<br />
We built the electromechanical coupling model <strong>of</strong> the<br />
airborne radar pitch servo system in this paper. The<br />
coupled model was the optimization model <strong>of</strong> the<br />
structure/control integration design. We used the Genetic<br />
Algorithm to optimize the parameter <strong>of</strong> the integrated<br />
optimization model and received fine impression. It<br />
proved the method used in the task is right and the<br />
practicability <strong>of</strong> Genetic Algorithm. This method can be<br />
used as a new method in servo system design and can be<br />
developed in the long ran.<br />
The model used classical three close loop control<br />
method. The simulation model is built up by dynamic<br />
simulation tool Matlab/Simulink and the simulation<br />
curve lines <strong>of</strong> reflecting the system performances are<br />
acquired. According to the simulation model,the effect<br />
<strong>of</strong> nonlinear factors on system performances is analyzed,<br />
and the measures <strong>of</strong> improving the system performances<br />
are given. The simulation results show the feasibility and<br />
effectiveness <strong>of</strong> the structure/control integration design<br />
in the servo system, and it is the stage for further<br />
research.<br />
Structure and control integration design is not the<br />
design <strong>of</strong> simple superposition between mechanical<br />
module and control module[10]. It analyzes deeply the<br />
coupling <strong>of</strong> structure and control and sets up the coupled<br />
model. Integrated concurrent design is done for the<br />
structure and control parameters. Modern structure with<br />
many complexities in itself has a strong dynamic<br />
coupling, coupled with the control role in the regulation.<br />
The overall performance <strong>of</strong> the system can be achieved<br />
under the combined effects <strong>of</strong> structure and control.<br />
Therefore, the establishment <strong>of</strong> the modern electrical and<br />
© 2011 ACADEMY PUBLISHER<br />
Structure<br />
Design<br />
variable<br />
TABLE Ⅱ. OPTIMIZATION RESULTS<br />
during the system being designed. The simulation proved<br />
that the integrated design method was feasible to design<br />
the radar servo system.<br />
Parameters Original design value Integrated Design value<br />
Control<br />
design variable<br />
Z1 18 19<br />
Z2 54 68<br />
Z3 18 18<br />
Z4 150 129<br />
b1 62.5 76.778<br />
b2 75 132.59<br />
K_i 5 14.682<br />
Tao_i 0.0028 0.36888<br />
K_v 47.876 21.399<br />
T_v 0.32265 0.08367<br />
Kp 59.178 58.585<br />
Ti 789.62 366.47<br />
K_A 0.3333 0.35943<br />
PW 4.6 2.3192<br />
K_B 0.5 0.23749<br />
W 0.2558 0.2495<br />
Mp 5% 0%<br />
tr 0.4 0.21<br />
ts 1 0.8<br />
tp 0.8 0.7<br />
e 1.5e-3 2.8e-4<br />
mechanical system coupling model can better reflect the<br />
actual situation. It will be the focus <strong>of</strong> the future<br />
research.In the future, research and analysis should be<br />
done further deeply in the structure/control integration<br />
design for radar servo system from the following:<br />
a) During the process <strong>of</strong> establishing this model, we<br />
ignore the support bearings and box, and other gear<br />
stiffness and damping and surface friction and other<br />
factors. On establishing the future model, we should take<br />
full account <strong>of</strong> these factors.<br />
b) This radar servo system is a three-axis system. It<br />
contains orientation axis, pitch axis and the<br />
horizontal-roller axis. It only studies the coupling <strong>of</strong> the<br />
mechanical part and the control part in the pitch servo<br />
system. It doesn’t consider the coupling <strong>of</strong> the three-axis.<br />
The future research should be taken into account the<br />
coupling factor <strong>of</strong> the three-axis movement.<br />
REFERENCES<br />
[1] Job van Amerongen,Peter Breedveld. Modelling <strong>of</strong> Physical<br />
Systems for the Design and Control <strong>of</strong> Mechatronic<br />
Systems. Annual Reviews in Control.2003, (27):87-117.<br />
[2] Liu D K,Yang Y L,Li Q S.Optimum positioning <strong>of</strong><br />
actuators in tall buildings using genetic algorithm[J].<br />
<strong>Computers</strong> and Structures,2003,81: 2823-2827.<br />
[3] YU Ling, JIA Chun-qiang. Functions and Examples in<br />
Matlab GA Toolbox. Mechanical Engineer.2004,(11):<br />
27-28.<br />
[4] Lu Jianwei,Zhang Xianmin, Shen Yunwen. Integrated<br />
Structral and Noise Control Design For EIC Linkage<br />
Mechanism. <strong>Journal</strong> <strong>of</strong> Mechanical Engineering.<br />
2003,39(3):40-43.<br />
[5] Giacomini D,Bianconi E,Martino L,Palma M.A now fully<br />
integrated power module for three phase servo motor<br />
driver applications[J].IEEE Industry Applications Society,
1912 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
2001,2:981-987.<br />
[6] IJA RM. FONSECA, PETERM. BAINUM. Integrated<br />
Structure and Control Optimization. <strong>Journal</strong> <strong>of</strong> Vibration<br />
and Control.2004,(10):1377-1391.<br />
[7] Wang G J,Fong C T,Chang K J.Neural-network-based<br />
self-tuning PI controller for precise motion control <strong>of</strong><br />
PMAC motor[J].IEEE Transactions on Industrial<br />
Electronics,2001,48(2):408-415<br />
[8] ZHAO Guo-feng,FAN Wei-hua,CHEN Qing-wei. A<br />
Survey on Backlash Nonlinearity. Acta Armamentarii.<br />
2006,27(6):1072-1080.<br />
[9] Long Kai, Cheng Ying. The Research <strong>of</strong> Parameters by<br />
The Simulation <strong>of</strong> Exciting Force in Gears. Computer<br />
Simulation. 2002,19(6):87-89.<br />
[10] Ahmad Al-shyyab. Non-Linear Dynamic Analysis <strong>of</strong> a<br />
Multi-Mesh Gear Train Using Multi-Term Harmonic<br />
Balance Method.The University <strong>of</strong> Toledo.2003.<br />
[11]Xin M,Balakrishnan S N,Ohlmeyer E J.Integrated guidance<br />
and control <strong>of</strong> missiles with method[J].IEEE Tmns on<br />
Control Systerns Technology.2006,14(6):981-992.<br />
[12] Fawzi Belblidia, Ernest Hinton, fully integrated design<br />
optimization <strong>of</strong> plate structures, Finite Elements in<br />
Analysis and Design 38 (2002) 227-244.<br />
[13] Li Q S,Liu D K,Tang J,Zhang N,Tam C M.Cornbinatorial<br />
optimal design <strong>of</strong> number and positions <strong>of</strong> actuators in<br />
actively controlled structures using genetie algorithms[J].<br />
<strong>Journal</strong> <strong>of</strong> Sound and Vibration,2004, 270:611-624.<br />
[14] Jahng-Hyon Park,Haruhiko Asada.Integrated Structure/C-<br />
ontrol Design <strong>of</strong> a Two-Link Non-rigid Robot Arm for<br />
High Speed Positioning. Proceedings <strong>of</strong> the 1992 IEEE<br />
International Conference on Robotics and Automation,<br />
1992,(5):735-741.<br />
[15] Joseph C.Chen, Jacob Chen. Testing a New Approach for<br />
Learning Teamwork Knowledge and Skills in Technical<br />
Education. Industrial Technology.2004, 20(2): 1-10.<br />
[16] Tan Ping,Dyke S J,Richardson A,et a1.Integrated device<br />
placement and control design in civil structures using<br />
genetic algorithms[J]. <strong>Journal</strong> <strong>of</strong> Structural Engineering,<br />
ASCE,2005,131(10):1489-1496.<br />
[17] David A S,Paul N R,Lin Peiyang.GA-optimized fuzzy<br />
logic control <strong>of</strong> a large-scale building for seismic<br />
© 2011 ACADEMY PUBLISHER<br />
loads[J].Engineering Structures,2007,30(2):436-449.<br />
[18] CUI Ling-li, GAO Li-xin, ZHANG Jian-yu, XIAO<br />
Zhi-quan. Integrated Structure and Control Design for<br />
Flexible Manipulator System. <strong>Journal</strong> <strong>of</strong> Beijing<br />
University <strong>of</strong> Technology.2007,33(8).<br />
[19] George O’ Neal, An Analytical Approach to Integrated<br />
Structural and Control Design, PHD,University <strong>of</strong><br />
Michigan,2001.<br />
[20]Anton C.Pil, haruhiko H. Asada. Integrated<br />
structure/control Design <strong>of</strong> Mechatronic Systems Using<br />
a Recursive Experimental Optimization Method.<br />
IEEE/ASME Transactions on Mechatronics.1996,9,<br />
1(3):191-203.<br />
Dingzhen Li (1972~), female, Nanyang,<br />
Henan province, China. She received the<br />
B.S. degree in automatic control <strong>of</strong><br />
electrical engineering from Northeast<br />
Heavy Machinery Institute, Qiqihaer,<br />
China and the M.S. degree in Test and<br />
Measurement Technology and Instrument<br />
from Nanjing University <strong>of</strong> Science and<br />
Technology, Nanjing, China, respectively.<br />
She is currently an associate pr<strong>of</strong>essor with<br />
Department <strong>of</strong> Electronics and Electrical Engineering, Nanyang<br />
Institute <strong>of</strong> Technology, Nanyang, China. Two books and thirty<br />
papers are published. Her main research interests include<br />
design <strong>of</strong> mechatronic systems and automation control <strong>of</strong><br />
intelligent equipments.<br />
Ruimin Jin (1967~), male, Nanyang, Henan province, China.<br />
Ph.D. Three books and thirty papers are published. He is now a<br />
pr<strong>of</strong>essor in Nanyang Institute <strong>of</strong> Technology. His main<br />
research interests are in the fields <strong>of</strong> in solar cells preparation<br />
and applications.
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1913<br />
A Model to Select System Core and Its<br />
Application<br />
Chongming LI<br />
College <strong>of</strong> Management, Huazhong Normal University, Wuhan, 430079, China.<br />
Email: lichongming@eyou.com.<br />
Yue DING<br />
College <strong>of</strong> Management, Huazhong Normal University, Wuhan, 430079, China.<br />
Email: dingyue_2003@163.com.<br />
Abstract—The theory <strong>of</strong> system core is given a method to<br />
determine key elements <strong>of</strong> the system based on graph<br />
theory, but it is difficult to apply in practice for the question<br />
<strong>of</strong> how to change a system to a graph and multi-core in a<br />
system. This paper gives a method to change a system to a<br />
graph based on correlation analysis, and gives a modle to<br />
select system core based on cluster analysis. More, in the<br />
case <strong>of</strong> the real estate system <strong>of</strong> Wuhan, a diffusion index<br />
curve is given based on the elements <strong>of</strong> system core selected<br />
by the model, the result illustrates that Wuhan real estate<br />
arises in 1991 and to its peak in 1993, and then declines to<br />
the bottom in 1996; next, it fluctuates in a small range and<br />
becomes smooth since 1999, but a slight upward trend<br />
during 2000, this conclusion is consistent with the actual<br />
development status <strong>of</strong> the real estate in Wuhan and prove<br />
the validity <strong>of</strong> the model.<br />
Index Terms—System Core; Cluster Analysis; Correlative<br />
Analysis; Real Estate System; Wuhan<br />
I. INTRODUCTION<br />
When making research on system problem, it usually<br />
draws support from some indexes to analyze the whole<br />
system, that is to say, to establish an index system<br />
describing the system. But for many systems, especially<br />
those complex systems, they involve a lot <strong>of</strong> elements<br />
which have extremely complex relations within<br />
themselves. In order to study the system, it has to analyze<br />
the elements and relations <strong>of</strong> the system, for the structure<br />
and function <strong>of</strong> system determined by them. Under a<br />
complete system information situation, it can choose<br />
those indexes which related to the system as many as<br />
possible, but along with indexes increase, the redundant<br />
information which has nothing to do with the system also<br />
will increase. Besides, redundant information will<br />
submerge which we needed; meanwhile it will increase<br />
the analysis and computation difficulties. Because the<br />
elements <strong>of</strong> system play different roles in the system (XU<br />
Jin, WANG Yingluo, 1993), some are very important and<br />
some are unimportance.There are a large number <strong>of</strong><br />
objective facts in nature and human society show that any<br />
system have some key elements, key elements <strong>of</strong> system<br />
play a dominant role to the system, so it hopes to discover<br />
the essential elements which have the key role through<br />
© 2011 ACADEMY PUBLISHER<br />
doi:10.4304/jcp.6.9.1913-1919<br />
some methods, and then uses these key elements to study<br />
the system. The theory <strong>of</strong> system core using the<br />
knowledge <strong>of</strong> graph theory to present a method that can<br />
determine the essential elements <strong>of</strong> system, which is also<br />
called the core <strong>of</strong> the system (XU Jin,1993), by the theory<br />
it can discover the essence, the main body and the key<br />
elements <strong>of</strong> the system.<br />
The theory was applied in complex system (WANG<br />
Jingguang, 2001), fault diagnosis (CAI Bing, ZHOU<br />
Liuding, 1994) and reliability <strong>of</strong> communication (CAO<br />
Qiguo, SUN Yugeng, 1997).But it's few used in<br />
social-economic systems, for the relations <strong>of</strong> elements in<br />
social-economic systems is too complex to explain by<br />
vertex cut sets and components, it's hard to be turned a<br />
social-economic system into a graph. Another problem is<br />
the lack <strong>of</strong> uniqueness to the system core, there is no<br />
explain to how a system core important than another in<br />
system core theory, many research describe the destroy <strong>of</strong><br />
connected graph by vertex cut sets and components based<br />
on toughness <strong>of</strong> a graph (Chvata1 ,1973), then denote the<br />
importance <strong>of</strong> vertexes through the damage <strong>of</strong> the<br />
connected graph, just as relative rupture <strong>of</strong><br />
graph(OUYANG Kezhi, OUYANG Keyi,1993) and the<br />
relation between degree <strong>of</strong> rupture and rupture<br />
number(ZHANG Shenggui, WANG Ziguo,1995), which<br />
give academic base to settle the question <strong>of</strong> multi-core.<br />
This paper gives a method to turn a system into a graph<br />
and formulates a model to select system core within the<br />
cluster analysis and grey correlative analysis, then the<br />
model was used in the real estate system <strong>of</strong> Wuhan, the<br />
conclusion according with the actual development status<br />
<strong>of</strong> the real estate <strong>of</strong> Wuhan, thereby demonstrating the<br />
validity <strong>of</strong> the model.<br />
II. SYSTEM CORE AND ITS MULTIPLE VALUED<br />
The system core theory given a method to study<br />
complex systems, it is describe the system center through<br />
the system core by qualitative and quantitative method. It<br />
is regard system as graph in system core theory and the<br />
system core is composed <strong>of</strong> vertex cut set which is<br />
important or has dominant role to the system function, it
1914 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
is used the number <strong>of</strong> sets and connected components to<br />
calculate system core.<br />
The basic idea <strong>of</strong> system core theory is that such<br />
component elements couldn’t have the same effect on a<br />
given system, while some elements are minor and some<br />
are very important to the system (XU Jin, 1993).<br />
Eliminating or destroying these key elements will make<br />
the system break down. These several essential elements<br />
are called the core <strong>of</strong> this system.<br />
Suppose that X is a system, its elements<br />
are n x x x ,..., , 1 2 , if x i has relation with x j in X ,<br />
then denote by x i x j , and connect this two with an<br />
edge e ij . In that way, we can use a graph to express the<br />
system X. The vertex set can represent element(index) <strong>of</strong><br />
X, edges in the graph express the relations <strong>of</strong> them, then<br />
structuring out graph G <strong>of</strong> system X , its vertex set<br />
is ( ) { 1, 2,...,<br />
n}<br />
x x x G V = , edge set<br />
is E ( G)<br />
= { xix<br />
j | xiandxj<br />
have relation } , in practice the<br />
relations between xi and x j are broad, they depend on<br />
the nature <strong>of</strong> researched system and the problem.<br />
Definition1. Suppose that G is a connected<br />
graph, V ( G)<br />
≥ 4 , then<br />
h( G)<br />
= max{ ω ( G − S)<br />
− S ; S ⊂ C(<br />
G)}<br />
( 1<br />
)<br />
h (G)<br />
is called core degree which is the value <strong>of</strong><br />
system core <strong>of</strong> system graph G , C (G)<br />
are all vertex<br />
cut sets in graph G , ω ( G − S)<br />
reflect the branches<br />
when graph G is cut <strong>of</strong>f by vertex cut sets S ; S is<br />
the number <strong>of</strong> vertex in vertex cut sets S .<br />
∗<br />
If there are vertexes cut sets S satisfied<br />
∗ ∗<br />
h( G)<br />
= ω ( G − S ) − S ;<br />
(2)<br />
then ∗<br />
S is the core <strong>of</strong> systemG .<br />
∗<br />
S are the vertex-cut sets that satisfy the value <strong>of</strong> core<br />
degree, core degree meaning in graph theory is the most<br />
destructive measurement for graphG , so the definition<br />
<strong>of</strong> system core is based on that <strong>of</strong> vertex-cut sets and the<br />
connected graph, thus it first requests that graph G is<br />
connected, at the same time, there is not core in<br />
full-connected graph K n for the lack <strong>of</strong> vertex-cut sets<br />
in K n .<br />
From the definition <strong>of</strong> system core, any system which<br />
contains a binary relation can regard as a graph, vertexes<br />
are elements <strong>of</strong> the system, and edges are relations <strong>of</strong><br />
elements. Since the born <strong>of</strong> system core theory, it has<br />
developed a lot in theory as well as practice, however, for<br />
the lack <strong>of</strong> uniqueness for most <strong>of</strong> the systems, there<br />
exists a problem concerning the option <strong>of</strong> core in<br />
© 2011 ACADEMY PUBLISHER<br />
application. As following two simple graphs, 1 G<br />
andG 2 .<br />
Figure1. Simple system graph<br />
from definition 1, it's know easily that the core degree <strong>of</strong><br />
G 1 is 2, namely ( 1 ) 2 = G h ,the core <strong>of</strong> G 1 is unique,<br />
∗<br />
S = { x1}<br />
; for G 2 , ( 2 ) 1 = G h ,but there are 3<br />
∗<br />
∗<br />
∗<br />
cores, S 1 = { x2}<br />
, S 2 = { x3}<br />
, S 3 = { x2<br />
, x3}<br />
.<br />
Because the core are the vertex-cut sets <strong>of</strong> graph, so for<br />
some graphs <strong>of</strong> complex system, the number <strong>of</strong> the core<br />
is more than that <strong>of</strong> the vertex is also possible, under such<br />
circumstances, the most commonly used method is based<br />
on the nature <strong>of</strong> studied system and the studied problem<br />
to choose which core to analyze the system, if we want to<br />
know which core can mostly reflect the nature <strong>of</strong> system<br />
and solve the studied system problem, it need to analyze<br />
every core, if the number <strong>of</strong> core is more than that <strong>of</strong> the<br />
vertex, then taking advantage <strong>of</strong> system core to analyze<br />
the system is not only failed to simplify the problem, but<br />
make it more complex, there it need another kind <strong>of</strong><br />
method to solve the problem, namely find out the core<br />
that can solve the studied system problem.<br />
III. GRAPH OF SYSTEM AND CLUSTER OF INDEXES<br />
Although there has a specific algorithm to solve the<br />
question <strong>of</strong> system core Identification, but the algorithm<br />
is only used to small connected graph and it is difficult to<br />
be used in complex system, In addition, system core<br />
theory and its specific applications main used in the<br />
communication network, the function communication<br />
network is transfer Information and connectivity is the<br />
basis <strong>of</strong> transfer Information, So the dominant vertex is<br />
core. However, much relations <strong>of</strong> social network is<br />
difficult to describe by vertex cut sets and connected<br />
graph, such as friendship, trust, acquaintance, like or<br />
other relations.<br />
In the course <strong>of</strong> resolving some problem about system,<br />
especially, socioeconomic system, it can transform the<br />
system into the form <strong>of</strong> graph. In the process <strong>of</strong><br />
constructing graphs, the key <strong>of</strong> a system graph is how to<br />
make sure the edge between vertices (system index), as to<br />
a system, the relation between indexes should reflect<br />
integrity <strong>of</strong> system at first that is the correlation between<br />
indexes. So the graph <strong>of</strong> a system can be determined by<br />
the method <strong>of</strong> associative analysis.
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1915<br />
As to indexes xi ( i = 1,<br />
2,...,<br />
n)<br />
taken as<br />
characterization <strong>of</strong> system X , there is relation<br />
between xi and x j which can be seen from the integrity<br />
<strong>of</strong> system (MIAO Dongsheng, 1998), the relation<br />
between xi and x j can be expressed by gray correlative<br />
analysis, gray correlative analysis is the measurement <strong>of</strong><br />
relationship between two systems or two elements, and it<br />
describes the status including the magnitude <strong>of</strong> change,<br />
comparative change in direction and velocity, and so on <strong>of</strong><br />
comparative change between elements in the process <strong>of</strong><br />
system development. If the trend <strong>of</strong> change between two<br />
elements is consistent in the process <strong>of</strong> development, the<br />
two elements change in a high degree <strong>of</strong> synchronization,<br />
the degree <strong>of</strong> relationship is comparatively high, vice<br />
versa. Gray correlative analysis is a kind <strong>of</strong> method with<br />
which can be used to analyze and determine the degree <strong>of</strong><br />
mutual effects between system elements and the degree in<br />
which elements contribute to system.<br />
The basic idea <strong>of</strong> gray correlative analysis is to<br />
determine the behavior <strong>of</strong> a system, find out numeric<br />
array <strong>of</strong> behavior, search for the elements which affect<br />
the behavior <strong>of</strong> system, collect the element data arrays<br />
affecting behavior <strong>of</strong> system, calculate the correlative<br />
degree between data array <strong>of</strong> each element and data array<br />
<strong>of</strong> behavior, so the relation between xi and x j can be<br />
expressed by correlation coefficient r ij<br />
r<br />
ij<br />
=<br />
∑<br />
k=<br />
1<br />
m<br />
∑<br />
k=<br />
1<br />
m<br />
( x ( k)<br />
− x )( x ( k)<br />
− x )<br />
( x ( k)<br />
− x )<br />
i<br />
i<br />
i<br />
i<br />
2<br />
m<br />
∑<br />
k=<br />
1<br />
j<br />
( x ( k)<br />
− x )<br />
j<br />
j<br />
j<br />
2<br />
(3)<br />
here, i , j = 1,<br />
2,...,<br />
n , is number <strong>of</strong> indexes <strong>of</strong> system;<br />
k = 1,<br />
2,...,<br />
m is number <strong>of</strong> indexes data; x i , x j is<br />
average <strong>of</strong> indexes data.<br />
Using formula (3), correlation coefficient r ij between<br />
xi and x j <strong>of</strong> system X can be determined, the<br />
correlation matrix <strong>of</strong> all indexes is R .<br />
⎡r11<br />
⎢<br />
⎢<br />
r21<br />
R =<br />
⎢ ...<br />
⎢<br />
⎣r1n<br />
r12<br />
r22<br />
...<br />
r2n<br />
...<br />
...<br />
...<br />
...<br />
1 , r = r , i,<br />
j = 1,<br />
2,...,<br />
rii ij ji<br />
here, = n .<br />
r1n<br />
⎤<br />
r<br />
⎥<br />
2n<br />
⎥<br />
... ⎥<br />
⎥<br />
rnn<br />
⎦<br />
In the process <strong>of</strong> constructing graph <strong>of</strong> a system, the<br />
relation between xi and x j is generalized, in this paper,<br />
according to the character <strong>of</strong> system and the problem <strong>of</strong><br />
system, it chooses a critical value r0 <strong>of</strong> a correlation<br />
coefficient, when rij ≥ r0<br />
, there is correlation between<br />
© 2011 ACADEMY PUBLISHER<br />
xi and x j , using an edge to join<br />
index i, j correspondingly. So it can gain graph G <strong>of</strong><br />
system X. In order to make sure the graph <strong>of</strong> system is<br />
connected, choose the smallest correlation coefficient r c<br />
that can make all indexes <strong>of</strong> system to be a connected<br />
graph as critical value.<br />
As it can be seen from the correlation matrix R <strong>of</strong><br />
indexes, some correlation coefficient are very big, the<br />
other are opposite smaller. So indexes with big<br />
correlation coefficient have strong correlation in system,<br />
then indexes can be grouped by clustering <strong>of</strong> index<br />
correlation: according to correlation matrix R , it's can<br />
search from the first row to find out r1 j ( j = 1,<br />
2,...<br />
n )<br />
bigger than r c ( r1 j ≥ rc<br />
), and take the indexes xi , x1<br />
as<br />
one group, for example, if r12 , r15,<br />
r19<br />
are all bigger than<br />
x , x , x , x can be as one group, then<br />
r c ,the indexes 1 2 5 9<br />
searching from the second row to find out r2 j ≥ rc<br />
, then<br />
indexes xi , x2<br />
as one group, and so on, till last row <strong>of</strong><br />
R ; then take the index with the smallest tab as<br />
representative stands for each group, if two group has the<br />
same index, merge the primary classification, for<br />
example, in the primary classification, x1 , x2<br />
, x5,<br />
x9<br />
is<br />
as one group, x7 , x9<br />
, x12<br />
is as another group, here<br />
, x , x<br />
x , finally take<br />
merge x7 9 12 into the group in which 1<br />
the indexes unlisted as one group respectively, by this<br />
way, n indexes can be divided to M groups, each index<br />
aggregate J i ( J i ∩ J j = Φ,<br />
i,<br />
j = 1,<br />
2,<br />
3,...,<br />
M )reflect<br />
system from one side, so index system which used to<br />
describe system should include one index per group at<br />
least.<br />
IV. MODELTO SELECT SYSTEM CORE<br />
The aim <strong>of</strong> turning a actual system into a graph is<br />
taking advantage <strong>of</strong> the system core theory to identify the<br />
key elements and simplify the indexes system <strong>of</strong> complex<br />
system. Therefore, after getting the system graph G by<br />
correlation analysis, we need to calculate the core <strong>of</strong><br />
G .According to the formula (1) and (2), it can get the<br />
core <strong>of</strong> graph G , but the core is uniqueness. The<br />
definition <strong>of</strong> the core shows that the core is the main body<br />
<strong>of</strong> the system and it can completely describe the system,<br />
at the same time, because <strong>of</strong> the clustering process <strong>of</strong><br />
system index, needing at least one in every kind <strong>of</strong> index<br />
to describe the system, if we want to completely describe<br />
a system, so, in the case that the core is not the only one,<br />
∗<br />
0<br />
S should satisfy<br />
∗<br />
S ≥<br />
∗<br />
0 M And S0 ∩ J i ≠ Φ (4)
1916 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
∗<br />
S 0 is the core value (the number <strong>of</strong> core element)<strong>of</strong><br />
the system, M represents the kind number <strong>of</strong> indexes<br />
cluster when the critical value <strong>of</strong> correlation coefficient<br />
is r c , J i is the indexes set <strong>of</strong> class i , i = 1,<br />
2,<br />
3,...,<br />
M .<br />
It is can combine the correlation clustering with the<br />
core theory to solve the problem <strong>of</strong> multi-core. but in<br />
application, there is another problem that should be<br />
solved ,namely whether it's can find out the core that<br />
satisfy formula (4), if system X do not contain the core<br />
that satisfy formula (4), indexes will be farther clustered,<br />
if<br />
xk ∈ A , xl ∉ A , A ∉ M<br />
there is min( r kl ) , xl ∈ B ,take A and B as one<br />
cluster, then cluster indexes <strong>of</strong> system into M ′<br />
categories( M ′ < M ),so it's can get a new clustering,<br />
seeing if there is a core<br />
then need further clustering till finding out the core<br />
∗<br />
∗<br />
S 0 , constitute elements <strong>of</strong> S 0 are the key ones <strong>of</strong><br />
studied system.<br />
∗<br />
S 0 satisfy formula (4), if not,<br />
V. A CASE STUDY IN REAL ESTATE SYSTEM OF<br />
WUHAN<br />
In this paper, the model <strong>of</strong> select system core is applied<br />
to the real estate system <strong>of</strong> Wuhan, establishing the index<br />
system <strong>of</strong> Wuhan real estate system, numbering these<br />
indexes(YE Yanbing, DING Lieyun, 2001) ,and getting<br />
index data from the year 1990 to 2001 as TABLE I.<br />
In order to make every index comparable, A<br />
dimensionless treatment is needed, it is use the method <strong>of</strong><br />
initial value, and it’s very simple and easily<br />
understandable.<br />
Suppose that there is an original numeric array listed as<br />
follows<br />
( 0)<br />
( 0)<br />
( 0)<br />
( 0)<br />
x ( i)<br />
= { x ( 1),<br />
x ( 2),...,<br />
x ( n)}<br />
( 1)<br />
( 0)<br />
x ( i)<br />
is produced by x ( i)<br />
which processed with<br />
initial value method by formula (5).<br />
( o ) ( o )<br />
( o )<br />
( 1)<br />
x ( 1)<br />
x ( 2)<br />
x ( n)<br />
x ( i)<br />
={ ( o ) , ( o ) ,..., ( o ) }<br />
x<br />
( 1)<br />
x<br />
( 1)<br />
x<br />
( 1)<br />
x x x n<br />
(5)<br />
( 1)<br />
( 1)<br />
( 1)<br />
= { ( 1),<br />
( 2),...,<br />
( )}<br />
using formula ( 3 ) to get the related coefficient<br />
between indexes, when r 0.<br />
85 , take indexes as<br />
c<br />
vertexes v i , the relations <strong>of</strong> indexes as edges e ij , it's can<br />
get a connected graph <strong>of</strong> Wuhan real estate system, as<br />
Figure2.<br />
Take advantage <strong>of</strong> formula (1) and (2), calculating out<br />
that the value <strong>of</strong> system core is 6, obtaining 43 cores as<br />
TABLE II.<br />
According to the method <strong>of</strong> clustering, when the real<br />
estate coefficient critical Value r 0 = 0.<br />
85 , obtaining<br />
one clustering from the index system <strong>of</strong> the real estate<br />
© 2011 ACADEMY PUBLISHER<br />
=<br />
in Wuhan,{1, 7, 15}, {2, 6, 9, 13}, {3, 4, 8}, {5,<br />
10,11,12,14,16}.<br />
Compared to the clustering <strong>of</strong> indexes with the core in<br />
TABLE II, it can get the core which satisfies formula(4<br />
)is C31, indexes included in C31 can be substituted for<br />
the ones in table 1 to analyze the real estate system <strong>of</strong><br />
Wuhan. It takes advantage <strong>of</strong> the Diffusion Index(DI) to<br />
describe the situation about the real estate system in<br />
Wuhan city from the year 1990 to 2001. It also verifies if<br />
the indexes included in C31 can describe the actual<br />
situation <strong>of</strong> real estate system in Wuhan, according to the<br />
indexes from C31, it gets the Diffusion Index curve as<br />
Figure 3.<br />
TABLE I. INDEXES DATA OF WUHAN REAL ESTATE SYSTEM<br />
Number 1 2 3 4<br />
Index Investment Residence Quantity <strong>of</strong> Construction<br />
Year<br />
investment Employment area<br />
1990 33452 84525 6793 641<br />
1991 34190 92853 7256 624<br />
1992 55821 112277 7576 671<br />
1993 246322 245672 12827 924<br />
1994 570618 470834 10652 1310<br />
1995 916194 630835 11173 1560<br />
1996 972609 678365 12743 1597<br />
1997 1065729 693015 9981 1585<br />
1998 1067997 814564 9894 1655<br />
1999 967130 814564 16049 1741.3<br />
2000 1013105 975576 11211 2086.5<br />
2001 1153400 861100 12421 2836.9<br />
CONTNUED TABLE I<br />
Number 5 6 7 8<br />
Index Housing Trade area Trade value House<br />
completed<br />
price index<br />
Year building area<br />
1990 534 27 10315 99.4<br />
1991 507.1 39 16025 98.9<br />
1992 488.5 82 39940 174.3<br />
1993 549.2 93 63601 132.8<br />
1994 692.6 179 113956 99.9<br />
1995 844.6 285 158145 109.9<br />
1996 1006.2 336 691315 208.6<br />
1997 1180.3 286 255678 98.3<br />
1998 1260.5 676 675331 98.1<br />
1999 1294.5 735 537350 104.4<br />
2000 1274.3 774 622500 110.8<br />
2001 1255.78 892 741400 100.1
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1917<br />
CONTNUED TABLE I<br />
Number 9 10 11 12<br />
Index Total real Total Per capita Consume<br />
estate GDP GDP income price<br />
Year<br />
index<br />
1990 2.12 176.83 1555.8 103<br />
1991 2.85 207.95 1771.68 107.3<br />
1992 3.47 255.42 2116.9 111.4<br />
1993 5.17 357.23 2872.9 119.8<br />
1994 8.96 485.76 3769.8 126.3<br />
1995 10.24 606.91 4453.9 118.4<br />
1996 12.7 782.13 4915.86 112.2<br />
1997 14.52 912.33 5573.04 103.1<br />
1998 24.14 1015.89 5912.52 97.4<br />
1999 26.32 1085.68 6198.24 96.1<br />
2000 30.28 1206.84 6953.94 100.6<br />
2001 37.08 1347.8 7305.05 99.5<br />
CONTNUED TABLE I<br />
Number 13 14 15 16<br />
Index Capital Land sale Construction Per capita<br />
cost area industry living space<br />
Year<br />
output value<br />
1990 142665 89.42 207581 6.1<br />
1991 177894 98.8 237301 6.2<br />
1992 324181 130.8 303877 6.3<br />
1993 471368 293.05 502134 6.5<br />
1994 751348 275.72 795827 6.9<br />
1995 1114959 142.64 1071878 7.2<br />
1996 1353224 161.6 1113665 7.5<br />
1997 1220022 282.17 1088014 7.8<br />
1998 1297707 190 1165540 8.1<br />
1999 1322586 196.54 1352129 8.6<br />
2000 1411232 212.6 1553941 8.8<br />
2001 1653600 223.1 1098300 9.65<br />
From Figure 3, it is show that the development <strong>of</strong><br />
real estate <strong>of</strong> Wuhan coincides with the development <strong>of</strong><br />
the real estate in China, from1990 to 2002, the real estate<br />
market <strong>of</strong> Wuhan experienced a complete cycle which<br />
has 4 period, underway development <strong>of</strong> real estate in<br />
years 1978~1991. Since 1991, there is a quick<br />
development <strong>of</strong> real estate in China, for the development<br />
<strong>of</strong> real estate in the city near the sea and the increase <strong>of</strong><br />
investment from abroad. Then there is a quick<br />
development <strong>of</strong> real estate in Wuhan, due to quick arising<br />
<strong>of</strong> real estate price. Real estate is excessive development<br />
and bubbles in real estate <strong>of</strong> China which leads the<br />
overheated economy, so The Chinese Government began<br />
to deflate money and restrict credit <strong>of</strong> real estate. The<br />
Chinese real estate has entered the adjustment period<br />
since the last <strong>of</strong> 1993, then the real estate <strong>of</strong> Wuhan<br />
© 2011 ACADEMY PUBLISHER<br />
decline from the peak in 1993 to the bottom in1996.<br />
Since the early <strong>of</strong> 1998, the government <strong>of</strong> China<br />
published series policies to arouse the development <strong>of</strong><br />
real estate, the real estate <strong>of</strong> China has entered the golden<br />
age with a steady development, so the real estate <strong>of</strong><br />
Wuhan has a steady development, since 1999 and a slight<br />
upward trend since2000.<br />
The result shows that Wuhan city cycle in real estate is<br />
general agreement with the national real estate cycle, this<br />
shows that elements which were chosen can be used to<br />
describe the real estate system <strong>of</strong> Wuhan, so the model to<br />
select system core is effective.<br />
TABLE II. SYSTEM CORES OF WUHAN REAL ESTATE SYSTEM<br />
Core Index<br />
C1 3,4,10,11,13,14<br />
C2 3 , 4,5,10,11,13,14<br />
C3 3, 4, 6 ,10 ,11, 13,14<br />
C4 3, 4,8 ,10, 11, 13,14<br />
C5 3 , 4 ,9,10,11,13,14<br />
C6 3, 4,10,11,12,13,14<br />
C7 3, 4,10,11,13,14,16<br />
C8 3,4,5,6,10,11,13,14<br />
C9 3,4,5,8, 10,11,13,14<br />
C10 3,4,5,8,11,13,14,16<br />
C11 3,4,5,9,10,11,13,14<br />
C12 3,4,5,10,11,12,13,14<br />
C13 3,4,5,10,11,13,14,16<br />
C14 3,4,6,8,10,11,13,14<br />
C15 3,4,6,10,11,12,13,14<br />
C16 3,4,6,10,11,13,14,16<br />
C17 3,4, 8,9,10,11,13,14<br />
C18 3,4,8,10,11,12,13,14<br />
C19 3,4, 8,10,11,13,14,16<br />
C20 3,4, 9,10,11,12,13,14<br />
C21 3,4,9,10,11,13,14,16<br />
C22 3,4,10,11,12,13,14,16<br />
C23 3,4,5,6,8,10,11,13,14<br />
C24 3,4,5,6,8,11,13,14,16<br />
C25 3,4,5,6,10,11,12,13,14<br />
C26 3,4,5,6,10,11,13,14,16<br />
C27 3,4,5,8,9, 10,11,13,14<br />
C28 3,4,5,8,9,11,13,14,16<br />
C29 3,4,5,8,10,11,12,13,14<br />
C30 3,4,5,,8,10,11,13,14,16<br />
C31 3,4,5,8,11,13,14,15,16<br />
C32 3,4,5,9,10,11,12,13,14<br />
C33 3,4,5,9,10,11,13,14,16<br />
C34 3,4,5,10,11,12,13,14,16<br />
C35 3,4,6,8,10,11,12,13,14<br />
C36 3,4,6,8,10,11,13,14, 16<br />
C37 3,4,6,10,11,12,13,14,16<br />
C38 3,4,8,9,10,11,12,13,14<br />
C39 3,4,8,9,10,11,13,14, 16<br />
C40 3,4,8,10,11,12,13,14,16<br />
C41 3,4,9,10,11,12,13,14,16<br />
C42 3,4,5,6,8,10,11,12,13,14<br />
C43 3,4,5,8,9,10,11,12,13,14
1918 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
Figure 2. Connected graph <strong>of</strong> Wuhan real estate system<br />
Figure 3. The diffusion index curve <strong>of</strong> real estate <strong>of</strong> Wuhan in<br />
1990-2001<br />
VI. CONCLUSIONS<br />
Commonly, any one complex system has multiple<br />
cores and each core describes the system in different way,<br />
but according to system problems to be studied, it is can<br />
find a perfect core to study the system based on some<br />
method and model. From the case study in real estate<br />
system <strong>of</strong> Wuhan, the core <strong>of</strong> real estate system <strong>of</strong><br />
Wuhan can be used to study the real estate system, for it<br />
is accords with the actual development status <strong>of</strong> the real<br />
estate <strong>of</strong> Wuhan and China, this means that the system<br />
can be replaced by the elements which are called system<br />
core, thereby demonstrating the validity <strong>of</strong> the model to<br />
select system core.<br />
With the development <strong>of</strong> system core theory, it will be<br />
widely used in the social and economic system. As an<br />
effective way to simplify the complex system, many<br />
economic problems can be studied and solved by the<br />
system core theory, just as the theory can be used in<br />
researching the stability <strong>of</strong> the society and the structure <strong>of</strong><br />
economic system(WANG Jingguang,2001), core<br />
© 2011 ACADEMY PUBLISHER<br />
competence <strong>of</strong> enterprise(ZHAO Binxin, ZHAO<br />
Jinghua,2000), etc.<br />
REFERENCES<br />
[1] Xu jin. The theory <strong>of</strong> system core and its application.<br />
XiDian University Press,1994.<br />
[2] W. Duckworth, B. Mans. Connected domination <strong>of</strong> regular<br />
graphs. Discrete Mathematics, Volume 309, Issue 8, 28<br />
April 2009, Pages 2305-2322.<br />
[3] Bruno Esc<strong>of</strong>fier, Laurent Gourvès, Jérôme Monnot.<br />
Complexity and approximation results for the connected<br />
vertex cover problem in graphs and hypergraphs .<strong>Journal</strong><br />
<strong>of</strong> Discrete Algorithms, Volume 8, Issue 1, March 2010,<br />
Pages 36-49.<br />
[4] CAO Qiguo,SUN Yugeng. The Hypergraph Design<br />
Method <strong>of</strong> Multibus Structures <strong>of</strong> Reliable<br />
Communication Networks. Acta Electronica Sinica,<br />
1997,(10):55-62.<br />
[5] Konstantin Avrachenkov, Vivek Borkar, Danil<br />
Nemirovsky. Quasi-stationary distributions as centrality<br />
measures for the giant strongly connected component <strong>of</strong> a<br />
reducible graph .<strong>Journal</strong> <strong>of</strong> Computational and Applied<br />
Mathematics, Volume 234, Issue 11, 1 October 2010,<br />
Pages 3075-3090.<br />
[6] OUYANG Kezhi, OUYANG Keyi. Relative Breaktivity <strong>of</strong><br />
Graphs. <strong>Journal</strong> <strong>of</strong> Lanzhou University(Natural Science<br />
Edition), 1993(3):78-82.<br />
[7] ZHANG Shenggui, WANG Ziguo. On Using Concept <strong>of</strong><br />
Degree <strong>of</strong> Rupture for Designing Reliable Network.<br />
<strong>Journal</strong> <strong>of</strong> northwestern polytechnical universty,<br />
1995(2):310-313.<br />
[8] Firdovsi Sharifov, Hakan Kutucu. Minimum Cost ≤k<br />
Edges Connected Subgraph Problems.Electronic Notes in<br />
Discrete Mathematics, Volume 36, 1 August 2010, Pages<br />
25-32.<br />
[9] Hu Xuefeng. On the establishment and improvement <strong>of</strong> the<br />
statistical indexes system <strong>of</strong> Real estate, The <strong>Journal</strong> <strong>of</strong><br />
ShanXi Einance and Economics University, 2002.2,86-89.<br />
[10] Ye Yanbing. Ding Lieyun, Design and study <strong>of</strong> real estate<br />
early warning indexes system. Optimization <strong>of</strong> Capital<br />
Construction, 2001.3,1-3.<br />
[11] Wang Xiaobo. Study on economy cycle and early warning.<br />
Press <strong>of</strong> Metallurgy Industry,1993.<br />
[12] Xu jin, XI Youmin, WANGYingluo. system core and core<br />
degree(I). <strong>Journal</strong> <strong>of</strong> Systems Science and Mathematical<br />
Sciences 1993,(02):20-28.<br />
[13] W. Ananchuen, N. Ananchuen, R.E.L. Aldred. The<br />
structure <strong>of</strong> 4-γ-critical graphs with a cut vertex .Discrete<br />
Mathematics, Volume 310, Issues 17-18, 28 September<br />
2010, Pages 2404-2414.<br />
[14] SHOU Jilin, LI Fei, Point Weighted Core and Coritivity <strong>of</strong><br />
Network System and Its Applications, Systems<br />
engineering--theory and practice, 1996(6):58-63.<br />
[15] WANG Jingguang. Study <strong>of</strong> the Relation Between<br />
Reliability & Complexity <strong>of</strong> Information Systems<br />
Structure. Measurement & Control Technology,<br />
2001,(02):26-34.<br />
[16] ZHAO Bingxin; ZHAO Jinghua. Researching on the Core<br />
Competence with Networks. Chinese <strong>Journal</strong> <strong>of</strong><br />
Management Science, 2000,(S1):45-51.<br />
[17] Michael R. Fellows, Guillaume Fertin, Danny Hermelin,<br />
Stéphane Vialette. Upper and lower bounds for finding<br />
connected motifs in vertex-colored graphs .<strong>Journal</strong> <strong>of</strong><br />
Computer and System Sciences, In Press, Corrected<br />
Pro<strong>of</strong>, Available online 3 August 2010.
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1919<br />
[18] Stephen P. Borgatti.Identifying sets <strong>of</strong> key players in a<br />
social network.Computational & Mathematical Organization<br />
Theory, 2006, Volume 12, Number 1, Pages 21-34.<br />
[19] Yong Yeon Shin and Jai Sang Koh.An algorithm for<br />
generating minimal cutsets <strong>of</strong> undirected graphs. <strong>Journal</strong> <strong>of</strong><br />
Applied Mathematics and Computing, 1998, Volume 5, Number<br />
3, Pages 681-693.<br />
[20] Chang C. Y. Dorea and Ary V. Medino.Anomalous<br />
Diffusion Index for Lévy Motions.<strong>Journal</strong> <strong>of</strong> Statistical<br />
Physics, 2006, Volume 123, Number 3, Pages 685-698.<br />
Chongming LI earned a B.S. in Electrical Technology from<br />
Shandong University <strong>of</strong> Technology in 1995, M.S. in<br />
Philosophy <strong>of</strong> Science and Technology from Wuhan University<br />
<strong>of</strong> Technology in 2001, and Ph.D. in Systems Engineering from<br />
© 2011 ACADEMY PUBLISHER<br />
Huazhong University <strong>of</strong> Science and Technology in 2005. He is<br />
Associate Pr<strong>of</strong>essor <strong>of</strong> the College <strong>of</strong> Management, Huazhong<br />
Normal University.<br />
His current research interests include Information<br />
Management, Land Resource Management.<br />
Yue DING earned a B.S. in Accounting and Auditing from<br />
Wuhan University in 1992, M.S. in Industrial Economy from<br />
Wuhan University in 1998, She is currently working toward<br />
her Ph.D. degree in Economic Management at Zhongshan<br />
University.<br />
Her current research interests include Information<br />
Management, Land Resource Management.
1920 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
De-noise Comprehensive Research<br />
On Airplane Cockpit Signals Recorded by CVR<br />
Dao-Lai Cheng<br />
College <strong>of</strong> Urban Construction and Safety Engineering, Shanghai Institute <strong>of</strong> Technology, Shanghai, 201418, China<br />
daolaicheng@163.com<br />
Chui-JieYi<br />
Qingdao R&D Center <strong>of</strong> Energy Equipment, Qingdao Technological University, Qingdao, China<br />
chuijieyi@vip.163.com<br />
Hong-Yu Yao<br />
Aviation Safety Technical Center, General Civil Aviation Administration <strong>of</strong> China (CAAC), Beijing, China<br />
yaohy@mail.castc.org.cn<br />
Abstract—The characteristic <strong>of</strong> cockpit sound recorded by<br />
CVR is the key evidence in investigating accident causes for<br />
wrecked aircraft. However, cockpit signals ( or CVR sound<br />
information) are complex, they include crew's voices (or<br />
pilot's voices), environmental noise and different kinds <strong>of</strong><br />
backgrounds sound signals , and many factors from inside<br />
and outside cockpit affects the analysis results, especially<br />
noise. To obtain the pure cockpit signal (no noise) from<br />
mixed cockpit signals, after CVR Signals’ classification, the<br />
comprehensive de-noise research for cockpit signals are<br />
made, including the DWT threshold de-noise, the cockpit<br />
sound’ ICA de-noising based on BSS and wavelet de-noise<br />
are put forward. Through different de-noise methods<br />
comparative research for cockpit signals, some valuable<br />
conclusion can be drawn in the end <strong>of</strong> the paper. These<br />
conclusions are very useful for judging and diagnosing the<br />
wreckage aircraft by pure cockpit signals (background<br />
sound signals).<br />
Index Terms—Information process; airplane; Blind Source<br />
Separation (BSS); Independent Components Analysis<br />
(ICA); Cockpit signals<br />
I. INTRODUCTION<br />
In order to record flight information and to<br />
reconstruct or diagnosis aircraft accident, most all <strong>of</strong><br />
large commercial aircraft (airplane) and other aircraft are<br />
necessary equipped with “black boxes, CVR & FDR”.<br />
Both CVR (Cockpit Voice Recorder) and FDR (Flight<br />
Date Recorder) play an indispensable role in aircraft<br />
accident investigation [1-3] . Compared with FDR, CVR<br />
(Fig.1) is one <strong>of</strong> the key evidence in the aircraft accident<br />
investigation. It is not only able to judge the unit's<br />
control, consciousness, determination, physical and<br />
mental state, but also can analyze the aircraft status and<br />
their environment by CVR. Cockpit signals (or CVR<br />
This paper is supported by the project <strong>of</strong> National Natural Science<br />
Foundation <strong>of</strong> China (Grant No. 61071203, 60772149).<br />
Copyright belongs to the papers all author and the units.<br />
Corresponding author: daolaicheng (daolaicheng@163.com)<br />
© 2011 ACADEMY PUBLISHER<br />
doi:10.4304/jcp.6.9.1920-1925<br />
sound information) includes crew's voices (or pilot's<br />
voices), environmental noise and different kinds <strong>of</strong><br />
backgrounds sound signals (such as switch sounds, an<br />
overspend warning signal). To effectively identify the<br />
background sound signals, de-noise must be done firstly<br />
for airplane cockpit signals. To solve de-noise problems,<br />
comprehensive de-noise researches are made<br />
systematically in the paper.<br />
Figure 1. Cockpit voice recorder hull (CVR)<br />
The paper structures are arranged as follows: Firstly,<br />
CVR Signals are classified into speech & non-speech<br />
signals; then, the discrete wavelet transform (DWT)<br />
threshold denoise for cockpit signals are described;<br />
thirdly, principles <strong>of</strong> blind source separation (BSS) and<br />
the principle <strong>of</strong> de-noising analysis based on ICA <strong>of</strong> BSS<br />
are made in details, including blind source separation<br />
and analysis <strong>of</strong> OGWE, maximum ratio <strong>of</strong> signal to noise<br />
<strong>of</strong> blind source separation algorithm; Fourthly; process <strong>of</strong><br />
cabin sound de-noising analysis based on ICA is done;<br />
finally , some research conclusions are obtained in the<br />
end <strong>of</strong> the paper.<br />
II. CVR SIGNALS’CLASSIFICAT [4-6]<br />
As we known that the CVR signals are recorded on<br />
4 channels connected by four wires: channel 1 from the<br />
cockpit area microphone <strong>of</strong> the CVR records non -speech<br />
information; channel 2 and channel 3 <strong>of</strong> the CVR record<br />
speech audio information from the captain and first<br />
<strong>of</strong>ficer’s audio selector panels; channel 4 records the
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1921<br />
audio information from the jump seat/observer’s radio<br />
panel.<br />
To analyze and de-noise CVR’ signals,<br />
conveniently, here the signals from different channels are<br />
divided into speech and no-speech signals. Speech<br />
information means the voice <strong>of</strong> captain and first <strong>of</strong>ficer or<br />
other crew; non-speech information mean noises and<br />
background sounds. On-speech signals can be divided to<br />
noises and background sound signals. For noises, they<br />
include engine noise, exterior air flow noise, sliding<br />
noise, selector noise, motor noise, loud frequency<br />
noises induced by tape itself or the recording circuit, etc.;<br />
The background sound signals include overspend<br />
warning, fire alarm, flight altitude advice, wing flutter,<br />
turbulence, landing gear extension and retraction or<br />
flaps/slaps up, switch sound, wheel landing, thumps,<br />
clicks, squeaks, rattles, airframe vibration or whirl<br />
flutter, etc. .<br />
III. DISCRECTE WAVELET TRANSFORM(DWT)<br />
THRESHOLDING DENOISE [5-9]<br />
Threshold is a technique used for signal and image<br />
denoising. The discrete wavelet transform uses two types<br />
<strong>of</strong> filters: (1) averaging filters, and (2) detail filters.<br />
When we decompose a signal using the wavelet<br />
transform, we are left with a set <strong>of</strong> wavelet coefficients<br />
that correlates to the high frequency subbands. These<br />
high frequency subbands consist <strong>of</strong> the details in the data<br />
set. If these details are small enough, they might be<br />
omitted without substantially affecting the main features<br />
<strong>of</strong> the data set. Additionally, these small details are <strong>of</strong>ten<br />
those associated with noise; therefore, by setting these<br />
coefficients to zero, we are essentially killing the noise.<br />
This becomes the basic concept behind threshold-set all<br />
frequency subband coefficients that are less than a<br />
particular threshold to zero and use these coefficients in<br />
an inverse wavelet transformation to reconstruct the data<br />
set.<br />
There are the double-density DWT and doubledensity<br />
complex DWT for 1-D signals. Here, the doubledensity<br />
DWT method is only discussed in the following<br />
papers simply.<br />
The method can be implemented by program. This<br />
program takes as input two parameters, one <strong>of</strong> which is<br />
the noisy input signal (to be threshold) and the other <strong>of</strong><br />
which is the threshold point. A sample CVR noisy signal<br />
is shown below (Fig.2), whose length is 512. To denoise<br />
the signal, we first take the forward double-density DWT<br />
over four scales. Then a denoising method, knows as s<strong>of</strong>t<br />
threshold, is applied to the wavelet coefficients though<br />
all scales and sub bands. The s<strong>of</strong>t threshold method sets<br />
coefficients with values less than the threshold T to 0,<br />
and then subtracts T from the non-zero coefficients. The<br />
double-density DWT method results in the following the<br />
CVR denoised signal (Fig.3).<br />
© 2011 ACADEMY PUBLISHER<br />
Figure 2. a sample CVR noisy signal<br />
Figure 3. the CVR signal by DWT threshold denoise<br />
IV. PRINCIPLES OF BLIND SOURCE<br />
SEPARATION((BSS) [10-13]<br />
Blind Source Separation (Blind Source Separation,<br />
BSS) is a process that the source signal is extracted from<br />
the mixed-signal. Normally, cockpit signals (CVR sound<br />
signals) are mixed signals, so, the pure cockpit signal<br />
(no noise) can be acquired by BSS. Some researches are<br />
made in the following paragraph.<br />
A Blind source separation<br />
Blind signal need to be in the certain conditions,<br />
and the "Blind" has double meanings: the source signal is<br />
unknown; how the source signals mixed is also unknown.<br />
(1) Definition <strong>of</strong> blind source separation<br />
Blind Source Separation (Blind Source Separation,<br />
BSS) is a process that the source signal is extracted from<br />
the mixed-signal. Blind signal need to be in the certain<br />
conditions, and the "Blind" has double meanings: the<br />
source signal is unknown; how the source signals mixed<br />
is also unknown.<br />
(2)The mathematical model <strong>of</strong> blind source<br />
separation<br />
The observation signal is M signals which are<br />
mixed, that is from the N statistical mixture <strong>of</strong> unknown<br />
source. Purpose <strong>of</strong> the BBS study is to separate the source<br />
signal from the signal mixture, which mathematical<br />
model described as follows:<br />
There are N independent sound sources, sound<br />
source signal is s i (t) ( i=1, 2, …… , N ), in order to
1922 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
separate the source signals, M(M ≥ N) measurement<br />
points are measured, the measured signal is y j (t)<br />
( j=1 , 2 , ……, M ), the mathematical expression as<br />
follows(1):<br />
N<br />
∑ aij i<br />
i=1<br />
y j () t = s () t + Q<br />
Here: yj() t – number j masured signal;<br />
si() t - number i sound source;<br />
ij<br />
(j=1, 2, ……, M) (1)<br />
a - The corresponding coefficient;<br />
Q - Noise vector.<br />
This is the simplest model <strong>of</strong> blind source separation<br />
and is the separation <strong>of</strong> instantaneous linear mixing<br />
model. If the blind signal is collected in a silent room,<br />
the noise vector is negligible, That is to say: Q = 0,<br />
Equation 1 can be written in matrix form (2):<br />
Y=AS (2)<br />
Here: A - coefficient matrix constituted by a ij , unknown;<br />
S - Source signal, unknown;<br />
Y - Observed signal, known.<br />
(3) The uncertainty <strong>of</strong> source signal<br />
Blind source separation theory based on the<br />
observed signals and source signals which are<br />
independent only. Because the weak conditions for blind<br />
source separation, amplitude and the location <strong>of</strong> source<br />
signal separated are uncertain. Then, equation (1) is<br />
written as follows (3):<br />
N aij<br />
y j () t = d i si<br />
() t (j=1, 2, ……, M) (3)<br />
d<br />
∑<br />
i=<br />
1<br />
j<br />
We can see that di is a constant, disi(t) as the source<br />
signal, the equation still holds, which reflects the<br />
uncertainty <strong>of</strong> the source signal amplitude, and secondly,<br />
when exchange location <strong>of</strong> coefficient and corresponding<br />
signal at the same time, the equation still holds. If the<br />
signal information is mainly contained in the signal<br />
waveform, the uncertainty does not affect the separation<br />
<strong>of</strong> the signals.<br />
(4) Simplify the problem <strong>of</strong> blind source separation<br />
BSS is to identify A and S in unknown<br />
circumstances <strong>of</strong> A and S, according to the independence<br />
<strong>of</strong> Y and the source signal, the separation process is to<br />
find a separation matrix W(4):<br />
Y=WX=WAS=CS, C=WA (4)<br />
For simplicity, the time t did not be written (the<br />
same below). If you can indeed find such a matrix,<br />
makes the C is the unit diagonal matrix (M = N), there<br />
are yi=is(i=1 , 2 , 3 , …, M) , then solution <strong>of</strong> the blind<br />
source separation problem is transformed to find the<br />
matrix W.<br />
B Independent components analysis (ICA)<br />
(1) Definition <strong>of</strong> independent components analysis<br />
(ICA)<br />
© 2011 ACADEMY PUBLISHER<br />
The current BSS description <strong>of</strong> the problem are<br />
mostly based on ICA model, BSS and ICA is equivalent<br />
normally, the difference is that ICA is a mathematical<br />
model that can solve various different problems, but BSS<br />
model is a real problem model that can be applied to<br />
other solutions, not only in ICA method.<br />
(2)The understanding <strong>of</strong> the independent<br />
components analysis (ICA)<br />
What is the independent components analysis (ICA)<br />
Look at a simple example: three people xi(t) are talking<br />
in a room at the same time , three microphones are placed<br />
at different locations, three speech signals si(t) are get<br />
through the three microphones, expression as follows(5):<br />
3<br />
() ()<br />
x t = ∑ a s t<br />
(5)<br />
i ij i<br />
j=<br />
1<br />
Here: a ij (i, j = 1, 2, 3 )is the mixing coefficient.<br />
According to the distance between the room<br />
microphone and the people (ignore the delay and other<br />
additional factors such as the sound wave diffraction and<br />
refraction, etc.), the problem <strong>of</strong> a similar cocktail party<br />
is solved, only under the condition that three <strong>of</strong> their<br />
words can get on the basis <strong>of</strong> xi (t), . If more than three<br />
people and different locations, equation (5) can be<br />
expressed as the generalized mathematical model(6):<br />
xi ( t)<br />
= ai1<br />
s1(<br />
t)<br />
+ ai<br />
2s<br />
2 ( t)<br />
+ … + ains<br />
n ( t)<br />
(i, j = 1, 2, 3……, n) (6)<br />
Here, the independent component si () t and the<br />
mixed matrix a ij are unknown, the observation signal<br />
xi ( t)<br />
is known. ( t)<br />
xi is to be used to estimated ( t)<br />
and a ij in practical applications. That is to say, in order<br />
to make Y = S, the separation matrix W must be sought.<br />
In the blind source separation algorithm, such as the<br />
kurtosis maximization algorithm (based on higher-order<br />
cumulates), the minimum mutual information method,<br />
maximum likelihood estimation method, the joint<br />
digitalization method, the maximum SNR method, are<br />
based on ICA.<br />
V. PRINCIPLE OF DE-NOISING ANALYSES BASED<br />
ON ICA<br />
According to the theoretical research, ICA process<br />
is to separate the independent component mostly<br />
approaching each source signal, that is, set the objective<br />
function to achieve the approximation. To achieve this<br />
approximation, we establish an objective function J (W)<br />
with separation matrix W according to information theory,<br />
statistical theory and other methods, which is based on<br />
the definition <strong>of</strong> the objective function in this paper. W is<br />
found to make J (W) into a maximum (or minimum)<br />
value.<br />
In order to find both the best objective function J (W)<br />
and an effective algorithm for solution, sound de-noising<br />
analysis module is made based on ICA process in this<br />
paper. And the speech de-noising analysis is made in the<br />
pilot cabin based on ICA, including two aspects: one is<br />
OGWE blind source separation algorithm with the usage<br />
s i
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1923<br />
<strong>of</strong> higher-order cumulates; the other is based on the<br />
largest ratio in signal to noise.<br />
A Blind source separation and analysis <strong>of</strong> OGWE<br />
Mixed signal need to be reduced into the mean and<br />
the whitened treatment in OGWE independent<br />
component analysis algorithm; because the mean is to<br />
simplify the calculation, while the whitening treatment<br />
can simplify the matrix, decreasing the problem<br />
complexity. Minimal contrast function (objective<br />
function) is used in OGWE algorithm; the rotation vector<br />
and higher-order cumulates matrix is calculated; and<br />
given transformations and angle calculations are<br />
combined for the realization <strong>of</strong> mixed-signal separation.<br />
B Maximum ratio <strong>of</strong> signal to noise <strong>of</strong> blind source<br />
separation algorithm<br />
Signal to noise ratio function (7):<br />
T<br />
s•s SNR = 10log T<br />
e•e (7)<br />
T<br />
s•s = 10log T<br />
( s− y) •( s− y)<br />
Here: S is the source signal, y is the estimated<br />
signal, noise signal is expressed as e = s − y 。 As the<br />
source signal is unknown, y (n) concludes noise, it is<br />
estimated that moving-average y ~ <strong>of</strong> y (n) is used to<br />
replace the source signal s, y = Wx ; ~ y = W~<br />
x ; W is<br />
the separation matrix; x ~ is processed by the moving<br />
p<br />
average, ~ 1<br />
xi<br />
( n)<br />
= ∑ xi<br />
( n − j)<br />
, i=0, 1, 2, …, p-1<br />
p j=<br />
0<br />
Therefore, the objective functions <strong>of</strong> maximized<br />
signal to noise ratio can expressed as follows (8):<br />
T T<br />
Wxx W<br />
SNR = (8)<br />
T T<br />
W ( ~ x − x)(<br />
~ x − x)<br />
W<br />
The purpose <strong>of</strong> the algorithm is to find theW ~ .<br />
VI. PROCESS OF CABIN SOUND DE-NOISING<br />
ANALYSES BASED ON ICA [17-18]<br />
To take advantage <strong>of</strong> the theory <strong>of</strong> ICA, at least the<br />
same sampling signal points must be re-structured, one is<br />
the translated source signal, and the other is the pure<br />
noise signal in cabin. These two signals could be denoised<br />
by above-mentioned method.<br />
A Sound de-noising analysis by OGWE<br />
Here, a pilot's voice signal (the cockpit signals<br />
recorded by CVR inside the cabin) and a pure noise<br />
signal in cabin are given. They are separately shown in<br />
Fig.4 and Fig.5.The two kinds <strong>of</strong> signals are firstly<br />
analyzed through using OGWE based on higher-order<br />
cumulates ICA blind source separation. Fig.6 and Fig.7<br />
are separated signals respectively.<br />
© 2011 ACADEMY PUBLISHER<br />
Figure 4. Waveform <strong>of</strong> a typical cabin sound (source signal)<br />
Figure.5 Noise waveforms in cabin<br />
Figure. 6 Waveform <strong>of</strong> pilot voice signal<br />
Figure. 7 Noise Waveform<br />
From the research and comparisons, we found that<br />
the uncertainty <strong>of</strong> blind source separation is order and<br />
amplitude. Fig.5 and Fig.7 show that the clear distinction<br />
<strong>of</strong> the amplitude <strong>of</strong> the noise signals. In fact, obvious<br />
difference also is included in amplitude <strong>of</strong> the pilot voice<br />
signal.<br />
B Sound de-noising module by the maximum signal to<br />
noise ratio<br />
Maximum SNR can be used to analyze the blind<br />
source separation. Matrix W is got as follows: W= [-
1924 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
.048306233, -0.965696385-0.914864537, 0.050990232].<br />
Absolute value <strong>of</strong> the separation matrix W in the second<br />
diagonal approximates to 1, which being effectively<br />
separate the pilot voice signal.<br />
C Comparative analysis <strong>of</strong> de-noising module sound<br />
based on wavelet<br />
To illustrate the accuracy <strong>of</strong> de-noising by the ICAbased<br />
blind source separation method, the effects <strong>of</strong> denoising<br />
was compared based on wavelet in MATLAB.<br />
Three different methods are made for different<br />
signals, which are separately based on OGWE after denoising;<br />
based on maximum signal to noise ratio after denoising;<br />
based on Mat lab wavelet toolbox after denoising.<br />
Their wave charts are shown from Fig.8~Fig.10.<br />
FFigure.8 Pilot Speech Signal Waveform<br />
Based on OGWE after de-noising<br />
Figure.9 The Pilot Speech Signal Waveform<br />
Based on SNR after de-noising<br />
Figure. 10 The pilots voice signal aveform<br />
Based on SNR after de-noising<br />
From these charts, we can conclude that blind<br />
source separation method based on ICA has the<br />
approximate effects with the successful wavelet for denoising<br />
except small difference in amplitude; and, that<br />
blind source separation method based on ICA has the<br />
approximate effects with the successful wavelet for denoising<br />
except small difference in amplitude.<br />
© 2011 ACADEMY PUBLISHER<br />
D Comparative analyses <strong>of</strong> denoising cabin sound on<br />
Short-term zero-crossing rate and short-term<br />
energy<br />
Short-term zero-crossing rate and short-term energy<br />
are further used to accurate the analysis <strong>of</strong> the three<br />
methods <strong>of</strong> de-noising effect.<br />
In order to analyze CVR source signal by typical<br />
short-time zero-crossing module, the 2984 points is<br />
sampled in source signal. A unit has 600 points in shortterm<br />
analysis; two units with 300 points overlapping<br />
regions, and the signal are divided into 9 units. For the<br />
short-term zero-crossing number; OGWE; the maximum<br />
SNR method and Wavelet De-noising, their comparative<br />
analysis results can be seen in Table.1, Table 2.<br />
From Table 1, OGWE blind source separation<br />
method is much closer to wavelet in the Zero-crossing<br />
numbers and de-noising effect.<br />
From Table 2, OGWE blind source separation<br />
method is much closer to wavelet in the short-term<br />
energy and de-noising effect. De-noising effect based on<br />
maximum signal to noise ratio is somehow less.<br />
TABLE Ⅰ RESULTS OF THE DIFFRERENT<br />
ANALYSES MODULE<br />
ShortZero OGWE Maximum SNR Method Wavelet<br />
1 26 18 26<br />
2 29 28 29<br />
3 29 34 29<br />
4 28 23 26<br />
5 33 35 34<br />
6 34 39 33<br />
7 29 25 27<br />
8 31 29 29<br />
9 30 27 27<br />
TABLE II SHOR ENERGY, OGWE, MAXIMUM SNR METHOD,<br />
WAVELET DENOISE RESULTS OF SOUND<br />
Short<br />
Energy<br />
OGWE Maximum SNR Method Wavelet<br />
1 19.027 34.324 14.221<br />
2 39.04 17.329 29.87<br />
3 34.362 17.852 22.56<br />
4 34.156 20.756 24.455<br />
5 41.723 27.137 33.812<br />
6 37.916 23.129 33.026<br />
7 31.498 19.81 24.235<br />
8 14.418 23.966 12.089<br />
9 5.2407 13.585 4.651<br />
VII. SPECTRAL ANALYSES OF CANIN SOUND<br />
SIGNAL BY ADOBE AUDITION SOFTWARE<br />
Pilots’ voice signal and pure noise signal are both<br />
directly separated by ICA-based blind source separation<br />
method and spectra analysis by Adobe Audition s<strong>of</strong>tware.<br />
Their results can be got in Fig.11. The three different<br />
spectrum (the source signal spectrum, pilot voice signal<br />
spectrum and the noise signal spectrum can be seen in<br />
figure11 from left to right.<br />
-<br />
1<br />
0
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1925<br />
From the spectra <strong>of</strong> three signals in Fig.11, we can<br />
conclude that 1) higher energy intensity are occurs for<br />
separating the pilot speech signal than the source noisy<br />
signal in voice by ICA-based blind source separation<br />
methods;2)the noise is at a relatively high frequency band.<br />
Figure.11 Cabin sound spectrums<br />
VIII. CONCLUSIONGS<br />
Through above typical cockpit signals de-noise<br />
analysis and research; some conclusions can be drawn as<br />
followings:<br />
(1)The cockpit signals are complex, can be<br />
classified into speech information and non -speech<br />
information; There are noise signal and background<br />
sound signals;<br />
(2)Discrete wavelet transform (DWT) shareholding<br />
denoise play an vital role for cockpit signals;<br />
(3)The pilots’ source signals <strong>of</strong> aircraft cockpit voice<br />
are separated based on the principle <strong>of</strong> ICA blind source<br />
separation methods.<br />
(4)Three different blind source separation methods<br />
(based on OGWE, maximum signal to noise ratio after<br />
de-noising and Mat lab wavelet toolbox after de-noising)<br />
have the approximate effects except small difference in<br />
amplitude.<br />
(5)Through the spectra <strong>of</strong> three signals (the source<br />
signal spectrum, pilot voice signal spectrum and the<br />
noise signal spectrum), we can conclude that 1) higher<br />
energy intensity are occurs for separating the pilot speech<br />
signal than the source noisy signal in voice by ICA-based<br />
blind source separation methods;2)the noise is at a<br />
relatively high frequency band.<br />
REFERNCES<br />
[1] Shu Ping, Zhong Minzhu, Yang Lin, Amelioration <strong>of</strong><br />
Cockpit Voice Recorder Decoding System [M].Beijing:<br />
Aviation Industry Press, 2004.97-100.<br />
[2] Kendall W.Neville (USA).“Research on Flight Techniques<br />
and Aviation Safety (ISBN 7-5364-6055-4), ” Published by<br />
Si Chuan Science and Technology Press, Chengdu,<br />
China, pp.40-46, August 2006.<br />
[3] Sound Spectrum Study Cockpit Voice Recorder – 12.<br />
Statistical Summary <strong>of</strong> Commercial Jet Airplane Accidents<br />
Worldwide Operations 1959 – 2008.Boeing, 2009<br />
Statistical Summary, JULY 2009.<br />
[4] McKinney Martin F, Jeroen Breebear.Feature for Audio<br />
and Music Classification [EB/OL]. http://mckinney.<br />
© 2011 ACADEMY PUBLISHER<br />
philps.com, 2008<br />
[5] Dao Laicheng, Chui Jieyi, Hongyu Yao.Nonstationary<br />
quiver spindle background sound analysis <strong>of</strong> airplane vie<br />
Wigner-Ville and Wavelet time-scale distribution[J] .<br />
Published by China Machine Press, Beijing, China,<br />
vol.43, no.5 pp.150-154, May 2007.<br />
[6] Hairong Guo, Daolai Cheng, Zhoufeng Liang.Sound<br />
spectral analyses <strong>of</strong> black boxes based on Matlab and VC<br />
program [J].Micro computer information, Beijing, China,<br />
2008, 24(3):299 -301\<br />
[7] Zhibing Gao, Chuijie Yi, Yangming Zhou The information<br />
management system’ design and implement for the<br />
characteristics <strong>of</strong> cockpit background sound [J].Micro<br />
computer information, Beijing, China, 2009, 25(34):18 -<br />
19.<br />
[8] Daolai Cheng, Chuijie Yi, Hongyu Yao, .Sound Signal<br />
Analysis <strong>of</strong> CVR Based on CVDS, WT & CZT<br />
Algorithm[J].DCDIS Series B, Vol.14(S5), 2007: 129-<br />
134.<br />
[9] CHENG Daolai, YI Chuijie, Zhang Zhiqiang.<br />
Comparative Analyses and Experiment Verification on<br />
Cockpit Background Sound’ Characteristic Frequency [J].<br />
Fourth International Conference on Innovative Computing,<br />
Information and Control December 7-9, 2009, Kaohsiung,<br />
Taiwan.<br />
[10] DaoLai Cheng, QingCheng Wang ChuiJie Yi, HongYu<br />
Yao.Analysis and Research for Airplane Cockpit<br />
Sound’ICA Denoise Based on Blind Source Separation<br />
Principle.2011 International conference on Industry,<br />
information System and Material Engineering<br />
(IISME2011)(C).April 16-17, 2011, Guangzhou, China<br />
[11] Malgorzata Zygarlicka, Janusz Mroczka. Reduction <strong>of</strong> the<br />
cross-terms <strong>of</strong> the Wigner Ville distribution by image<br />
processing. X international PhD workshop OWD’2008,<br />
215-220.<br />
[12] ZhiQiang Zhang Characteristics acquirement and analyses<br />
aircraft black boxes cockpit voice [P]. a dissertation for the<br />
master degree in engineering.Qingdao, Shangdong<br />
province, China. Qingdao Technological University,<br />
2010.12<br />
[13] Zhibing Gao, Chuijie Yi, Yangming Zhou The information<br />
management system’ design and implement for the<br />
characteristics <strong>of</strong> cockpit background sound [J].Micro<br />
computer information, Beijing, China, 2009, 25(34):18 -<br />
19.<br />
Daolai Cheng, born in 1965, PhD,<br />
Pr<strong>of</strong>essor, the vice director <strong>of</strong> College <strong>of</strong><br />
Urban Construction & Safety<br />
Engineering, Shanghai Institute <strong>of</strong><br />
Technology, Shanghai (201418), China.<br />
His research interests include thermal<br />
energy and power engineering, signal<br />
process.<br />
Tel:+86-21-60873631, 13311998959<br />
E-mail:daolaicheng @163 . com<br />
Chuijie Yi, born in 1958, PhD, pr<strong>of</strong>essor, president <strong>of</strong> Qingdao<br />
Technological University, China. His research interests include<br />
noise and vibration control.<br />
Hongyu Yao, born in 1963, PhD, researcher, chief engineer <strong>of</strong><br />
General Civil Aviation Administration <strong>of</strong> China (CAAC),<br />
China. His research interests include cockpit signal analyses.
1926 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
Fuzzy Support Vector Machines Control<br />
for 6-DOF Parallel Robot<br />
Dequan Zhu<br />
Institute <strong>of</strong> Intelligent Machines, Chinese <strong>Academy</strong> <strong>of</strong> Sciences, Hefei, China<br />
Department <strong>of</strong> Automation, University <strong>of</strong> Science and Technology <strong>of</strong> China, Hefei, China<br />
College <strong>of</strong> Engineering, Anhui Agricultural University, Hefei, China,<br />
Email: adqzhu@sina.com<br />
Tao Mei<br />
Institute <strong>of</strong> Intelligent Machines, Chinese <strong>Academy</strong> <strong>of</strong> Sciences, Hefei, China<br />
Email: tmei@iim.ac.cn<br />
Lei Sun<br />
College <strong>of</strong> Engineering, Anhui Agricultural University, Hefei, China<br />
Email: SL961102@163.com<br />
Abstract—In order to realize the trajectory tracking control<br />
<strong>of</strong> six degrees <strong>of</strong> freedom parallel robot, the dynamics<br />
equation <strong>of</strong> six degrees <strong>of</strong> freedom parallel robot was<br />
established. The parallel robot has obvious nonlinear,<br />
uncertainty characteristics and external disturbance, so the<br />
sliding mode variable structure theory was introduced into<br />
the system control. A fuzzy support vector machines control<br />
strategy based on sliding mode control was designed to<br />
reduce the oscillation <strong>of</strong> the sliding mode control.<br />
Parameters <strong>of</strong> fuzzy support vector machines controller<br />
were optimized by hybrid learning algorithm, which<br />
combines least square algorithm with improved genetic<br />
algorithm, to get the optimal control performance for the<br />
controlled object. The controller designed consists <strong>of</strong> a fuzzy<br />
sliding mode controller and a fuzzy support vector machines<br />
controller, and the compensation controller is selected by<br />
comparing switching function with the thickness <strong>of</strong><br />
boundary layer. Simulation results show that under the<br />
condition <strong>of</strong> model error and external disturbance, the<br />
control strategy designed gets tracking effect with high<br />
precision and speed.<br />
Index Terms—parallel robot, fuzzy control, support vector<br />
machines, sliding mode control, dynamics equation<br />
I. INTRODUCTION<br />
The six degrees <strong>of</strong> freedom (6-DOF) Stewart platform<br />
parallel robot is a closed-loop mechanism in which the<br />
end-effector (mobile platform) is connected to the base<br />
by six extensible legs [1]. Compared with serial robot,<br />
the parallel robot has potential advantages in terms <strong>of</strong><br />
compliance, accuracy, high speed and payload. Therefore,<br />
it has been used in precision lathe, assembly robotic<br />
manipulator and electronics manufacture [2,3]. Because<br />
<strong>of</strong> the complex condition and the uncertain object, the<br />
parallel robot is not only a complicated nonlinear<br />
multivariable and strong coupling system, but also a time-<br />
© 2011 ACADEMY PUBLISHER<br />
doi:10.4304/jcp.6.9.1926-1934<br />
varying system. The parallel robot is not controlled<br />
accurately and its track is not kept better by the general<br />
model-based control method [4,5].<br />
In the domain <strong>of</strong> artificial intelligence techniques for<br />
the system control, different control algorithms have been<br />
used to realize the trajectory tracking control for the<br />
parallel robot. The sliding mode control algorithm has<br />
complete adaptability for system disturbance and stirring,<br />
which is extensively applied in the control <strong>of</strong> the parallel<br />
robot [6,7]. In the fuzzy control, the mathematical model<br />
for the system does not be set up precisely and the joints<br />
<strong>of</strong> robots can be decoupled, but fuzzy control system is<br />
easily influenced by nonlinear, time-varying and random<br />
disturbance [8]. Neural network control algorithm has<br />
many advantages, such as self-learning, self-organizing,<br />
self-adaptive capacity, nonlinear and parallel distributed<br />
processing, and so on. However, it also has the congenital<br />
defects, such as it falls into local minimum easily, and it<br />
is weakly normalized for few samples [9-11]. These<br />
defects make it difficult to meet control precision for<br />
parallel robot.<br />
Thus, some control methods are combined to realize<br />
the trajectory tracking control for the parallel robot. A<br />
cascade-control algorithm based on a sliding mode in the<br />
legspace was proposed by Hongbo Guo, Yongguan Luo,<br />
Guirong Liu and Hongren Li to realize the trajectory<br />
tracking control <strong>of</strong> hydraulically driven six degrees <strong>of</strong><br />
freedom parallel robotic manipulator [6]. A control<br />
approach which is based on the coupling <strong>of</strong> sliding mode<br />
and multi-layers perceptron neural networks was<br />
proposed by Achili B, Daachi B, Amirat Y and Ali-cherif<br />
A to deal with the robust adaptive control tracking <strong>of</strong> a 6<br />
degree <strong>of</strong> freedom parallel robot [9]. A sliding mode<br />
control with discontinuous projection-based adaptation<br />
laws was proposed by Yangjun Pi and XuanyinWang to<br />
improve the tracking performance <strong>of</strong> the parallel robot
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1927<br />
manipulator [10]. A robust neural-fuzzy-network control<br />
system was presented by Rongjong Wai and Pochen<br />
Chen to realize the joint position control <strong>of</strong> an n-rod<br />
robotic manipulator for periodic motion in order to deal<br />
with the uncertainties in application, such as friction<br />
forces, external disturbances, and parameter variations<br />
[12]. A new discrete sliding mode control approach for<br />
parallel robot was presented by Shaocheng Qu and<br />
Yongji Wang to achieve accurate servo tracking in the<br />
presence <strong>of</strong> load variations, parameter variations and<br />
nonlinear dynamic interactions [13]. A fuzzy-PI<br />
compound control system for three-cylinder hydraulic<br />
parallel robot was designed by Qidan Zhu, Xunyu Zhong<br />
and Bo Xu [14].<br />
Support vector machines, which is based on the<br />
structural risk minimization rule, overcomes the<br />
shortcoming that neural network structure relies on the<br />
experience <strong>of</strong> designer. Its topology structure is decided<br />
by support vectors. It solves these problems well, such as<br />
high dimension, local minimum and small samples, and<br />
has advantages <strong>of</strong> both neural network and traditional<br />
model [15,16]. So the support vector machines is<br />
combined with fuzzy control to design a fuzzy support<br />
vector machines controller for parallel robot to reduce the<br />
chattering in sliding mode control. It is important to<br />
select the proper SVM parameters for improving the<br />
learning and generalizing capacity <strong>of</strong> the control system<br />
[16]. Thus, the fuzzy proportional coefficients were<br />
adjusted with the controlled object, and the parameters <strong>of</strong><br />
the controller were optimized by least square (LS)<br />
learning algorithm and improved genetic algorithm (IGA)<br />
in order to improve control precision and working<br />
stability <strong>of</strong> the parallel robot.<br />
II. DYNAMICS MODEL OF 6-DOF STEWART PLATFORM<br />
PARALLEL ROBOT<br />
l6 l l l<br />
5<br />
4 3<br />
l1<br />
l2<br />
Figure 1. Schematic diagram <strong>of</strong> 6-DOF Stewart platform parallel robot<br />
A. Reference frame<br />
The 6-DOF Stewart platform parallel robot is shown in<br />
Fig.1. It consists <strong>of</strong> mobile platform, base platform and<br />
six extensible leg, each <strong>of</strong> which is connected with the<br />
two platforms by spherical joints [2,3]. The legs are<br />
driven by six servo-electromotors. To describe the motion<br />
<strong>of</strong> the mobile platform, two reference frames are chosen:<br />
a fixed reference frame {B, XB, YB, ZB} attached to the<br />
base platform and a mobile reference frame {P, XP, YP, ZP}<br />
attached to the mobile platform, as shown in Fig.1. Six<br />
coordinates are used to further describe the position and<br />
© 2011 ACADEMY PUBLISHER<br />
the orientation <strong>of</strong> the mobile platform in detail. Three<br />
coordinates are the positional displacements [Xp, Yp, Zp] T ,<br />
which describe the position <strong>of</strong> a fixed point in the mobile<br />
platform with respect to the fixed reference frame. The<br />
other three coordinates are the angular displacements,<br />
represented by Euler angles [γ, β, α] T , which describe the<br />
orientation <strong>of</strong> the mobile platform with respect to the<br />
fixed reference frame. Therefore, the generalized<br />
coordinate vector, whose elements are the six variables<br />
chosen to describe the position and orientation <strong>of</strong> the<br />
mobile platform, can be defined as [XP, YP, ZP, γ, β, α] T .<br />
The rotation matrix from mobile reference frame to fixed<br />
reference frame can be described as follow:<br />
⎡cαcβ<br />
cαsβsγ<br />
− sαcγ<br />
sαsβcγ<br />
⎤<br />
R =<br />
⎢<br />
⎥<br />
⎢<br />
sαcβ<br />
cαcγ<br />
+ sαsβsγ<br />
sαsβsγ<br />
− cαsγ<br />
⎥<br />
⎢⎣<br />
− sβ<br />
cβsγ<br />
cβcγ<br />
⎥⎦<br />
B<br />
(1)<br />
P<br />
where c(⋅) denotes cos(⋅); s(⋅) denotes sin(⋅).<br />
The ith leg vector li & with respect to the fixed reference<br />
frame can be described as<br />
B<br />
l&<br />
i = c&<br />
+ RP<br />
B&<br />
i − P&<br />
i ,i=1,2,,6 (2)<br />
where c is the translation vector <strong>of</strong> the origin <strong>of</strong> the<br />
mobile reference frame with respect to the fixed reference<br />
frame; Bi is the position vector <strong>of</strong> the ith joint point <strong>of</strong><br />
base platform with respect to the fixed reference frame;<br />
Pi is the position vector <strong>of</strong> the ith joint point <strong>of</strong> mobile<br />
platform with respect to the mobile reference frame.<br />
The extended length <strong>of</strong> the ith leg is described as<br />
Δ l = l&<br />
− l<br />
(3)<br />
i<br />
i<br />
where li0 is the original length <strong>of</strong> the ith leg.<br />
Well-controlled lengths <strong>of</strong> six legs make the mobile<br />
platform follow the desired trajectory.<br />
B. Dynamics equation<br />
In order to solve the kinetic energy and the potential<br />
energy <strong>of</strong> parallel robot, the whole system is separated<br />
into the mobile platform and the six legs with the base<br />
platform.<br />
Suppose the angle velocity <strong>of</strong> mobile platform is ωk ,<br />
and then the kinetic energy <strong>of</strong> mobile platform KEh,<br />
which includes translational and rotating kinetic energy,<br />
can be described as<br />
1 2 2 2<br />
KEh = ( mu<br />
( x&<br />
P + y&<br />
P + z&<br />
P ) + ωk<br />
Ichωh<br />
) (4)<br />
2<br />
where mu is the mass <strong>of</strong> mobile platform; xP, yP, zP is the<br />
displacement about the axis XP, YP, ZP respectively; o I ch is<br />
the inertia matrix <strong>of</strong> the mobile platform with respect to<br />
mobile reference frame. It can be computed by<br />
o h T<br />
I ch = RIc<br />
R<br />
(5)<br />
where R is the corresponding rotating matrix, defined by<br />
the angle rotating rule <strong>of</strong> Roll-Pitch-Yaw; h<br />
I c is the<br />
rotating inertia with respect to the mobile reference frame.<br />
It is described as<br />
⎡I<br />
⎤<br />
X 0 0<br />
P<br />
h ⎢<br />
⎥<br />
Ic<br />
= ⎢ 0 IY<br />
0<br />
P ⎥<br />
(6)<br />
⎢<br />
⎥<br />
⎣<br />
0 0 I ZP<br />
⎦<br />
i 0<br />
T<br />
o
1928 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
where I , I , I is the rotating inertia with respect to<br />
X P<br />
YP<br />
Z P<br />
the axis XP, YP, ZP respectively.<br />
The angle velocity <strong>of</strong> mobile platform is described as<br />
v r r<br />
ω = & γR<br />
( α ) R ( β ) X + & βR<br />
( α)<br />
Y + & αZ<br />
k<br />
Z B<br />
YB<br />
⎡CαCβ<br />
− Sα<br />
0⎤<br />
⎡ & γ ⎤<br />
=<br />
⎢<br />
⎥ ⎢ ⎥<br />
⎢<br />
SαCβ<br />
Cα<br />
0 &<br />
⎥ ⎢<br />
β<br />
(7)<br />
⎥<br />
⎢⎣<br />
− Sβ<br />
0 1⎥⎦<br />
⎢⎣<br />
& α ⎥⎦<br />
where R , Z R is the corresponding rotating matrix<br />
B YB<br />
respectively, defined by the angle rotating rule <strong>of</strong> Roll-<br />
Pitch-Yaw; c(⋅) denotes cos(⋅); s(⋅) denotes sin(⋅).<br />
The kinetic energy <strong>of</strong> mobile platform is rewritten as<br />
1 T<br />
KEh = q&<br />
M h ( q)<br />
q&<br />
(8)<br />
2<br />
where Mh(q) is defined as<br />
⎡mu<br />
0 0 0 0 0 ⎤<br />
⎢<br />
⎥<br />
⎢<br />
0 mu<br />
0 0 0 0<br />
⎥<br />
⎢ 0 0 mu<br />
0 0 0 ⎥<br />
M h ( q)<br />
= ⎢<br />
⎥<br />
⎢ 0 0 0 M h44<br />
M h45<br />
− I xsβ<br />
⎥<br />
⎢ 0 0 0 M<br />
⎥<br />
h54<br />
M h55<br />
0<br />
⎢<br />
⎥<br />
⎢⎣<br />
0 0 0 − I xsβ<br />
0 I X ⎥⎦<br />
(9)<br />
2<br />
2 2<br />
2 2<br />
where M h44<br />
= I X sin β + IY<br />
sin γ cos β + I Z cos γ cos β ,<br />
M h45<br />
= ( I x − IZ<br />
) cosγ<br />
sinγ<br />
cos β , M h454<br />
= ( IY<br />
− I Z ) cosγ<br />
sin γ cos β ,<br />
2<br />
2<br />
M h55<br />
= IY<br />
cos γ + I Z sin γ .<br />
The potential energy <strong>of</strong> mobile platform is described as<br />
T<br />
P = m gz = [ 0 0 m g 0 0 0]<br />
q (10)<br />
h<br />
u<br />
p<br />
u<br />
where g is the gravity acceleration.<br />
The extensible legs, which are driven by the servoelectromotors,<br />
are separated into the cylinders and the<br />
rods. They are regarded as the rigid parts with rotating<br />
inertia with respect to themselves. Each leg is represented<br />
by the centroid point <strong>of</strong> it.<br />
The position <strong>of</strong> the ith leg centroid point Gi is<br />
described as<br />
1<br />
r m i<br />
BiGi<br />
m ilBi<br />
m i Li<br />
l i ui<br />
lˆ<br />
2 r<br />
= [ 1 + 2 ( − 2 )] = [ i + Li<br />
] ui<br />
m1<br />
i + m2i<br />
m1i<br />
+ m2i<br />
(11)<br />
where l im<br />
i l im<br />
i<br />
lˆ<br />
1 1 − 2 2<br />
i =<br />
, l1i is the distance between the<br />
m1i<br />
+ m2i<br />
center point <strong>of</strong> the ith base joint and the centroid point <strong>of</strong><br />
cylinder <strong>of</strong> the ith leg; l2i is the distance between the ith<br />
upper joint and the rod <strong>of</strong> the ith leg; m1i is the mass <strong>of</strong><br />
cylinder <strong>of</strong> the ith leg, m2i is the mass <strong>of</strong> rod <strong>of</strong> the ith leg;<br />
Li is the length <strong>of</strong> the ith leg; ui is the orientation <strong>of</strong> the ith<br />
leg. ui can be defined as<br />
r BiPi<br />
u =<br />
(12)<br />
i<br />
L<br />
i<br />
The velocity <strong>of</strong> centroid point <strong>of</strong> the ith leg VGi r<br />
described as<br />
is<br />
r<br />
VGi<br />
lˆ<br />
r r<br />
i<br />
r r m2<br />
= [ VP<br />
− ( V<br />
i P ⋅ u<br />
i i ) ui<br />
] +<br />
L<br />
m + m<br />
r<br />
V (13)<br />
Pi<br />
i<br />
where VPi r is the velocity vector <strong>of</strong> the centroid point <strong>of</strong><br />
upper joint <strong>of</strong> the ith leg Pi.<br />
© 2011 ACADEMY PUBLISHER<br />
Z B<br />
1<br />
2<br />
The kinetic energy <strong>of</strong> the ith rod is described as<br />
1 r r<br />
T<br />
K L = ( m<br />
i 1 i + m2i<br />
) VG<br />
V<br />
i G<br />
(14)<br />
i<br />
2<br />
Then, substituting equation (13) for equation (14), the<br />
kinetic energy <strong>of</strong> rod can be rewritten as<br />
1<br />
r r r<br />
T<br />
T r r r<br />
T<br />
K L = ( m<br />
i 1i<br />
+ m2i<br />
)[ hiVb<br />
V<br />
i b − k<br />
i iVb<br />
u<br />
i i ( ui<br />
) Vb<br />
(15)<br />
i<br />
2<br />
lˆ<br />
2<br />
where i m2<br />
i<br />
m2<br />
i 2<br />
hi<br />
= [ + ] ; ki<br />
= hi<br />
−[<br />
] .<br />
Li<br />
m1i<br />
+ m2i<br />
m1i<br />
+ m2i<br />
The total mass <strong>of</strong> six legs can be written as[3]<br />
T<br />
2<br />
T<br />
1<br />
6<br />
2]<br />
∑<br />
i=<br />
1<br />
M = [ J ( H − J K J ) J ( m + m ) (16)<br />
legs<br />
legs<br />
where Klegs denotes the total kinetic energy <strong>of</strong> six legs,<br />
rT<br />
rT<br />
rT<br />
rT<br />
rT<br />
rT<br />
−1<br />
J1<br />
= diag{<br />
u1<br />
, u2<br />
, u3<br />
, u4<br />
, u5<br />
, u6<br />
} , J 2 = Vbq&<br />
r<br />
.<br />
The Jacobian matrix J for parallel robot may be<br />
described as<br />
J = J1 J2 (17)<br />
Suppose the structure <strong>of</strong> six legs is same, and then the<br />
total kinetic energy <strong>of</strong> six legs may be expressed as<br />
6<br />
1 T<br />
Klegs = ∑ K L = q&<br />
M legs ( q)<br />
q&<br />
(18)<br />
i 2<br />
i=<br />
1<br />
The potential energy <strong>of</strong> six legs may be written as<br />
6<br />
lˆ<br />
m2i<br />
Plegs<br />
=∑ [( m1i<br />
+ m2i<br />
)( + )( z p − xb′<br />
Sβ<br />
+ y CβSγ<br />
z CβCγ<br />
)]<br />
i b′<br />
+ ′<br />
i<br />
L m<br />
1<br />
1 m<br />
i<br />
i i +<br />
=<br />
2i<br />
(19)<br />
Lagrange equation for parallel robot can be expressed<br />
as<br />
d ⎡∂<br />
L(<br />
q,<br />
q&<br />
) ⎤ ∂L(<br />
q,<br />
q&<br />
)<br />
⎢ ⎥ − = τ i , i=1,2,• • •,.n (20)<br />
dt ⎣ ∂q&<br />
i ⎦ ∂qi<br />
where q∈R n is nominal coordinate; L is the Lagrange<br />
function <strong>of</strong> mechanism system; τi is the force on the ith<br />
nominal coordinate.<br />
The dynamic Lagrange equation <strong>of</strong> 6-DOF rigid<br />
parallel robot is described as<br />
M ( q)<br />
q&<br />
& ( t)<br />
+ V ( q,<br />
q&<br />
) q&<br />
( t)<br />
+ G(<br />
q)<br />
+ τ = J τ ( t)<br />
(21)<br />
6<br />
m<br />
where q, q&<br />
, q&<br />
& ∈ R is the position, the velocity and the<br />
acceleration <strong>of</strong> centroid point <strong>of</strong> mobile platform<br />
respectively; M(q) is the mass <strong>of</strong> mobile platform and six<br />
legs; ( q,<br />
q&<br />
) is the velocity vector <strong>of</strong> mobile platform<br />
V m<br />
and six legs; G(q)∈R 6 is the gravity vector <strong>of</strong> mobile<br />
platform and six legs; τ(t)∈R 6 is the control force vector<br />
<strong>of</strong> mobile platform and six legs; τd∈R n is the model error<br />
and external disturbance; M(q), Vm ( q,<br />
q&<br />
) , G(q) can be<br />
computed by the equations <strong>of</strong> the kinetic energy and the<br />
potential energy <strong>of</strong> mobile platform and six legs.<br />
If the model designed is precise, control law <strong>of</strong> robot is<br />
expressed as<br />
T<br />
( t) = j [ M ( q)(<br />
qd<br />
− kve<br />
− k pe)<br />
+ V ( q,<br />
q)<br />
q + G(<br />
q)]<br />
− τ &<br />
&<br />
& & (22)<br />
where qd is expect angle, e = q − qd, e& = q&<br />
− q&<br />
. d<br />
Then, substituting equation (22) for equation (21), the<br />
control equation <strong>of</strong> stable close-loop system is expressed<br />
as<br />
1<br />
d<br />
1i<br />
T<br />
2i
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1929<br />
+ k e + k e = 0<br />
& (23)<br />
e v p &<br />
Because it is difficult to build the real model <strong>of</strong> the<br />
object precisely, desired model is only built. Its control<br />
law is expressed as<br />
T<br />
( t) = J [ M o ( q)(<br />
qd<br />
− kve<br />
− k pe)<br />
+ Vo<br />
( q,<br />
q)<br />
q + Go<br />
( q)]<br />
− τ<br />
&<br />
&<br />
& &<br />
(24)<br />
Then, substituting equation (24) for equation (21), the<br />
below equation is established.<br />
M ( q)<br />
q&<br />
& + C(<br />
q,<br />
q&<br />
) q&<br />
+ G(<br />
q)<br />
= M o ( q)(<br />
q&<br />
& d − kve&<br />
− k pe)<br />
+ ( q,<br />
q&<br />
) q&<br />
+ G ( q)<br />
+ F(<br />
t)<br />
(25)<br />
Vo o<br />
If ⊿M=MO − M, ⊿V=VO − V, ⊿G=GO − G, the<br />
below equation is obtained.<br />
−1 e& & + kve&<br />
+ k pe<br />
= M O ( ΔMq&<br />
& + ΔVq&<br />
+ ΔG)<br />
(26)<br />
From equation (26), the decline <strong>of</strong> the control<br />
performance is brought up partly by the parameter and<br />
non-parameter uncertainty. Thus, the compensation for<br />
the uncertainty is needed for improving the control<br />
precision <strong>of</strong> the robot.<br />
Ⅲ. STRUCTURE OF CONTROL SYSTEM OF PARALLEL<br />
ROBOT<br />
According to dynamics equation <strong>of</strong> the robot and its<br />
control law, sliding mode control method and fuzzy<br />
support vector machines are used to compensate for the<br />
uncertainty. Structure <strong>of</strong> control system for robots is<br />
shown in Fig.2. In the figure, q denotes the real track <strong>of</strong><br />
robot; qd denotes the expect track; e denotes the error<br />
vector; FSVMC denotes the fuzzy support machines<br />
controller; FSMC denotes the fuzzy sliding mode<br />
controller; R(e) denotes the switching function, whose<br />
inputs are e and e& .<br />
qd<br />
q& & d<br />
q& d<br />
kv<br />
k<br />
kv<br />
p<br />
M O (q)<br />
u<br />
u<br />
1<br />
2<br />
GO (q)<br />
τ<br />
CO ( q,<br />
q&<br />
)<br />
Figure 2. Structure <strong>of</strong> control system for robots<br />
Compensation control law for fuzzy sliding mode is<br />
defined as<br />
τ i ( t ) = M ( q)(<br />
q&<br />
& d − kve&<br />
− k pe)<br />
+ V ( q,<br />
q&<br />
) q&<br />
+ G(<br />
q)<br />
+ ui<br />
, i = 1.<br />
2<br />
(27)<br />
where u1 is the control compensation <strong>of</strong> fuzzy sliding<br />
mode controller; u2 is the control compensation <strong>of</strong> fuzzy<br />
support vector machines controller .<br />
Function R(e) is used to decided which controller as<br />
the compensation controller. Suppose the thickness <strong>of</strong><br />
boundary layer is Q ; if R ( e)<br />
> Q , FSMC is used for<br />
control compensation; if R ( e)<br />
< Q , FSVMC is used for<br />
© 2011 ACADEMY PUBLISHER<br />
q&<br />
q<br />
control compensation; if R ( e)<br />
= Q , the below sliding<br />
algorithm is used for control compensation.<br />
u ( t)<br />
= ( 1 − d ( e))<br />
u2<br />
( t)<br />
+ d ( e)<br />
u1(<br />
t)<br />
+ τ (28)<br />
where τ is the control torque; u1 and u2 are the outputs <strong>of</strong><br />
FSMC and FSVMC respectively; d(e) is the sliding<br />
function, whose function is to make FSMC and FSVM<br />
switch smoothly.<br />
⎧ 0,<br />
E(<br />
t)<br />
∈AFSVMC<br />
⎪<br />
d(<br />
e)<br />
= ⎨0<br />
< d(<br />
e)<br />
< 1,<br />
E(<br />
t)<br />
∈ A − AFSVMC<br />
(29)<br />
⎪<br />
⎩ 1,<br />
AFSMC<br />
where AFSVMC is the control range <strong>of</strong> fuzzy support<br />
machines controller; AFSMC is the control range <strong>of</strong> sliding<br />
mode controller. They are defined as<br />
A = E | e ≤ Q<br />
(30)<br />
{ } p<br />
{ E e ≤ + ζ }<br />
FSVMC<br />
A = | Q<br />
(31)<br />
p<br />
where ζ is the thickness <strong>of</strong> switching layer, 0 < ζ < Q . P<br />
norm is defined as<br />
1/<br />
p<br />
p ⎪⎫<br />
i ⎬<br />
⎪⎧<br />
e = e<br />
p ⎨∑<br />
(32)<br />
⎪⎩ i ⎪⎭<br />
According to P norm definition, the sliding function<br />
used is described as<br />
d( e)<br />
= max( 0,<br />
sat(<br />
R − Q)<br />
/ ζ ) (33)<br />
⎧ y,<br />
where sat ( y)<br />
= ⎨<br />
⎩sgn(<br />
y),<br />
y < 1,<br />
y = e .<br />
p<br />
If R ( e)<br />
< Q , e∈AFSVMC, and d(e)=0; if R ( e)<br />
> Q + ζ ,<br />
e∈AFSVMC, and d(e)=1. If A′ = A − AFSVMC<br />
, 0
1930 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
Suppose sliding mode plane s=0, the below equation is<br />
obtained.<br />
ce + e&<br />
= 0<br />
(35)<br />
The control compensation <strong>of</strong> sliding mode may be<br />
defined as<br />
u = τ = −Ksign(<br />
s)<br />
(36)<br />
1<br />
1<br />
where K is the coefficient matrix, K≥M ( ΔMq<br />
+ ΔCq<br />
+ ΔG+<br />
τd)<br />
− && & .<br />
Lyapunov function is defined as<br />
1<br />
V = (37)<br />
2<br />
2s<br />
The below equation is obtained by equation (1), (7)<br />
and (8).<br />
V = ss&<br />
≤ −K(<br />
t)<br />
| s | < 0<br />
(38)<br />
Under the new control law, sliding mode exists and<br />
may be attained. Sliding mode switches at the sliding<br />
mode plane s=0, which brings the strong oscillation.<br />
In the sliding mode control law, the switching gain<br />
K(t), which is used to compensate the uncertainty, easily<br />
arises the chattering. For reducing the chattering, K(t)<br />
should be varied with the time.<br />
Fuzzy rules may be expressed as:<br />
If s>0, K(t) should be increased; If s0.<br />
© 2011 ACADEMY PUBLISHER<br />
Substituting K ˆ ( t)<br />
for K(t) in equation (36), control law<br />
can be rewritten as<br />
u = −Kˆ<br />
× sign(<br />
s)<br />
(40)<br />
1<br />
Ⅴ. FUZZY SUPPORT VECTOR MACHINES CONTROLLER<br />
A. Structure <strong>of</strong> FSVM controller<br />
Structure <strong>of</strong> FSVM controller is shown in Fig.5. Inputs<br />
<strong>of</strong> FSVM system are { qd, q& d }; the output compensation<br />
<strong>of</strong> FSVM system for uncertainty is u2; qd is the desired<br />
positions <strong>of</strong> two joints; q is the real positions <strong>of</strong> two<br />
joints; e is the position error <strong>of</strong> two joints; e& is the varying<br />
rate <strong>of</strong> position error <strong>of</strong> two joints.<br />
qd<br />
+<br />
_<br />
e<br />
LS & IGA hybrid<br />
optimization algorithm<br />
d/dt<br />
e&<br />
Ke<br />
E<br />
e& Ke& E& FSVM controller<br />
K(<br />
x,<br />
x1)<br />
K<br />
K(<br />
x,<br />
x1)<br />
e<br />
t<br />
.<br />
.<br />
.<br />
K x,<br />
x )<br />
( 1<br />
U 2<br />
Ku<br />
u2<br />
Figure 5. Structure <strong>of</strong> FSVM control system<br />
Parallel<br />
robot<br />
Inputs and outputs <strong>of</strong> control system are fuzzified; {e,<br />
e& , u2} are fuzzified respectively as {E, E& , U2}; their<br />
fuzzy subsets are {NB, NM, NS, Z, PS, PM, PB}, which<br />
respectively denotes {negative big, negative middle,<br />
negative small, zero, positive small, positive middle,<br />
positive big}; quantified grades <strong>of</strong> them are {-6, -5, -4, -3,<br />
-2, -1, 0, 1, 2, 3, 4, 5, 6}. Triangle distribution function is<br />
selected as their membership function. According to<br />
varying range <strong>of</strong> inputs and outputs <strong>of</strong> control system, the<br />
proportional coefficients <strong>of</strong> fuzzy control are e K , e K & and<br />
Ku respectively. Decision process <strong>of</strong> fuzzy control is<br />
2<br />
described by the three-layer SVM model.<br />
1) Input layer<br />
The function <strong>of</strong> input layer is that input variables are<br />
fuzzified as the input <strong>of</strong> control system x.<br />
⎧E<br />
= Kee<br />
⎪<br />
⎨E&<br />
= Kece&<br />
(41)<br />
⎪<br />
⎩x<br />
= ( E,<br />
E&<br />
)<br />
2) Hidden layer<br />
The function <strong>of</strong> inner layer is to realize the kernel<br />
computation <strong>of</strong> four-dimension input x and SVM. The<br />
radial basis function kernel function is expressed as<br />
2 2<br />
K( x,<br />
x ) = exp( − x − x / 2σ<br />
) (42)<br />
i<br />
where σ is the kernel width, which reflects the radius <strong>of</strong><br />
close boundary.<br />
3) Output layer<br />
The function <strong>of</strong> output layer is to obtain the real input<br />
control value <strong>of</strong> the controlled object by computing SVM<br />
regression value.<br />
i<br />
q
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1931<br />
N ⎛<br />
⎜ y ( xi<br />
) = ∑ akK<br />
( xi,<br />
xk<br />
) + b i = 1,<br />
2 (43)<br />
⎜<br />
k = 1<br />
⎜<br />
⎝ui<br />
= KUi<br />
y(<br />
xi<br />
)<br />
The controller parameters are optimized by the hybrid<br />
learning algorithm. First, least square algorithm is used<br />
for <strong>of</strong>f-line optimize the parameters <strong>of</strong> support vector<br />
machines. Then, improved genetic algorithm is used for<br />
on-line optimizing the parameters <strong>of</strong> support vector<br />
machines and fuzzy proportional coefficients.<br />
B. Parameter optimization <strong>of</strong> control system<br />
The parameters <strong>of</strong> affecting SVM performance are<br />
number <strong>of</strong> training sample set D, penalty coefficient γ,<br />
kernel width σ and insensitive coefficient ε , and so on.<br />
The system performance is also affected by fuzzy<br />
proportional relations between real values and fuzzy<br />
values in decision-making process <strong>of</strong> control system.<br />
Only when these parameters are in finite range, the<br />
system has the better control performance. Optimal<br />
parameter combination varies with the object.<br />
1) Off-line optimization <strong>of</strong> γ and σ<br />
Because ε may be preset by the noise, which reflects<br />
the prediction <strong>of</strong> data noise level by support vector<br />
machines, least square algorithm was only used to <strong>of</strong>fline<br />
optimize γ and σ.<br />
First, the method <strong>of</strong> rising exponent was used to search<br />
the proper γ set and σ set. For example, γ = 2 -4 , 2 -2 , …,<br />
2 10 ;σ = 2 -10 , 2 -8 ,… , 2 -2 .<br />
Second, using the method <strong>of</strong> net search, the parameter<br />
combination(γ , σ) was selected to verify it crossly. The<br />
training sample set was divided into S groups {G1,<br />
G2,⋅⋅⋅GS}. Selecting randomly S-1 groups as training set,<br />
and another as verifying set; generalization capability was<br />
evaluated with the following equation.<br />
S<br />
N N 2<br />
P = ∑∑ y − y x i<br />
i= 1 y∈Gi<br />
)) x | ( ( (44)<br />
where Gi is ith group verifying set; y N is the sample <strong>of</strong><br />
verifying set; xi is the parameter vector [a,b] where D - Gi<br />
was set as train sample; y(x|θi) is the output <strong>of</strong> SVM<br />
system.<br />
Final, the parameter combination was selected<br />
circularly to verify it crossly and the performance index P<br />
was computed until the optimal parameter combination<br />
(γ , σ) is obtained.<br />
2) On-line optimization <strong>of</strong> fuzzy SVM parameters<br />
(a) Encoding<br />
Because <strong>of</strong> the complexity and continuity <strong>of</strong><br />
optimizing process <strong>of</strong> SVM parameters, the coding<br />
method <strong>of</strong> float number is used to avoid the effect <strong>of</strong><br />
binary coding on the evaluation <strong>of</strong> algorithm performance<br />
and computing precision.<br />
(b) Selection <strong>of</strong> fitness function<br />
In improved genetic algorithm, the individual<br />
evolution is decided by individual fitness value. Thus the<br />
individual fitness value need be computed. The individual<br />
is sequenced by fitness value and sequenced population is<br />
© 2011 ACADEMY PUBLISHER<br />
lined out by the upper limit and lower limit. Fitness<br />
function is used to evaluate SVM individual and fitness<br />
function designed influences directly the performance <strong>of</strong><br />
genetic algorithm. According to the feature <strong>of</strong> robot<br />
system, fitness function was described by the sum <strong>of</strong><br />
error among given system input and real output. It was<br />
expressed as<br />
M<br />
Fi = Ei<br />
max<br />
k = 1<br />
i i<br />
− Σ S ( k)<br />
−T<br />
( k)<br />
(45)<br />
where i = 1,2,⋅⋅⋅, N is the number <strong>of</strong> individual in<br />
population; k is the number <strong>of</strong> individual variable.<br />
Mean error EMSE <strong>of</strong> system track was expressed as<br />
N1<br />
1<br />
2 2 ∑ [( Ti<br />
− f ( xi<br />
)) ]<br />
i=<br />
1 EMSE<br />
=<br />
(46)<br />
N<br />
(c) Genetic operation<br />
Genetic operations include selection, crossover and<br />
mutation. Its objective is to substitute the new generation<br />
population into next generation population. The<br />
procedure <strong>of</strong> operation is given as follows:<br />
Step1: Generation <strong>of</strong> initial population;<br />
Step2: Re-evaluation and adding age;<br />
Step3: Selection <strong>of</strong> parents: prior selection <strong>of</strong> elder<br />
individuals;<br />
Step4: Crossover and mutation: generation <strong>of</strong> new<br />
individuals;<br />
Step5: Evaluation: evaluation <strong>of</strong> new individuals;<br />
Step6: Natural selection: selection considering the<br />
diverseness <strong>of</strong> individuals;<br />
Step7: Steps 2 to 6 are repeated until the convergence<br />
is achieved.<br />
In genetic operation, set population is 200; set<br />
crossover probability is 0.75; set mutation probability is<br />
0.02. Each parameter was set by following:<br />
1 255<br />
1 31<br />
D ∈[<br />
1,<br />
512]<br />
, γ ∈[<br />
, 255 ] σ ∈[<br />
, 127 ]<br />
256 256 , 32 32 ,<br />
1 63 1 15<br />
ε ∈[<br />
, 255 ] e [ , 255 ]<br />
64 64 , 16 16<br />
∈ K<br />
1 63<br />
Ke& ∈[<br />
, 255 ]<br />
, 64 64 ,<br />
1 127<br />
Ku<br />
∈[<br />
, 15 ]<br />
128 128 .<br />
where D, γ, σ and ε are coded respectively by 8 bit (8 bit<br />
integer), 9 bit (9 bit integer),14 bit(8 bit integer, 6 bit<br />
decimal),16bit (8 bit integer, 8 bit decimal) binary strings;<br />
fuzzy proportional coefficients Ke, e K & , Ku are coded<br />
respectively by 12 bit (8 bit integer, 4 bit decimal), 10 bit<br />
(4 bit integer, 6 bit decimal), 11bit (4 bit integer, 7 bit<br />
decimal), binary strings. Thus, they are coded by 72 bit<br />
binary strings and their values are discrete. Their units are<br />
1, 1/256, 1/64, 1/32, 1/16, 1/64, 1/128 respectively. After,<br />
the individual fitness function are computed using these<br />
parameters, the individuals in the new population are<br />
selected by the desired value.<br />
Ⅵ. SIMULATION AND APPLICATION<br />
To verify the effectiveness <strong>of</strong> the presented control<br />
strategy for the parallel robot, the comparative simulation<br />
experimental researches were carried out between the<br />
1
1932 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
designed control strategy and the fuzzy sliding mode<br />
control strategies using experimental simulation. In<br />
simulation experiment <strong>of</strong> control performance, the mobile<br />
platform is driven by six asymmetric cylinders with a<br />
cylinder diameter <strong>of</strong> 85mm and a rod diameter <strong>of</strong> 64mm,<br />
and a full stroke <strong>of</strong> 840mm, which are controlled by six<br />
servo-electromotors. The installed sensors measure the<br />
leg lengths and forces between the centroid point <strong>of</strong> rods<br />
and the heads <strong>of</strong> the cylinders. The radius <strong>of</strong> the base<br />
platform and the mobile platform are 1250 and 540mm<br />
respectively. The simulation experiments <strong>of</strong> parallel robot<br />
were conducted by the simulation s<strong>of</strong>tware. In the<br />
simulation experiments, the experimental values (100, 15,<br />
1.0, 0.01, 100,1.0,0.1) are set as the initial values <strong>of</strong><br />
control parameter combination (D, γ, σ, ε, Ke, e K & , Ku) ,<br />
trace error EMSE is 2.134; after hybrid optimization,<br />
optimal parameter combination (D, γ, σ, ε, Ke, e K & , Ku) is<br />
(1.6, 3.2, 0.2, 1.5, 65, 0.4, 0.07), trace error EMSE is 0.014.<br />
The experiments concerned position tracking <strong>of</strong> centroid<br />
point <strong>of</strong> mobile platform for the following reference<br />
trajectories qd(t)=1.0+0.20sin(2πt) mm and<br />
qd(t)=1.0+0.40sin(2πt) mm by the designed control<br />
system. The experimental results are shown in Fig.6 and<br />
Fig.7.<br />
Figure 6. Position tracking (qd(t)=1.0+0.20sin(2πt))<br />
Figure 7. Position tracking (qd(t)=1.0+0.40sin(2πt))<br />
The experiments concern the position tracking error <strong>of</strong><br />
centroid point <strong>of</strong> mobile platform for the following<br />
reference trajectories qd(t)=1.0+0.20sin(2πt) mm and<br />
qd(t)=1.0+0.40sin(2πt) mm by the designed control<br />
system. The experimental results are shown in Fig.8 and<br />
Fig.9.<br />
© 2011 ACADEMY PUBLISHER<br />
It can be seen that the designed control system<br />
performs much better than the fuzzy sliding mode control<br />
methods from Fig.6, Fig.7, Fig.8 and Fig 9. It can be<br />
obtained that position tracking error <strong>of</strong> centroid point <strong>of</strong><br />
mobile platform for the following reference trajectory<br />
qd(t)=1.0+0.40sin(2πt) mm is smaller than that for the<br />
following reference trajectory qd(t)=1.0+0.20sin(2πt) mm<br />
by the designed control system. From above figures,<br />
using designed controller, track error is low. Considering<br />
uncertainty and complexity <strong>of</strong> the system, the track error<br />
may be permitted.<br />
Figure 8. Position tracking error (qd(t)=1.0+0.20sin(2πt))<br />
Figure 9. Position tracking error (qd(t)=1.0+0.40sin(2πt))<br />
The experiments concern the velocity tracking <strong>of</strong><br />
centroid point <strong>of</strong> mobile platform for the following<br />
reference trajectories qd(t)=1.0+0.20sin(2πt) mm and<br />
qd(t)=1.0+0.40sin(2πt) mm by the designed control<br />
system. The experimental results are shown in Fig.10 and<br />
Fig.11.<br />
Figure 10. Velocity tracking (qd(t)=1.0+0.20sin(2πt))
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1933<br />
Figure 11. Velocity tracking (qd(t)=1.0+0.40sin(2πt))<br />
The experiments concern the velocity tracking error <strong>of</strong><br />
centroid point <strong>of</strong> mobile platform for the following<br />
reference trajectories qd(t)=1.0+0.20sin(2πt) mm and<br />
qd(t)=1.0+0.40sin(2πt) mm by the designed control<br />
system. The experimental results are shown in Fig.12 and<br />
Fig.13.<br />
Figure 12. Velocity tracking error (qd(t)=1.0+0.20sin(2πt))<br />
Figure 13. Velocity tracking error (qd(t)=1.0+0.40sin(2πt))<br />
It can be seen that the designed control system<br />
performs much better than the fuzzy sliding mode control<br />
methods in the velocity tracking from Fig.10, Fig.11,<br />
Fig.12 and Fig.13. It can be obtained that velocity<br />
tracking error <strong>of</strong> centroid point <strong>of</strong> mobile platform for the<br />
following reference trajectory qd(t)=1.0+0.40sin(2πt) mm<br />
is smaller than that for the following reference trajectory<br />
qd(t)=1.0+0.20sin(2πt) mm by the designed control<br />
system.<br />
© 2011 ACADEMY PUBLISHER<br />
The experiments concern control input <strong>of</strong> legs for the<br />
following reference trajectories qd(t)=1.0+0.20sin(2πt)<br />
mm and qd(t)=1.0+0.40sin(2πt) mm by the designed<br />
control system. The experimental results are shown in<br />
Fig.14 and Fig.15. It can be seen that the control input <strong>of</strong><br />
legs are different and the control input for the following<br />
reference trajectory qd(t)=1.0+0.40sin(2πt) mm is<br />
different from that for the following reference trajectory<br />
qd(t)=1.0+0.20sin(2πt) mm by the designed control<br />
system.<br />
Figure 14. Control input <strong>of</strong> leg (qd(t)=1.0+0.20sin(2πt))<br />
Figure 15. Control input <strong>of</strong> leg (qd(t)=1.0+0.40sin(2πt))<br />
Ⅶ. CONCLUSION<br />
In this paper, the dynamics equation <strong>of</strong> 6-DOF parallel<br />
robot was established. According to the dynamics<br />
equation, a fuzzy support vector machines control<br />
strategy based on the sliding mode control was proposed.<br />
The proposed controller consists <strong>of</strong> a fuzzy sliding mode<br />
controller and a fuzzy support vector machines controller.<br />
The compensation controller is decided by comparing the<br />
switching function with the thickness <strong>of</strong> boundary layer.<br />
Using improved GA and FL algorithm to optimize the<br />
performance parameters <strong>of</strong> support vector machines and
1934 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
the fuzzy proportional parameters, a better control system<br />
was obtained. The system uncertainty and external<br />
disturbance was compensated. Experimental simulation<br />
was carried out with 6-DOF parallel robot to investigate<br />
the effectiveness <strong>of</strong> the proposed control method. The<br />
simulation results show that the control method designed<br />
gets tracking effect with high precision and speed, as well<br />
as reduces the chattering under the condition <strong>of</strong> existing<br />
model error and external disturbance.<br />
REFERENCES<br />
[1] I. A. Bonev and J. A. Ryu, “New method for solving the<br />
direct kinematics <strong>of</strong> general 6-6 Stewart platforms using<br />
three linear extra sensors.” Mechanism and Machine<br />
Theory, vol. 35, pp. 425- 436, March 2000.<br />
[2] I. Davliakos and E. Papadopoulos, “Model-based control<br />
<strong>of</strong> a 6-d<strong>of</strong> electro-hydraulic Stewart–Gough platform.”<br />
Mechanism and Machine Theory, vol.43, pp.1385-1400,<br />
November 2008.<br />
[3] M. R. Sirouspour and S. E. Salcudean, “Nonlinear control<br />
<strong>of</strong> hydraulic robots.” IEEE Transactions on Robotics and<br />
Automation, vol.17, pp.173-182, April 2001.<br />
[4] Xiufeng Zhang and Lining Sun, “Research <strong>of</strong> precise<br />
parallel robot control method and system.” Chinese<br />
<strong>Journal</strong> <strong>of</strong> Mechanical Engineering, vol. 40, pp.177-180,<br />
April 2004.<br />
[5] B. Dasgupta and T. S. Mruthyunjaya, “The Stewart<br />
platform manipulator: A review.” Mechanism and Machine<br />
Theory, vol.35, pp.15-40, January 2000.<br />
[6] Hongbo Guo, Yongguang Liu, Guirong Liu and HongRen<br />
Li, “Cascade control <strong>of</strong> a hydraulically driven 6-DOF<br />
parallel robot manipulator based on a sliding mode.”<br />
Control Engineering Practice, vol.16, pp.1055-1068,<br />
September 2008.<br />
[7] S. Islam , P X Liu, “Output feedback sliding mode control<br />
for robot manipulators.” Robotica, vol.28, pp.975-987,<br />
December 2010.<br />
[8] J. S. Oh, J. B. Park and Y. H. Choi, “Stable path tracking<br />
control <strong>of</strong> a mobile robot using a wavelet based fuzzy<br />
neural network.” International <strong>Journal</strong> <strong>of</strong> Control<br />
Automation and Systems, vol.3, pp.552-563, December<br />
2005.<br />
[9] B. Achili, B. Daachi, Y. Amirat and A. Ali-cherif, “A<br />
robust adaptive control <strong>of</strong> a parallel robot.” International<br />
<strong>Journal</strong> <strong>of</strong> control, vol.83, pp.2107-2119, October 2010.<br />
[10] Yangjun Pi and Xuanyin Wang, “Trajectory tracking<br />
control <strong>of</strong> a 6-DOF hydraulic parallel robot manipulator<br />
with uncertain load disturbances.” Control Engineering<br />
Practice, vol.11, pp.185-193, January 2011.<br />
© 2011 ACADEMY PUBLISHER<br />
[11] T. Dierks and S. Jagannathan, “Neural Network Output<br />
Feedback Control <strong>of</strong> Robot Formations.” IEEE<br />
Transactions on Systems Man and Cybernetics part B-<br />
Cybernetic, vol.40, pp. 383-399, April 2010.<br />
[12] Rongjong Wai and Pochen Chen, “Robust Neural-Fuzzy-<br />
Network Control for Robot Manipulator Including<br />
Actuator Dynamics.” IEEE transactions on industrial<br />
electronics, vol.53, pp.1328-1349, August 2006.<br />
[13] Shaocheng Qu and Yongji Wang, “Discrete sliding mode<br />
control for a parallel robot.” In Proceedings <strong>of</strong> the fifth<br />
international conference on machine learning and<br />
cybernetics, pp.871-874, Dalian, China, August 2006.<br />
[14] Qidan Zhu, Xunyu Zhong and Bo Xu: “Design <strong>of</strong> Fuzzy-PI<br />
Compound Control system for three-cylinder hydraulic<br />
parallel robot.” In Proceedings <strong>of</strong> the 2007 IEEE<br />
international conference on mechatronics and automation,<br />
pp.989-993, Harbin, China, August 2007.<br />
[15] C.-F. Lin, Sh.-D. Wan., “Fuzzy support vector machines.”<br />
IEEE Transactions on Neural Networks, vol.13, pp.464-<br />
471, March 2002.<br />
[16] F. Orabona, C. Castellini, B. Caputo, L. Jie, and G.<br />
Sandini , “On-line independent support vector machines.”<br />
Pattern Recognition, vol.43, pp.1402-1412, April 2010.<br />
Dequan Zhu received B.S. degree in agricultural mechanism<br />
from Anhui Agricultural University and received M.S. degree in<br />
mechanical and electronic engineering from Hefei University <strong>of</strong><br />
Technology in 1997 and 2005 respectively. Currently, he is an<br />
associate pr<strong>of</strong>essor at Anhui Agricultural University, and a PhD<br />
candidate in automation at University <strong>of</strong> Science and<br />
Technology <strong>of</strong> China. His major research experiences and<br />
interests include modern agricultural equipment and intelligent<br />
control.<br />
Tao Mei received B.S. degree in precision mechanism from<br />
Zhejiang University and received Ph.D. degree in mechanics<br />
from University <strong>of</strong> Science and Technology <strong>of</strong> China in 1982<br />
and 2001 respectively. Currently, he is a researcher at Institute<br />
<strong>of</strong> Intelligent Machines, Chinese <strong>Academy</strong> <strong>of</strong> Science.<br />
His major research experiences and interests include robotics<br />
and intelligent control.<br />
Lei Sun received M.S. degree in control theory and engineering<br />
from Hefei University <strong>of</strong> Technology and received Ph.D. degree<br />
in detect technology and automatic mechanism from University<br />
<strong>of</strong> Science and Technology <strong>of</strong> China in 2004 and 2008<br />
respectively. Currently, he is an lecturer at Anhui Agricultural<br />
University. His major research experiences and interests include<br />
modern agricultural equipment and intelligent control.
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1935<br />
Parameters Optimization <strong>of</strong> Least Squares<br />
Support Vector Machines and Its Application<br />
Chunli Xie 1,2<br />
1. Dalian University <strong>of</strong> Technology/ School <strong>of</strong> Electronic and Information Engineering, Dalian, 116024, China<br />
2. Dalian Nationalities University/College <strong>of</strong> Electromechanical and Information Engineering, Dalian, 116024, China<br />
Email: chunlix@dlnu.edu.cn<br />
Cheng Shao 1 and Dandan Zhao 3<br />
3. Dalian Nationalities University/School <strong>of</strong> Computer Science and Engineering, Dalian, 116024, China<br />
Email: cshao@dlut.edu.cn, zhaodd@dlnu.edu.cn<br />
Abstract—Parameters optimization plays an important role<br />
for the performance <strong>of</strong> least squares support vector<br />
machines (LS-SVM). In this paper, a novel parameters<br />
optimization method for LS-SVM is presented based on<br />
chaotic ant swarm (CAS) algorithm. Using this method, the<br />
optimization model is established, within which the fitness<br />
function is the mean square error (MSE) index, and the<br />
constraints are the ranges <strong>of</strong> the designing parameters.<br />
After having been validated its effectiveness by an artificial<br />
data experiment, the proposed method is then used in the<br />
identification for inverse model <strong>of</strong> the nonlinear underactuated<br />
systems. Finally real data simulation results are<br />
given to show the efficiency.<br />
Index Terms—Least Squares Support Vector Machines,<br />
Parameters Optimization, Chaotic Ant Swarm Algorithm<br />
I. INTRODUCTION<br />
A novel statistical learning method called Support<br />
Vector Machines (SVM) was presented by Vapnik in<br />
1995. Due to the advantages such as the complete<br />
statistical learning theory foundation and perfect study<br />
ability, SVM has become quite an active research field in<br />
machine learning and broadly used in many fields such as<br />
pattern recognition and regression estimation problems<br />
[1, 2]. The classical training algorithm <strong>of</strong> SVM is<br />
equivalent to solving a quadratic programming with<br />
linear and inequality constraints. Least squares support<br />
vector machines (LS-SVM) has been recently introduced<br />
by Suykens et al. as reformulations to standard SVM [3,<br />
4], which simplifies the training process <strong>of</strong> standard SVM<br />
in a great extent by replacing the inequality constraints<br />
with equality ones. The simplicity <strong>of</strong> LS-SVM promotes<br />
the applications <strong>of</strong> SVM, and many pattern recognition<br />
and function approximation problems have been tackled<br />
with LS-SVM in the last decade [5-9].<br />
The parameters in regularization item and kernel<br />
function are called parameters in LS-SVM, which play an<br />
important role for the algorithm performance. The<br />
Corresponding author: Chunli Xie.<br />
© 2011 ACADEMY PUBLISHER<br />
doi:10.4304/jcp.6.9.1935-1941<br />
existing techniques for tuning the parameters in LS-SVM<br />
can be summarized into two kinds: one is based on<br />
analytical techniques, the other is based on heuristic<br />
searches. The first kind <strong>of</strong> techniques determines the<br />
parameters with gradients <strong>of</strong> some generalized error<br />
measures [10]. And the second kind <strong>of</strong> techniques<br />
determines the parameters with modern heuristic<br />
algorithms including genetic algorithms (GA), simulated<br />
annealing algorithms (SA), particle swarm optimization<br />
algorithms (PSO) and other evolutionary algorithms [11-<br />
15]. Iterative gradient-based algorithms rely on smoothed<br />
approximation <strong>of</strong> a function. So, it does not ensure that<br />
the search direction points exactly to an optimum <strong>of</strong> the<br />
generalization performance measure which is <strong>of</strong>ten<br />
discontinuous. Grid search [16] is one <strong>of</strong> the conventional<br />
approaches to deal with discontinuous problems.<br />
However, it needs an exhaustive search over the space <strong>of</strong><br />
parameters, which must be time consuming. This<br />
procedure also needs to locate the interval <strong>of</strong> feasible<br />
solution and a suitable sampling step.<br />
In this paper, a novel algorithm <strong>of</strong> parameters<br />
optimization is presented based on the principles <strong>of</strong> the<br />
chaotic ant swarm (CAS) algorithm. Inspired by the<br />
chaotic and self-organizing behavior <strong>of</strong> the ants in nature,<br />
the novel CAS [17] algorithm is developed in 2006,<br />
which combines the chaotic behavior <strong>of</strong> individual ant<br />
with the intelligent foraging actions <strong>of</strong> ant colony via the<br />
organization variable for solving optimization problems.<br />
Similar to GA, EA and PSO, the CAS algorithm is a<br />
population-based optimization tool, which searches for<br />
optima by updating generations. However, unlike GA and<br />
EA, the CAS algorithm does not need evolutionary<br />
operators such as crossover and mutation. Compared to<br />
GA and EA, the advantages <strong>of</strong> CAS algorithm are that it<br />
possesses the capability to escape from local optima, is<br />
easy to be implemented, and has fewer parameters to be<br />
tuned. Compared to PSO, the advantages <strong>of</strong> CAS<br />
algorithm are that it has higher convergent precision. The<br />
CAS algorithm has been successfully applied to<br />
parameters estimation, artificial network training and<br />
fuzzy system control, etc [18-26]. The CAS algorithm is<br />
used to the parameters optimization <strong>of</strong> LS-SVM, and the
1936 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
feasibility <strong>of</strong> this approach is examined on the testing<br />
function and nonlinear under-actuated systems.<br />
This paper is organized as follows. The LS-SVM<br />
regression algorithm is briefly reviewed in Section 2.<br />
Parameters optimization algorithm based on the CAS<br />
algorithm is addressed in Section 3. The results <strong>of</strong> testing<br />
and simulation are presented to demonstrate the<br />
effectiveness <strong>of</strong> the proposed method in Section 4. The<br />
application <strong>of</strong> LS-SVM based on the CAS Algorithm is<br />
given in Section 5. Finally, the paper is concluded in<br />
Section 6.<br />
II. LS_SVM REGRESSION<br />
The LS-SVM, evolved from the SVM, changes the<br />
inequality constraint <strong>of</strong> a SVM into an equality constraint<br />
and forces the sum <strong>of</strong> squared error (SSE) loss function to<br />
become an experience loss function <strong>of</strong> the training set.<br />
Then the problem has become one <strong>of</strong> solved linear<br />
programming problems. This can be specifically<br />
described as follows [4]:<br />
Given the following training sample set (D):<br />
{ ( , y ) k = 1,<br />
2,<br />
L,<br />
N}<br />
D = x k k<br />
where N is the total number <strong>of</strong> training data pairs,<br />
k<br />
n<br />
R ∈ x is the regression vector and ∈ R is the<br />
n<br />
output. According to SVM theory, the input space R is<br />
mapped into a feature space, and then the linear equation<br />
in the feature space can be defined as:<br />
T<br />
f ( x) = w ϕ ( x)<br />
+ b<br />
(1)<br />
h<br />
where the nonlinear mapping ϕ : R → R maps the<br />
input data into a so-called high dimensional feature space<br />
(which can be infinite dimension). The regularized cost<br />
function <strong>of</strong> the LS-SVM is given as:<br />
where,<br />
1 T 1<br />
min J ( w , e)<br />
= w w + γ<br />
2 2<br />
T<br />
n<br />
y k<br />
N<br />
2<br />
∑ ek<br />
k = 1<br />
s. t.<br />
yk<br />
= w ϕ ( xk<br />
) + b + ek<br />
, k = 1,<br />
2,<br />
L,<br />
N (2)<br />
h n<br />
w ∈ R is the weight vector, ek ∈ R is slack<br />
variable, b ∈ R is a bias term and γ ∈ R is regularization<br />
item. The Lagrangian corresponding to Eq. (2) can be<br />
defined as follows:<br />
L ( w,b,e; α)<br />
=<br />
N<br />
−∑<br />
k = 1<br />
k<br />
T { w ( x ) + b + e y }<br />
J ( w,e) α ϕ<br />
− (3)<br />
where α k ∈ R(<br />
k = 1,<br />
2,<br />
L,<br />
N ) are the Lagrange multipliers.<br />
The KKT conditions can be expressed by<br />
N<br />
∑<br />
k = 1<br />
© 2011 ACADEMY PUBLISHER<br />
k<br />
w = α ϕ(x<br />
)<br />
(4)<br />
k<br />
k<br />
k<br />
n<br />
k<br />
T<br />
α = γe<br />
(5)<br />
N<br />
∑<br />
k = 1<br />
k<br />
k<br />
α = 0<br />
(6)<br />
k<br />
w ϕ ( x ) b + e − y = 0<br />
(7)<br />
k<br />
+ k k<br />
After elimination <strong>of</strong> w and e k , the solution <strong>of</strong> the<br />
optimization problem can be obtained by solving the<br />
following set <strong>of</strong> linear equations<br />
⎡b⎤<br />
⎡0<br />
⎢ ⎥ = ⎢v<br />
⎣α⎦<br />
⎢⎣<br />
T<br />
1 N<br />
T N<br />
, N ] ∈ R<br />
with y = [ y , L,<br />
y ] ∈ R ,<br />
v −1<br />
T ⎤ ⎡0⎤<br />
−1 ⎥ ⎢ ⎥<br />
Ω + γ I ⎥⎦<br />
⎣ y⎦<br />
r¡<br />
N<br />
= [ 1,<br />
L,<br />
1<br />
T<br />
]<br />
N<br />
∈ R<br />
α = [ α1 , L α and Ω is an N × N kernel matrix.<br />
By using the kernel trick [2], one obtains<br />
T<br />
Ω = ϕ ( x ) ϕ(<br />
x ) = ( x , x ) , ∀k,<br />
l = 1,<br />
2,<br />
L,<br />
N.<br />
kl<br />
k<br />
l<br />
K k l<br />
And the resulting LS-SVM regression model becomes<br />
N<br />
∑<br />
k = 1<br />
(8)<br />
f ( x) = α K(<br />
x,<br />
x ) + b<br />
(9)<br />
where α k , b are the solution to Eq. (8).<br />
k<br />
Note that the dot product ϕ ⋅) ϕ(<br />
⋅)<br />
in the feature space<br />
( T<br />
is replaced by a prechosen kernel function K( ⋅,<br />
⋅)<br />
due to<br />
the employment <strong>of</strong> the kernel trick. Thus, there is no need<br />
to construct the feature vector w or to know the nonlinear<br />
mapping ϕ(⋅) explicitly. Given a training set, the training<br />
<strong>of</strong> an LS-SVM is equal to solving a set <strong>of</strong> linear equations<br />
as Eq. (8). This greatly simplifies the regression problem.<br />
The chosen kernel function must satisfy the Mercer’s<br />
condition [2]. Possible kernel functions are, e.g.:<br />
Linear kernel<br />
K( x , x ) = x ⋅ x .<br />
k<br />
Polynomial kernel<br />
k<br />
l<br />
K ( x , x ) = ( x ⋅ x + 1)<br />
.<br />
l<br />
Gaussian RBF kernel<br />
k<br />
k<br />
l<br />
K( x , x ) = exp( − x − x / 2 ) .<br />
k<br />
l<br />
l<br />
k<br />
m<br />
k<br />
2 2<br />
l σ<br />
where d denotes the polynomial degree, σ is the kernel<br />
(bandwidth) parameter.<br />
It is well known that LS-SVM generalization<br />
performance depends on a good setting <strong>of</strong> regularization<br />
parameter and the kernel parameter. In order to achieve<br />
the better generalization performance, it is necessary to<br />
select and optimize these parameters.<br />
III. PARAMETERS OPTIMIZATION OF LS_SVM BASED ON<br />
CAS ALGORITHM
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1937<br />
A. Overview <strong>of</strong> CAS Algorithm.<br />
Ants have attracted many scientists’ significant<br />
interests because their colonies can achieve the selforganizing<br />
behavior and the high level <strong>of</strong> structure. Most<br />
<strong>of</strong> the existing ant-inspired optimization algorithms are<br />
based on the random meta-heuristic <strong>of</strong> nondeterministic<br />
probability theory. However, Cole suggested that ant<br />
colony exhibits a periodic behavior while single ant show<br />
low-dimensional deterministic chaotic activity patterns<br />
[27]. From the view <strong>of</strong> dynamics, the chaotic behavior <strong>of</strong><br />
single ant has some relation to the self-organizing and<br />
foraging behaviors <strong>of</strong> ant colony. The chaotic behavior <strong>of</strong><br />
individual ant and the intelligent organization actions <strong>of</strong><br />
ant colony are adaptations to the environment. These<br />
behaviors help the ants to find food and survive.<br />
According to the theory, a novel optimization algorithm,<br />
called CAS algorithm, was presented.<br />
In the CAS algorithm, the chaotic system<br />
θ + = θ exp( µ ( 1−θ<br />
)) [28] was introduced into the<br />
n 1 n<br />
n<br />
heuristic equation <strong>of</strong> the CAS algorithm for obtaining the<br />
chaotic search initially. The adjustment <strong>of</strong> the chaotic<br />
behaviour <strong>of</strong> individual ant is achieved by the<br />
introduction <strong>of</strong> a successively decrement <strong>of</strong> organization<br />
variable µ i and leads the individual to move to the new<br />
site acquired with the best fitness value eventually.<br />
( pid − θid<br />
) is introduced to achieve the information<br />
exchange <strong>of</strong> individuals and the movements to new site<br />
taken on the best fitness value. pid is selected based on<br />
the fitness theory which is very widely developed in<br />
optimization theory such as genetic algorithm and tabu<br />
search, and so on. θid is the state <strong>of</strong> the d th dimension<br />
<strong>of</strong> ant i .<br />
The CAS algorithm is a kind <strong>of</strong> iterative optimization<br />
algorithm, which is firstly employed in the optimization<br />
<strong>of</strong> sequential space. In the sequential space coordinates,<br />
the mathematic description [17] <strong>of</strong> the CAS algorithm as<br />
follows:<br />
⎧<br />
( 1+<br />
ri<br />
)<br />
µ i ( n)<br />
= µ i ( n −1)<br />
⎪<br />
⎪<br />
7.<br />
5<br />
⎪<br />
θid<br />
( n)<br />
= ( θid<br />
( n −1)<br />
+ × Vi<br />
) ×<br />
⎪<br />
ψ id<br />
⎨<br />
7.<br />
5<br />
aµ<br />
i ( n)(<br />
3−ψ<br />
id ( θid<br />
( n−1)<br />
+ × Vi<br />
))<br />
ψ<br />
⎪ ( 1−e<br />
id 7.<br />
5<br />
⎪<br />
e<br />
− × Vi<br />
+<br />
ψ<br />
⎪<br />
id<br />
⎪ ( −2aµ<br />
i ( n)<br />
+ δ )<br />
⎩e<br />
( pid<br />
( n −1)<br />
−θid<br />
( n −1))<br />
(10)<br />
where i = 1, 2,<br />
L,<br />
N , N is the size <strong>of</strong> the ant swarm;<br />
d = 1, 2,<br />
L,<br />
L , L is the dimension <strong>of</strong> the optimization<br />
space; n means the current iteration, and n −1<br />
is the<br />
previous iteration; µ i is the current state <strong>of</strong> the ith ant’s<br />
organization variable, µ i(<br />
0)<br />
= 0.<br />
999 ; ri is termed by us<br />
as the organization factor <strong>of</strong> ant i ; ψ id determines the<br />
selection <strong>of</strong> the search range <strong>of</strong> the dth element <strong>of</strong><br />
variable in the search space; V i determines the search<br />
region <strong>of</strong> ant i and <strong>of</strong>fers the advantage that ants could<br />
© 2011 ACADEMY PUBLISHER<br />
search diverse regions <strong>of</strong> the problem space. The value <strong>of</strong><br />
Vi should be suitably selected according to concrete<br />
optimization problems; a is a sufficiently large positive<br />
constant and can be selected as a = 200 ;<br />
δ ( 0 ≤ δ ≤ 2 / 3)<br />
is a constant; pid ( n −1)<br />
is the best<br />
position found by the ith ant and its neighbors within<br />
n −1<br />
steps; θid is the current state <strong>of</strong> the dth dimension<br />
<strong>of</strong> ant i , θ id ( 0)<br />
= ( 7.<br />
5/<br />
ψ id )( 1−Vi<br />
) R , where R is a<br />
uniformly distributed random number in R ∈[<br />
0,<br />
1]<br />
.<br />
r i and ψ id are two important parameters. r i is the<br />
organization factor <strong>of</strong> ant i , which affects the<br />
convergence speed <strong>of</strong> the CAS algorithm directly. If r i is<br />
very large, the iteration step <strong>of</strong> ‘‘chaotic” search is small<br />
then the system converges quickly and the desired optima<br />
or near-optima cannot be achieved. If r i is very small, the<br />
iteration step <strong>of</strong> ‘‘chaotic” search is large then the system<br />
converges slowly and the runtime will be longer. Since<br />
small changes are desired as iteration step evolves, the<br />
value <strong>of</strong> r i is chosen typically as 5 . 0 0 ≤ < r i . The format<br />
<strong>of</strong> r i can be designed according to concrete problems and<br />
runtime. Each ant could have different r i , such<br />
as ri = 0. 3 + 0.<br />
02⋅<br />
rand(<br />
1)<br />
. ψ id affects the search ranges<br />
<strong>of</strong> the CAS algorithm. If the interval <strong>of</strong> the search is<br />
ωid<br />
ωid<br />
[ − , ] , then we can obtain an approximate formula<br />
2 2<br />
7.<br />
5<br />
ω = .<br />
id<br />
ψ id<br />
In principle, a neighborhood can be any ordered finite<br />
set. These neighbors are not necessarily individuals who<br />
are near them in the parameter space, but rather ones that<br />
are near them in a topological space. In fact the CAS<br />
algorithm does not impose any limitation on the<br />
definition <strong>of</strong> the distance between two ants. In order to<br />
simulate the behaviors <strong>of</strong> ants, we use the Euclidian<br />
distance. Supposing there are two ants whose positions<br />
are θ , L, θ ) and θ , L , θ ) , respectively, where<br />
( i1<br />
iL<br />
( j1<br />
jL<br />
− L ¢<br />
i , j = 1,<br />
2,<br />
L,<br />
N (where, N is the size <strong>of</strong> ant swarm)<br />
and i ≠ j , the distance between the two ants is<br />
2 ( θi1 2<br />
θ j1)<br />
+<br />
2<br />
+ ( θiL<br />
−θ<br />
jL)<br />
In the CAS algorithm, the neighbor selection can be<br />
defined in two ways. The first is the nearest fixed number<br />
<strong>of</strong> neighbors. The nearest m ants are selected as the<br />
neighbors <strong>of</strong> single ant. The second way is to consider the<br />
situation in which the number <strong>of</strong> neighbors increasing<br />
with iterative steps. This is due to the influence <strong>of</strong> selforganization<br />
behavior <strong>of</strong> ant i . The impact <strong>of</strong><br />
organization will become stronger than before and the<br />
neighbor <strong>of</strong> the ant will increase. That is to say, the<br />
number <strong>of</strong> nearest neighbor is dynamically changed as<br />
time evolves or iterative steps increase. The number q <strong>of</strong>
1938 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
single ant is defined to increase for every T iterative<br />
steps.<br />
B. Parameters Optimization <strong>of</strong> LS-SVM Based on CAS<br />
Algorithm<br />
As stated before, the CAS algorithm has powerful<br />
global search ability to find exact or approximate<br />
solutions for optimization and search problems. Thus, a<br />
parameters selection approach using the CAS algorithm<br />
for LS-SVM is presented in this paper. There are two key<br />
factors to determine the optimized parameters using the<br />
CAS algorithm: one is how to represent the parameters as<br />
the ant’s position, namely how to encode. Another is how<br />
to define the fitness function which evaluates the<br />
goodness <strong>of</strong> an ant. These two key factors are given as<br />
follows:<br />
Encoding parameters: the optimized parameters for<br />
LS-SVM include kernel parameter and regularization<br />
parameter. In solving parameters selection by the CAS<br />
algorithm, each ant is requested to represent a potential<br />
solution, namely parameters combination. So let us<br />
denote an m -parameters combination as a vector <strong>of</strong><br />
dimension m . For example, if Gauss radial basis function<br />
(RBF) is chosen as a kernel function, we denote the<br />
vector as v = ( γ , σ ) .<br />
Fitness function: the fitness function is generalization<br />
performance measure. There are some different<br />
descriptions for the generalization performance measure.<br />
Therefore, the corresponding fitness can be determined.<br />
The fitness <strong>of</strong> an ant is evaluated by the mean square<br />
error (MSE) index, which is defined as the error between<br />
the function estimation <strong>of</strong> LS-SVM and the reference<br />
model. It can be expressed by<br />
N<br />
1<br />
N ∑<br />
i=<br />
1<br />
( y − f ( x))<br />
where N denotes the number <strong>of</strong> training data, y is the<br />
reference model, and f (x)<br />
is the function estimation <strong>of</strong><br />
LS-SVM.<br />
In the CAS algorithm one aims at minimizing the MSE<br />
through choosing the optimal parameters combination,<br />
that is<br />
2<br />
f (z , L,<br />
z ) = minMSE (11)<br />
min 1 i<br />
subject to the equality constraints<br />
gi i i<br />
≤ z ≤ h , i = 1,<br />
2<br />
where the optimization variables are γ and<br />
σ respectively, [ i, i ] h g denotes the value range for each<br />
variable, which is different with different reference model<br />
and training data.<br />
The flowchart <strong>of</strong> the CAS-based parameters selection<br />
algorithm for the LS-SVM is shown in Fig. 1.<br />
© 2011 ACADEMY PUBLISHER<br />
IV. SIMULATION RESEARCH<br />
Experiment <strong>of</strong> a typical test function estimation is<br />
performed to evaluate the performance <strong>of</strong> the proposed<br />
parameters selection method. All experiments are<br />
performed on a PC with Pentium IV 2.93GHz<br />
processor, 512MB <strong>of</strong> main memory and the Matlab 6.5<br />
simulation s<strong>of</strong>tware.<br />
Given one-dimensional Sinc function<br />
f ( x)<br />
= sinc( x)<br />
+ v,<br />
x ∈[-3,3]<br />
(12)<br />
where v is the Gaussian noise with zero mean and<br />
standard deviation 0.1. We select 100 pairs <strong>of</strong> data as the<br />
train set from the input variable range. One aims at<br />
minimizing the MSE via the CAS algorithm to select the<br />
optimal kernel parameter σ <strong>of</strong> Gauss RBF kernel<br />
function and regularization itemγ .The searching ranges<br />
are set as follows: γ ∈[<br />
0,<br />
30]<br />
, σ ∈[<br />
0,<br />
5]<br />
. The CAS<br />
algorithm parameters are chosen as follows: N = 20 , the<br />
maximum number <strong>of</strong> iterations is 200,<br />
δ = 2/<br />
3 , a = 200 , = 0. 05 + 0.<br />
02×<br />
rand()<br />
, ψ 0.<br />
25<br />
,<br />
r i<br />
1 = i<br />
ψ 1.<br />
5 . In simulation, the first way is used to select<br />
2 = i<br />
neighbours <strong>of</strong> single ant. The researching results <strong>of</strong><br />
parameters are γ = 7.<br />
7379 and σ = 0.8851<br />
, respectively.<br />
The training result for LS-SVM via the above parameters<br />
is shown in Fig. 2. It can be seen from Fig.2 that LS-<br />
SVM realizes very good function approximation, so the<br />
CAS algorithm successfully realizes the parameters<br />
optimization selection for the test function.
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1939<br />
Figure 2. Simulation result <strong>of</strong> sin c function<br />
In order to explain the effectiveness <strong>of</strong> this method, we<br />
adopt the genetic algorithm (GA, crossover rate is 0.8,<br />
mutation rate is 0.2%, population size is 30, the<br />
maximum number <strong>of</strong> iterations is 200) and particles<br />
swarm optimization algorithm (PSO, the population size<br />
and maximum number <strong>of</strong> iterations is the same as GA) to<br />
carry out many times’ experiments. The model <strong>of</strong> LS-<br />
SVM is tested with the testing set about 50 data produced<br />
by randomly initialized, the average results is recorded in<br />
Table 1. Table 1 shows the model testing MSE <strong>of</strong> this<br />
paper method is the minimum.<br />
Selection method Testing error<br />
GA<br />
PSO<br />
CAS<br />
TABLE I.<br />
…... Simulation<br />
— Real<br />
····Training data<br />
AVERAGE RESULTS OF PARAMETERS OPTIMIZATION OF LS-SVM<br />
BY DIFFERENT METHODS<br />
7.9057×10 -4<br />
5.5782×10 -4<br />
3.1550×10 -4<br />
V. APPLICATION OF LS_SVM BASED ON THE CAS<br />
ALGORITHM<br />
The inverted pendulum artificially created is a complex<br />
nonlinear system in order to deeply research the control<br />
for the nonlinear, high order and under-actuated system.<br />
Characterized as a typical nonlinear, high order, unstable<br />
and under-actuated system, it is very difficult to give a<br />
precise mathematical model. Therefore, the model<br />
identification research for the inverted pendulum system<br />
is very important.<br />
The GPIP2003 single planar inverted pendulum is<br />
considered as a plant in the paper, whose inverse model is<br />
identified by LS-SVM. We adopt the example provided<br />
by the inverted pendulum toolbox, where the pendulum is<br />
displaced from lower position to the upper. After the<br />
pendulum reaches the upper position, one applies the<br />
disturbance by plucking the pendulum. The experiment<br />
data from the process overcoming the disturbance to the<br />
© 2011 ACADEMY PUBLISHER<br />
stabilization is sampled to the workspace <strong>of</strong> Matlab<br />
environment by the communication interface, which is<br />
stored as the text document by the command “save~”.<br />
The data includes seven items such as the sampling period,<br />
the control variable, angle <strong>of</strong> the pendulum, position <strong>of</strong><br />
the cart, angular rate <strong>of</strong> the pendulum, velocity <strong>of</strong> the cart<br />
and displacement <strong>of</strong> the objective. Angle <strong>of</strong> the pendulum,<br />
angular rate <strong>of</strong> the pendulum, position <strong>of</strong> the cart, velocity<br />
<strong>of</strong> the cart and the control variable are selected as multiinput<br />
and single-output model for LS-SVM. 100 pairs<br />
data from the input variable are chosen as the training<br />
sample set, in which 40 pairs data are selected as the<br />
testing sample set. The minimum MSE error as the fitness<br />
function, we utilize the CAS algorithm to carry the<br />
optimization selection for the regularization item γ and<br />
kernel parameter σ . The researching results <strong>of</strong><br />
parameters are γ = 7.<br />
7379 and σ = 0.8851<br />
, and the<br />
testing error is 0.0022. The estimation for the inverse<br />
model is achieved using the above result. The simulation<br />
result is shown in Fig. 3. Fig. 3 shows that the estimation<br />
value approaches to the real sampling value. simulation<br />
results show the LS-SVM model has good generalization<br />
performance and stronger robust performance after<br />
optimized by the CAS algorithm.<br />
VI. CONCLUSION<br />
…... Simulation<br />
— Real<br />
Figure 3. Simulation result <strong>of</strong> inverse modeling<br />
for the inverted pendulum<br />
Appropriate parameters are very crucial to leastsquares<br />
support vector machines (LS-SVM) learning<br />
results and generalization ability. This paper presents a<br />
novel parameter selection method for LS-SVM is<br />
presented based on chaotic ant swarm (CAS) algorithm.<br />
The selection problem <strong>of</strong> LS-SVM parameters is<br />
considered as a swarm intelligence optimization problem,<br />
and a CAS optimization algorithm is employed to search<br />
the optimal objective function. CAS algorithm is global<br />
search method and it need not to consider LS-SVM<br />
dimensionality and complexity. Simulation and<br />
experiment results show that the proposed method is an<br />
effective approach for parameter optimization.
1940 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
ACKNOWLEDGMENT<br />
The authors are grateful to the anonymous referees for<br />
their valuable remarks and helpful suggestions, which<br />
have significantly improved the paper. This work was<br />
supported in part by a grant from the support <strong>of</strong> Key<br />
Project <strong>of</strong> Chinese National Programs for Fundamental<br />
Research and Development (973 Program) (2007CB7140<br />
06), National Nature Science Foundation <strong>of</strong> China<br />
(61074020) and the Fundamental Research Funds for the<br />
Central Universities (DC10040101).<br />
REFERENCES<br />
[1] V. N. Vapnik, The Nature <strong>of</strong> Statistical Learning Theory,<br />
chapter 6, New York: Springer-Verlag, 1995.<br />
[2] V. N. Vapnik, “An Overview <strong>of</strong> Statistical Learning<br />
Theory,” IEEE Trans on Neural Networks, vol. 10, p. 988-<br />
999, 1999.<br />
[3] J. A. K.Suykens, “Support vector machines: a nonlinear<br />
modeling and control perspective,” European <strong>Journal</strong> <strong>of</strong><br />
Control, vol. 7, p. 311-327, 2001.<br />
[4] J. A. K.Suykens, “Nonlinear Modeling and Support Vector<br />
Machines (Published Conference Proceedings style),”<br />
Proc. <strong>of</strong> 18th Annu. IEEE Conference on Instrumentation<br />
and Measurement Technology, Budapest, p. 287-294,<br />
2001.<br />
[5] K. S. Chua, “Efficient computations for large least square<br />
support vector machine classifiers,” Pattern Recognition<br />
Letters, vol. 24, p. 75-80, 2003.<br />
[6] I. Goethals, K. Pelckmans, J. A. K. Suykens and B. De<br />
Moor, “Identification <strong>of</strong> MIMO Hammerstein models<br />
using least squares support vector machines,” Automatica,<br />
vol. 41, p. 1263-1272, 2005.<br />
[7] L. Bako, G. Mercere, S. Lecoeuche and M. Lovera,<br />
“Recursive subspace identification <strong>of</strong> Hammerstein models<br />
based on least squares support vector machines,” IET<br />
Control Theory and Application, vol. 3, p. 1209-1216,<br />
2009.<br />
[8] L. K. Hou, Q. X. Yang and J. L. An, “Modeling <strong>of</strong> SRM<br />
Based on XS-LSSVR Optimized by GDS,” IEEE<br />
Transactions on Applied Superconductivity, vol. 20, p.<br />
1102-1105, 2010.<br />
[9] Z. J. Li, Y. N. Zhang and Y. P. Yang, “Support vector<br />
machine optimal control for mobile wheeled inverted<br />
pendulums with unmodelled dynamics,” Neurocomputing,<br />
vol. 73, p. 2773-2782, 2010.<br />
[10] N. E. Ayat, M. Cheriet and C. Y. Suen, “Automatic model<br />
selection for the optimization <strong>of</strong> SVM kernels,” Pattern<br />
Recognition, vol. 38, p. 1733-1745, 2005.<br />
[11] Y. W. Kang, J. Li, G. Y. Cao, H. Y. Tu, J. Li and J. Yang,<br />
“Dynamic temperature modeling <strong>of</strong> an SOFC using least<br />
squares support vector machines,” <strong>Journal</strong> <strong>of</strong> Power<br />
sources, vol. 179, p. 683-692, 2008.<br />
[12] P. F. Pai and W. C. Hong, “Support vector machines with<br />
simulated annealing algorithms in electricity load<br />
forecasting,” Energy Conversion andManagement, vol. 46,<br />
p. 2669-2688, 2005.<br />
[13] X. L. Tang, L. Zhang, J. Cai and C. B. Li, “Multi-fault<br />
classification based on support vector machine trained by<br />
chaos particle swarm optimization,” Knowledge-Based<br />
Systems, vol. 23, p. 486-490, 2010.<br />
[14] S. J. An, W. Q. Liu and S. Venkatesh, “Fast crossvalidation<br />
algorithms for least squares support vector<br />
machines and kernel ridge regression,” Pattern<br />
Recognition, vol. 40, p. 2154-2162, 2007.<br />
© 2011 ACADEMY PUBLISHER<br />
[15] W. T. Mao, G. R. Yan, L. L. Dong and D. Hu, “Model<br />
selection for least squares support vector regression based<br />
on small-world strategy,” Expert Systems with Applications,<br />
vol. 38, p. 3227-3237, 2011.<br />
[16] T. V. Gestel, J. A. K. Suykens and B. Baesens, S. Viaene,<br />
J. Vanthienen and G. Dedene, et al., “Benchmarking least<br />
squares support vector machine classifiers,” Machine<br />
Learning, vol. 54, p. 5-32, 2004.<br />
[17] L. X. Li, Y. X. Yang, H. P. Peng and X. D. Wang,<br />
“Parameters identification <strong>of</strong> chaotic systems via chaotic<br />
ant swarm,” Chaos, Solitons and Fractals, vol. 28, p.<br />
1204–1211, 2006.<br />
[18] L. X. Li, Y. X. Yang, H. P. Peng and X. D. Wang, “An<br />
optimization method inspired by chaotic ant havior,”<br />
International <strong>Journal</strong> <strong>of</strong> Bifurcation Chaos, vol. 16, p.<br />
2351-2364, 2006.<br />
[19] J. J. Cai, X. Q. Ma, L. X. Li, Y. X. Yang, H. P. Peng and X.<br />
D. Wang, “Chaotic ant swarm optimization to economic<br />
dispatch,” Electric Power Systems Research, vol. 77, p.<br />
1373-1380, 2007.<br />
[20] L. X. Li, Y. X. Yang and, H. P. Peng, “Computation <strong>of</strong><br />
multiple global optima through chaotic ant swarm,” Chaos,<br />
Solitons and Fractals, Vol. 40 (2009), p. 1399-1407.<br />
[21] Y. G. Tang, M. Y. Cui, Li L. X., H. P. Peng and X. P.<br />
Guan, “Parameter identification <strong>of</strong> time-delay chaotic<br />
system using chaotic ant swarm,” Chaos, Solitons and<br />
Fractals, vol. 41, p. 2097-2102, 2009.<br />
[22] L. X. Li, Y. X. Yang and, H. P. Peng, “Fuzzy system<br />
identification through chaotic ant swarm,” Chaos, Solitons<br />
and Fractals, vol. 41, p. 401-408, 2009.<br />
[23] H. Zhu, L. X. Li, Y. Zhao, Y. Guo and Y. X. Yang, “CAS<br />
algorithm-based optimum design <strong>of</strong> PID controller in<br />
AVR system,” Chaos, Solitons and Fractals, vol. 42, p.<br />
792-800, 2009.<br />
[24] Y. Y. Li, Q. Y. Wen, L. X. Li and H. P. Peng, “Hybrid<br />
chaotic ant swarm optimization,” Chaos, Solitons and<br />
Fractals, vol. 42, p. 880-889, 2009.<br />
[25] W. C. Hong, “Application <strong>of</strong> chaotic ant swarm<br />
optimization in electric load forecasting,” Energy Policy,<br />
vol. 38, p. 5830-5839, 2010.<br />
[26] A. Chatterjee, S. P. Ghoshal and V. Mukherjee, “Chaotic<br />
ant swarm optimization for fuzzy-based tuning <strong>of</strong> power<br />
system stabilizer”, Electrical Power and Energy Systems,<br />
in press.<br />
[27] B. J.Cole, “Is animal behavior chaotic? Evidence from the<br />
activity <strong>of</strong> ants.” Proc R Soc Lond Ser B biol Sco, vol. 244,<br />
p. 253-259, 1991.<br />
[28] R. V. Solé, O. Miramontes and B. C. Goodwill,<br />
“Oscillations and chaos in ant societies,” <strong>Journal</strong> <strong>of</strong> Theory<br />
Biology, vol. 161, p. 343-357, 1993.<br />
Chunli Xie Xie received his B.Sc. and M.Sc. degrees from<br />
Fushun Petroleum Institute and Liaoning Shihua University,<br />
Fushun, China, in 1995 and 2003, respectively. He is currently<br />
working toward the Ph.D degree with Dalian University <strong>of</strong><br />
Technology, Dalian, China.<br />
His research interests include adaptive control, robust control,<br />
machine learning, nonlinear systems, artificial intelligence and<br />
application.<br />
Cheng Shao was born in Shenyang, P. R. China, on June 7,<br />
1958. Shao received his B.Sc. degree from Liaoning University,<br />
Shenyang, China, in 1982. Then he received the M.Sc and Ph.D.<br />
degrees from Northeastern University, Shenyang, China, in<br />
1987 and 1992.<br />
He is currently a full-time Pr<strong>of</strong>essor and Ph.D. Advisor with<br />
School <strong>of</strong> Electronic and Information Engineering, Dalian
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1941<br />
University <strong>of</strong> Technology, China. His research interest covers<br />
complex system modeling and intelligence control.<br />
Dandan Zhao was born in Fuxin, P. R. China, on March 4,<br />
1975. Zhao received her B.Sc. and M.Sc. degrees from Fushun<br />
Petroleum Institute and Liaoning Shihua University, Fushun,<br />
© 2011 ACADEMY PUBLISHER<br />
China, in 1997 and 2003, respectively. She is currently a fulltime<br />
lecturer <strong>of</strong> School <strong>of</strong> Computer Science and Engineering,<br />
Dalian Nationalities University, Dalian, China. Her research<br />
interests include electronic commerce, semantic network,<br />
swarm intelligent and information processing.
1942 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
The Expected Value Model <strong>of</strong> Multiobjective<br />
Programming and its Solution Method Based on<br />
Bifuzzy Environment<br />
Mingfa Zheng<br />
College <strong>of</strong> Science, Air Force Engineering University, Xi'an, Shanxi, 710051, China<br />
Email: mingfazheng@126.com<br />
Bingjie Li and Guangxing Kou<br />
College <strong>of</strong> Science, Air Force Engineering University, Xi'an, Shanxi, 710051, China<br />
Email: {mingfa103, kouguangx }@ 163.com<br />
Abstract—In this paper, based on bifuzzy theory, we study<br />
the multiobjective programming problem under bifuzzy<br />
environment and present the expected value model to the<br />
problem. Furthermore, to the proposed model, the concepts<br />
<strong>of</strong> non-inferior solution are defined, and their relations are<br />
also discussed. According to practical decision-making<br />
process, a solution method, called the method <strong>of</strong> main<br />
objective function, has been studied, whose results can<br />
facilitate us to design algorithms to solve the bifuzzy<br />
multiobjective programming problem. Finally, a numerical<br />
example is given to explain the proposed method.<br />
Index Terms—credibility theory, bifuzzy variable,<br />
multiobjective programming, expected value model<br />
I. INTRODUCTION<br />
The multiobjective programming problems are studied<br />
by many researchers such as [3], [15], [16],[ [18]],[ [20].<br />
For given multiobjective problem, its absolute optimal<br />
solutions which optimize each objective functions<br />
simultaneously usually don not exist, so we consider their<br />
non-inferior solutions, which are Pareto optimal solutions<br />
in common use. There are various types <strong>of</strong> uncertainties<br />
in the real-world problem. As we known, random<br />
phenomena is one class <strong>of</strong> uncertain phenomena which<br />
has been well studied. Based on the probability,<br />
stochastic multiobjective programming problems have<br />
been presented such as [1], [17]. Besides randomness,<br />
fuzziness is a basic type <strong>of</strong> subjective uncertainty<br />
initiated by [26]. Since the pioneering work <strong>of</strong> Zadeh,<br />
possibility theory was developed and extended by many<br />
researchers such as [2],[4],[7],[23],[21],[24]. Based on<br />
possibility theory, an axiomatic approach, called<br />
credibility theory [6], was studied extensively. From a<br />
measure-theoretic viewpoint, credibility theory provides a<br />
theoretical foundation for fuzzy programming [9] just like<br />
the role <strong>of</strong> probability theory in stochastic programming<br />
[5]. In a practical decision-making process, we <strong>of</strong>ten face<br />
a hybrid uncertain environment where linguistic and<br />
frequent nature coexist. For the examples <strong>of</strong> two fold<br />
© 2011 ACADEMY PUBLISHER<br />
doi:10.4304/jcp.6.9.1942-1948<br />
uncertainty, we may refer to Liu [6], Liu [8], [10],<br />
Liu[11], Liu and Liu [13], Yazenin[22], Zhou[25]. To<br />
deal with this two fold uncertainty, it is required to<br />
employ bifuzzy theory[7]. The multiobjective<br />
programming under bifuzzy environment has not been<br />
developed well, therefore, following the idea <strong>of</strong> stochastic<br />
multiobjective programming, this paper devotes the<br />
bifuzzy multiobjective programming (BMOP) problems<br />
based on the random fuzzy theory. For the parameters <strong>of</strong><br />
bifuzzy, we consider their expectation which convert the<br />
BMOP problem into the expected value model <strong>of</strong> bifuzzy<br />
multiobjective (EVBMOP) which is a deterministic<br />
multiobjective problem. By the deterministic problem<br />
above, we can obtain the expected value efficient<br />
solutions or expected value weakly efficient solutions to<br />
the BMOP problem. In actual problem, we usually need<br />
to distinguish between primary and secondary <strong>of</strong> the<br />
objective functions to the BMOP problem, so the method<br />
<strong>of</strong> main objective function is presented to solve the<br />
BMOP problem in this paper, which can covert the<br />
EVBMOP problems corresponding to the BMOP<br />
problem into the deterministic single objective<br />
programming problems whose optimal solutions are<br />
expected value weakly efficient solutions to the BMOP<br />
problems.<br />
This paper is organized as follows. The next section<br />
provides a brief review on the related concepts and results<br />
in bifuzzy theory. Section 3 presents the BMOP problem<br />
and its expected value model. Furthermore, based on the<br />
expected value model, the expected value efficient<br />
solution and expected value weakly efficient solution to<br />
the BMOP are proposed, and their properties are<br />
discussed. To solve the BMOP problem, the method <strong>of</strong><br />
main objective function is introduced in Section 4.<br />
Finally, Section 5 provides a summary <strong>of</strong> the main results<br />
<strong>of</strong> this paper.
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1943<br />
II. BASIC CONCEPTS<br />
Given a universe Γ , ρ( Γ)<br />
is the power set <strong>of</strong> Γ , and<br />
a set function Pos defined on ρ( Γ)<br />
is called a possibility<br />
measure if it satisfies the following conditions[4]:<br />
(Pos1) P os( φ ) = 0, P os(<br />
Γ)<br />
= 1, and<br />
(Pos2) P os( ∪ i∈I Ai) = supi∈I P os( Ai)<br />
for any subclass<br />
{ Aii∈I}<strong>of</strong> ρ( Γ).<br />
The triplet ( Γ, ρ(<br />
Γ),<br />
Cr)<br />
is usually called a possibility<br />
space, which is called a pattern space by Nahimias [19].<br />
In addition, a self-dual set function, called credibility<br />
measure, is defined as follows [12]:<br />
1<br />
c<br />
C r( A) = (1 + P os( A) − P os( A )).<br />
2<br />
for any A ∈ ρ(<br />
Γ),<br />
where A c is the complement <strong>of</strong> A .<br />
A fuzzy variable X is defined as a function from a<br />
credibility space ( Γ, ρ(<br />
Γ),<br />
Cr)<br />
to the set <strong>of</strong> real numbers.<br />
Based on credibility measure, the expected value <strong>of</strong><br />
fuzzy variable X is defined as [12]<br />
∞<br />
∫ ∫−∞ 0<br />
0<br />
E[ ξ] = Cr( ξ ≥ r) dr − Cr( ξ ≤ r) dr (1)<br />
provided that one <strong>of</strong> the two integrals is finite.<br />
Given a credibility space ( Γ, ρ(<br />
Γ),<br />
Cr)<br />
, which is<br />
complete, we obtain the definition <strong>of</strong> bifuzzy variable as<br />
follows:<br />
Definitions 2.1.[7] Let ( Γ, ρ(<br />
Γ),<br />
Cr)<br />
be a credibility<br />
space. A map ( 1, 2,<br />
,<br />
) : v is said to be<br />
an bifuzzy vector if for any Borel subset B <strong>of</strong><br />
the function<br />
T<br />
n<br />
ξ = ξ ξ ξn<br />
T → F<br />
n−ary n<br />
'<br />
R , C{ r γ ∈Γ ξγ( γ ) ∈B}<br />
is measurable<br />
with respect to γ . As n = 1, ξ is called a bifuzzy<br />
variable.<br />
Definitions 2.2.[7] Suppose ξ is a bifuzzy variable,<br />
the expected value <strong>of</strong> ξ is defined as the mathematical<br />
expectation <strong>of</strong> the fuzzy variable E[ ξγ<br />
], i.e.,<br />
E( ξ) = ∫ E[ ξγ] Crd(<br />
γ)<br />
(2)<br />
Γ<br />
provided that the integrand E[ ξ γ ] defined by Eq.(1)<br />
exists almost surely with respected to γ , and is integral.<br />
From Eq.(2), we can provide the expectation <strong>of</strong><br />
bifuzzy variable, i.e.,<br />
E( ξ ) = Eγ[ Eγ '[<br />
ξγ( γ ')]].<br />
Ⅲ BIFUZZY MULTIOBJECTIVE PROGRAMMING PROBLEMS<br />
A. Expected Value Model <strong>of</strong> Bifuzzy Multiobjective<br />
Programming<br />
Considering the bifuzzy multiobjective programming<br />
(BMOP) problem as follows:<br />
⎧min<br />
F( x, ξ) = ( f1( x, ξ), f2( x, ξ), ⋅⋅⋅,<br />
fj( x,<br />
ξ))<br />
x∈R ⎪<br />
(BMOP) ⎨ st . . G( x, ξ) = ( g1( x, ξ), g2( x, ξ), ⋅⋅⋅ , gn( x,<br />
ξ))<br />
1944 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
Similarly, by the linear properties <strong>of</strong> fuzzy variable, we<br />
can obtain:<br />
G( λx + (1 −λ)<br />
x , ξ )<br />
1 2<br />
≤ λGx ( , ξ ) + (1 −λ)<br />
Gx ( , ξ )<br />
1 γ 2 γ<br />
Using the same method above, we can obtain:<br />
EG [ ( λx+ (1 −λ)<br />
x , ξ)]<br />
1 2<br />
≤ λEG [ ( x1, ξ)] + (1 − λ) EGx [ ( 2,<br />
ξ)]<br />
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1945<br />
F( x, ξ ) is strict convex vector function on D , and is<br />
also conmonotonic, by noting the inequality just given, it<br />
easy to know that<br />
*<br />
EF [ ( αx+ (1 −α)<br />
x,<br />
ξ)]<br />
*<br />
< α EF [ ( x, ξ)] + (1 −α)<br />
EF [ ( x,<br />
ξ)]<br />
*<br />
< EF [ ( x,<br />
ξ )],<br />
*<br />
which is a contradiction with x ∈ D . Thus,<br />
Dpa ⊃ Dwpa<br />
,<br />
wpa<br />
which proves the required theorem.<br />
Ⅳ SOLUTION METHOD<br />
A. Expected Value Model <strong>of</strong> Bifuzzy Multiobjective<br />
Programming<br />
In real-world problems, we just need to consider the<br />
main objective function to some real-life problems,<br />
therefore, a type <strong>of</strong> method, called the method <strong>of</strong> main<br />
objective function, is presented in the following. Without<br />
any loss <strong>of</strong> generality, let f1( x, ξ ) be regarded as main<br />
objective function to the BMOP problem, and wish the<br />
expectation <strong>of</strong> the other objective functions fj( x, ξ ) ,<br />
j=2,3,…,p, satisfy the following constraint-conditions:<br />
E[ fj( x, ξ )] ≤ αi<br />
, j=2,3,…,p. Then the BMOP problem<br />
can be transformed into the following SOP problem<br />
where<br />
min E[ f1( x,<br />
ξ )] (8)<br />
x∈ D<br />
D<br />
= { x∈ D E[ f ( x, ξ)] ≤ α , j = 2, 3,..., p},<br />
j i<br />
whose optimal solution set is denoted as D sab .<br />
Obviously, the constraint set D to problem (8) is a<br />
new set which is added into several constraint condition<br />
E[ fj( x, ξ )] ≤ αi,<br />
j = 2, 3, ⋅⋅⋅,<br />
p.<br />
Then we employ the<br />
method <strong>of</strong> solving the nonlinear programming which is a<br />
linear problem in particular to solve the transformed SOP<br />
problem whose optimal solution is the non-inferior<br />
solution to the BMOP problem verified by the following<br />
theorem.<br />
D ⊂ D<br />
Theorem 4.1. sab wpa<br />
* *<br />
Pro<strong>of</strong>. If x ∈ Dsab<br />
, and x ∉ Dwpa<br />
, then, by the<br />
definition <strong>of</strong> expected value weakly efficient solution,<br />
there<br />
must exist some x∈ D,<br />
such that<br />
*<br />
E[ fj( x, ξ )] < E[ fj( x , ξ )] for all j=1,2,...,p.<br />
*<br />
Since x ∈ D,<br />
we have<br />
*<br />
E[ fj( x , ξ)] ≤ α j,<br />
j = 2,3, ⋅⋅⋅,<br />
p.<br />
It follows from inequality above that<br />
*<br />
E[ f j( x, ξ)] ≤ E[ f j( x , ξ)] ≤ α j,<br />
j = 2,3, ⋅⋅⋅,<br />
p,<br />
which illuminates x∈ D,<br />
i.e., x is the feasible solution<br />
to SOP problem, therefore, it is easy to know<br />
*<br />
E[ f ( x, ξ )] < E[ f ( x , ξ )] ,<br />
j j<br />
© 2011 ACADEMY PUBLISHER<br />
which is a contraction with by the previous hypothesis<br />
* *<br />
that x ∈ Dsab<br />
. Hence, x ∈ Dwpa<br />
, which implies the<br />
required conclusion.<br />
Theorem 4.2. Without any loss <strong>of</strong> generality, assuming<br />
that f1( x, ξ ) is the main objective function, if H( x, ξ )<br />
is linear vector function, F( x, ξ ) and G( x, ξ ) are strict<br />
convex vector function on x . Furthermore, for any<br />
given x1 and x2 , F( x1, t)<br />
and F( x2, t)<br />
(correspondingly, G( x 1,<br />
t)<br />
and G( x2, t)<br />
are comonotonic<br />
on t , then Dsab ⊂ Dpa<br />
. In addition, if Dab ≠ φ,<br />
we can<br />
obtain:<br />
Dsab ⊂ Dab<br />
.<br />
Pro<strong>of</strong>. It follows from the assumed conditions that<br />
E[ f1( x, ξ )] is strict convex function, so the optimal<br />
*<br />
solution to SOP problem must be unique. If x is the<br />
unique optimal solution to the SOP problem, and<br />
*<br />
x ∉ D , there must exist x∈ D and x ≠ x<br />
* such that<br />
pa<br />
E[F( x, ξ )] ≤ E[F( x , ξ )] ,<br />
i.e.,<br />
*<br />
E[ fj( x, ξ)] ≤ E[ fj( x , ξ)],<br />
j = 1,2,3, ⋅⋅⋅,<br />
p.<br />
Obviously, there must exist some j0(1 =< j0<br />
1946 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
the α i according to the actual demand. Furthermore, if<br />
the α i is not well-found, then the feasible sets D may<br />
be empty set, which can't get the optimal solution <strong>of</strong> SOP<br />
problem, that is, we can’t obtain the expected value noninferior<br />
solutions to the BMOP problem, so we can take<br />
the following measure which can avoid that D is empty<br />
set:<br />
0 α j=E[ fj( x , ξ )], j = 2,3, ⋅⋅⋅,<br />
p,<br />
0<br />
for any given x ∈ D,<br />
which can guarantee that one<br />
0<br />
solution at least, i.e., there exist x ∈ Dat<br />
least.<br />
Furthermore, the optimal solutions <strong>of</strong> the SOP problem<br />
by the measure proposed above must be the expected<br />
value weakly efficient solution <strong>of</strong> the BMOP problem,<br />
and it may be not satisfying, but it is the practical<br />
technique to deal with real-life problem frequently.<br />
B. Expected Value Model <strong>of</strong> Bifuzzy Multiobjective<br />
Programming<br />
In particular, if the bifuzzy variable ξ involved in the<br />
problem (8) is a discrete one, we will illuminate how to<br />
calculate the E[ f1( x, ξ )]. Assume that the bifuzzy<br />
variable ξ is a discrete one such that γ is a discrete<br />
fuzzy variable taking on finite number <strong>of</strong> values with<br />
possibility μ i , i = 1, 3, ⋅⋅⋅,<br />
N , respectively, and<br />
N<br />
satisfying maxi= 1 μi<br />
= 1, j = 1, 3, ⋅⋅⋅,<br />
N , and for each i ,<br />
fuzzy variable ξ taking on the following values<br />
γ<br />
γ i<br />
ξ ( γ ) with possibility μ > 0;<br />
1<br />
'<br />
11<br />
'<br />
12<br />
ξγ ( γ 1<br />
……<br />
) with possibility μ 12 > 0;<br />
' ξγ( γ 1 1N<br />
) with possibility μ 1<br />
1N1<br />
> 0;<br />
ξ '<br />
( γ ) with possibility μ > 0;<br />
γ<br />
2<br />
21<br />
'<br />
22<br />
ξγ ( γ 2<br />
……<br />
) with possibility μ 22 > 0;<br />
' ξγ( γ 2 2N<br />
) 2<br />
with possibility μ 2N<br />
2<br />
……<br />
> 0;<br />
ξγ γ i i<br />
' ( 1 ) with possibility μ i1<br />
> 0;<br />
ξγ γ i i<br />
' ( 2 ) with possibility μ i2<br />
> 0;<br />
……<br />
ξγγ i iN i<br />
'<br />
( ) with possibility μ iNi<br />
……<br />
> 0;<br />
N N ' ξγ( γ 1)<br />
with possibility μ N1<br />
> 0;<br />
ξγγ N N<br />
' ( 2 ) with possibility μ N 2 > 0;<br />
……<br />
ξγγ NN<br />
'<br />
( ) with possibility μ iN N > 0;<br />
N N<br />
It is easy to obtain the expectation <strong>of</strong> fuzzy variable<br />
f1( x, ξγ( γ ')) as follows:<br />
© 2011 ACADEMY PUBLISHER<br />
11<br />
21<br />
Ni<br />
'<br />
1 ξγ = γ ' 1 ξγ γ =∑ij<br />
1 ξγ γ i ij<br />
j=<br />
1<br />
f ( x, ) E [ f ( x, ( '))] p f ( x,<br />
( )) (9)<br />
where pij are the weights <strong>of</strong> fuzzy<br />
variable f x ξγγ i ij calculated by the following<br />
formulas [14]:<br />
'<br />
1( , ( ))<br />
j j−1 Ni Ni+<br />
1<br />
1 1<br />
pij = (max μik − max μik ) + (max μik − max μik<br />
)<br />
2 k= 1 k= 0 2 k= j k= j+<br />
1<br />
(10)<br />
where μi0= μ iNi+ 1 = 0, i = 1, 2, ⋅⋅⋅ , N, j = 1, 2, ⋅⋅⋅,<br />
Ni,<br />
and satisfies the following constrains:<br />
p ≥ 0,<br />
Ni + 1<br />
p<br />
Ni<br />
+ 1 = max μ = 1.<br />
∑<br />
ij j ij j ij<br />
By the Eq.(2), the expectation <strong>of</strong> bifuzzy variable<br />
f1( x, ξγ<br />
) are given in the following<br />
E[ f ( x, ξ)] = E [ f ( x, ξ )] =∑p<br />
f ( x,<br />
ξ ) (11)<br />
1 γ 1 γ i 1 γ<br />
i=<br />
1<br />
where pi<br />
are the weights <strong>of</strong> fuzzy variable<br />
f1( x, ξγ<br />
) calculated similarly by the Eq.(10).<br />
Example 4.1. Solving the following bifuzzy<br />
multiobjective programming<br />
⎧ min( f1( x, ξ), f2( x,<br />
ξ))<br />
x<br />
⎪<br />
⎪ = (5x1 + 7x2 − 4x3 + 2 ξ, − 2x1 + 3x2 + 8x3 −3<br />
ξ)<br />
⎪<br />
⎨ st .. x1 + 2x2 −3x3 ≤ 5<br />
⎪<br />
−7x − x + x ≤<br />
⎪ 1 3 2 7 3 3<br />
⎪ 11x1 + 5x2 −6x3 ≤ 10<br />
⎩<br />
(12)<br />
where f1( x, ξ ) is the main objective function, and the<br />
limit value <strong>of</strong> E[ f2( x, ξ )] is 4.4, i.e.,<br />
E[ f2( x, ξ )] ≤ α2<br />
= 4.4.<br />
In addition, ξ is the discrete bifuzzy variable defined<br />
as<br />
⎧ X , with possibility 3/5<br />
ξγ<br />
1<br />
⎪<br />
= ⎨ X2 , with possibility 1/4<br />
⎪⎩ X , with possibility 1 .<br />
3<br />
Here the fuzzy variable X 1 assumes the value 3, 4,<br />
5with the possibility 1/4, 3/4, and 1, respectively; X 2<br />
assumes the value 1, 2, 3 with the possibility 5/12, 1 and<br />
7/12, respectively; and X 3 assumes the value 6, 8, 10<br />
with the possibility 2/7, 1/7 and 1, respectively.<br />
According to the method <strong>of</strong> main objective function<br />
discussed above, the problem (12) can be transformed<br />
into the following single objective problem<br />
⎧ min E[( f1( x, ξ)] = E[5x1 + 7x2 − 4x3 + 2 ξ]<br />
x<br />
⎪<br />
⎪ st .. E[ − 2x1 + 3x2 + 8x3 −3 ξ ] ≤ 4.4<br />
⎪<br />
⎨ x1 + 2x2 −3x3 ≤ 5 (13)<br />
⎪<br />
−7x − x + x ≤<br />
⎪ 1 3 2 7 3 3<br />
⎪ 11x1 + 5x2 −6x3 ≤ 10.<br />
⎩<br />
We can obtain the following results<br />
N
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1947<br />
'<br />
f1( x,<br />
ξγ( γ 1 11))]<br />
'<br />
= 5x1+ 7x2− 4x3+ 6, f1( x,<br />
ξγ( γ 1 12))]<br />
'<br />
= 5x1+ 7x2− 4x3+ 8, f1( x,<br />
ξγ( γ 1 13))]<br />
= 5x1+ 7x2− 4x3+ 10.<br />
It is easy to know<br />
' '<br />
'<br />
f1( x, ξγ ( γ f x f x<br />
1 11))] ≤ 1( , ξγ ( γ 1 12))] ≤ 1( , ξγ ( γ 1 13))].<br />
Therefore, we can obtain the distribution function <strong>of</strong><br />
' f1( x, ξγ( γ j<br />
fuzzy variable<br />
1 1 ))<br />
as<br />
⎧5x1+<br />
7x2− 4x3+ 6, with possibility 1/ 4<br />
' ⎪<br />
f1( x, ξγ( γ j = ⎨ x + x − x + with possibility<br />
1 1 )) 5 1 7 2 4 3 8, 3/4<br />
⎪<br />
⎩5x1+<br />
7x2− 4x3+ 10, with possibility 1<br />
with weights p11 = 1/ 8, p 12 = 1/ 4, and p13=<br />
5 / 8,<br />
respectively, which are calculated by Eq.(10).<br />
It follows the Eq.(9) that<br />
1 ξγ1 3<br />
= ∑<br />
j=<br />
1<br />
ij 1 ξγ1 ' γ1j<br />
f ( x, ) p f ( x,<br />
( ))<br />
1 1<br />
= (5x1+ 7x2− 4x3+ 6) + (5x1+ 7x2− 4x3+ 8)<br />
8 4<br />
5<br />
+ (5x1+ 7x2− 4x3+ 10)<br />
8<br />
= 5x1+ 7x2− 4x3+ 9,<br />
whose possibility is 3/5.<br />
Similarly, we can obtain<br />
f1( x, ξγ<br />
) = 5x + x − x +<br />
2 1 7 2 4 3 25/6,<br />
f1( x, ξγ<br />
) = 5x + x − x +<br />
3 1 7 2 4 3 132/7,<br />
with the possibility 1/4 and 1, respectively.<br />
Obviously,<br />
' '<br />
'<br />
f1( x, ξγ ( γ f x f x<br />
1 12))] ≤ 1( , ξγ ( γ 1 11))] ≤ 1( , ξγ ( γ 1 13))].<br />
Hence, without any loss <strong>of</strong> generality, the distribution<br />
function <strong>of</strong> fuzzy variable f1( x, ξγ<br />
) is the following<br />
i<br />
⎧ 5x1+ 7x2− 4x3+ 25/6, with possibility 1/4<br />
' ⎪<br />
f1( x, ξγ( γ j = ⎨ x + x − x + with possibility<br />
1 1 )) 5 1 7 2 4 3 9, 3/5<br />
⎪<br />
⎩5x<br />
1+ 7x2− 4x3+ 132/ 7, with possibility 1<br />
with weights<br />
p11 = 1/ 8, p12= 7 / 40, and p13<br />
= 7 /10,<br />
respectively, which are calculated by Eq.(10).<br />
By the Eq.(11), we can deduce<br />
ξ = γ ξγ = ∑ i ξγi<br />
i=<br />
3<br />
1 1 1<br />
1<br />
E[ f ( x, )] E [ f ( x, )] p f ( x,<br />
)<br />
1 7<br />
= (5x1+ 7x2− 4x3+ 25/ 6) + (5x1+ 7x2− 4x3+ 9)<br />
8 40<br />
7<br />
+ (5x1+ 7x2− 4x3+ 132 / 7)<br />
10<br />
= 5x1+ 7x2− 4x3 + 20.953.<br />
Using the same method, we can obtain the expectation<br />
<strong>of</strong> the bifuzzy variable f2( x, ξ ) as follows<br />
1 ξ = γ 2 ξγ 3<br />
= ∑<br />
i=<br />
1<br />
i 2 ξγi<br />
E[ f ( x, )] E [ f ( x, )] p f ( x,<br />
)<br />
=− 2x + 3x + 8x −23.25.<br />
1 2 3<br />
Therefore, problem (13) is equivalent to the following<br />
problem:<br />
© 2011 ACADEMY PUBLISHER<br />
⎧min<br />
5x1+ 7x2− 4x3 + 20.953<br />
x<br />
⎪<br />
⎪ st . . − 2x1 + 3x2 + 8x3 −23.25 ≤ 4.4<br />
⎪<br />
⎨ x1 + 2x2 −3x3 ≤ 5 (14)<br />
⎪<br />
−7x − x + x ≤<br />
⎪ 1 3 2 7 3 3<br />
⎪ 11x1 + 5x2 −6x3 ≤ 10,<br />
⎩<br />
whose optimal solution is<br />
( x1, x2, x 3) = ( -0.4286, 0,0)<br />
solved by LINGO s<strong>of</strong>tware. Furthermore, we can obtain<br />
*<br />
that x = ( -0.4286,0,0)<br />
is the expected value weakly<br />
efficient solution to problem (12) by the Theorem 4.1.<br />
Ⅴ CONCLUSIONS<br />
In this study, we mainly concerned the expected value<br />
model and the solution method <strong>of</strong> the multiobjective<br />
programming problem under bifuzzy environment. We<br />
first presented a new type <strong>of</strong> bifuzzy multiobjective<br />
programming problem. As we known, the non-inferior<br />
solutions play important role to multiobjective problem,<br />
so the expected value non-inferior solutions to the BMOP<br />
problem are presented and their relations are also studied.<br />
In addition, a solution method, called the method <strong>of</strong> main<br />
objective function, was discussed, which facilitated us to<br />
design algorithms to solve the BMOP problem.<br />
REFERENCES<br />
[1] F. Benabdelaziz, P. Lang and R. Nadeau, “Pointwise<br />
Efficiency in Multiobjective Stochastic Linear<br />
Programming,” <strong>Journal</strong> <strong>of</strong> Operational Research<br />
Society,vol.45, pp. 11-18, 2000.<br />
[2] G. De. Cooman, E.E. Kerre and F. Vanmassenhove,<br />
“Possibility theory: an Integral Theoretic Approach,”<br />
Fuzzy Sets Syst, vol. 46, pp. 287-299, 1992.<br />
[3] Y.D. Hu, “The Efficient Theory <strong>of</strong> Multiobjective<br />
Programming,” China: Shanghai Since and Technology<br />
Press, 1994.<br />
[4] G.J. Klir, “On fuzzy-set Interpretation <strong>of</strong> Possibility<br />
Theory. Fuzzy sets and Systems,” vol. 108, pp. 263-273,<br />
1999.<br />
[5] P. Kall and S.W. Wallace, “Stochastic Programming,”<br />
Chichester: Wiley, 1994.<br />
[6] B.D. Liu, Uncertainty theory, “An Introduction to its<br />
Axiomatic Foundations,”Germany: Springer-Verlag, 2004.<br />
[7] B.Liu, “Toward Fuzzy Optimization without Mathematical<br />
Ambiguity,”Fuzzy Optimization and Decision Making,”<br />
vol. 1, pp. 43-63, 2002.<br />
[8] B.Liu,“Fuzzy Random Dependent-Chance Programming,”<br />
IEEE Trans. Fuzzy Syst, vol. 9, pp. 721-726, 2001.<br />
[9] B.D.Liu,“Theory and Practice <strong>of</strong> Uncertainty Programming,<br />
Heidelberg, Physica-Verlag, 2002.<br />
[10] B. Liu, “Uncertain Programming,” New York: Wiley, 1999.<br />
[11] B. Liu, “ Random Fuzzy Dependent-Chance Programming<br />
and its Hybrid Intelligent Algorithm,” Information<br />
Sciences,” vol. 141, pp. 259-271,2002.<br />
[12] B. Liu and Y.K. Liu, “Expected Value <strong>of</strong> Fuzzy Variable<br />
and Fuzzy Expected Value Models,” IEEE Trans. Fuzzy<br />
Syst, vol. 10, pp. 445-450,2002<br />
[13] Y.K. Liu and B. Liu, “Expected Value Operator <strong>of</strong><br />
Random Fuzzy Variable Operator,” International <strong>Journal</strong>
1948 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
<strong>of</strong> Uncertainty, Fuzziness, Knowlledge-Based Systems, vol.<br />
11, pp. 195-295, 2003.<br />
[14] Y.K. Liu and S.Wang,“Theory <strong>of</strong> Fuzzy Random<br />
Optimization (China Agricultural University Press, Beijing<br />
2006).<br />
[15] C.Y. Lin and J.L. Dong, “The Efficient Theory and<br />
Method <strong>of</strong> Multiobjective Programming,” China: Jilin<br />
Educational Press, 2006.<br />
[16] B.J. Ma, “The Efficient Rate <strong>of</strong> Efficient Solution to Linear<br />
Multiobjective Programming,”<strong>Journal</strong> <strong>of</strong> Systems<br />
Engineering and Electronic Technology, vol. 2, pp. 68-106,<br />
2000.<br />
[17] I.M. Stancu-Minasian, “Stochastic Programming with<br />
Multiple Objective Functions,” Buckarest, 1984.<br />
[18] M.M. Munoz and F. Ruiz, “An interval reference pointbased<br />
method for stochastic multiobjective programming<br />
problems,” European <strong>Journal</strong> <strong>of</strong> Operational Research,<br />
vol. 197, pp. 25-35, 2009.<br />
[19] S. Nahmias, “Fuzzy variables,” Fuzzy Sets and<br />
Systems.,vol. 1, pp. 97-101, 1978.<br />
[20] H.C.Wu, “Solutions <strong>of</strong> Fuzzy Multiobjective Programming<br />
Problems Based on the Concept <strong>of</strong> Scalarization,” J Optim<br />
Theory Appl, vol. 139, pp. 361-378, 2008.<br />
© 2011 ACADEMY PUBLISHER<br />
[21] Z. Wang, and J. Klir, “Fuzzy Measure Theory, New York:<br />
Plenum Press, 1992.<br />
[22] A.V. Yazenin, “Fuzzy and Stochastic Programming,”<br />
Fuzzy Sets Syst, vol. 22, pp. 171-180, 1987.<br />
[23] R. R.Yager, “A Foundation for a Theory <strong>of</strong> Possibility,”<br />
<strong>Journal</strong> <strong>of</strong> Cybernetics, vol.10, pp. 177-204, 1980.<br />
[24] Q. Zhang, Shigeya Maeda andToshihiko Kawachi,<br />
“Stochastic multiobjective optimization model for<br />
allocating irrigation water to paddy fields,” Paddy Water<br />
Environ, vol. 5, pp. 93-99, 2007.<br />
[25] J. Zhou and B. Liu, “Analysis and Algorithms <strong>of</strong> Bifuzzy<br />
Systems,” International <strong>Journal</strong> <strong>of</strong> Uncertainty. Fuzziness<br />
and Knowledge-Based Systems, vol. 12, pp. 357-376, 2004.<br />
[26] L.A. Zadeh, “Fuzzy Sets as a Basis for a Theory <strong>of</strong><br />
Possibility,” Fuzzy Sets and Systems, vol. 1, pp. 3- 28,<br />
1978.<br />
.
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1949<br />
A Method for Building Partially Connected<br />
Neural Network<br />
Gang Li<br />
Management Department, Shanghai University for Science and Technology, Shanghai, China<br />
Email: sdlig@163.com<br />
Xingsan Qian, Chunming Ye,<br />
Management Department, Shanghai University for Science and Technology, Shanghai, China<br />
Email: qxsqg@126.com, yechm6464@163.com<br />
Abstract - This paper focuses mainly on application <strong>of</strong><br />
Partial Connected Back Propagation Neural Network<br />
(PCBP) instead <strong>of</strong> typical fully connected neural network<br />
(FCBP), as PCBP with less connections learns faster than<br />
FCBP. The initial neural network is fully connected, after<br />
training with sample data, a clustering method is employed<br />
to cluster weights between input to hidden layer and from<br />
hidden to output layer, and connections that are relatively<br />
unnecessary are deleted, thus the initial network becomes a<br />
PCBP network. PCBP can be used in prediction or data<br />
mining by training it with data that comes from database.<br />
At the end <strong>of</strong> this paper, several experiments are conducted<br />
to illustrate the effects <strong>of</strong> PCBP using the submersible pump<br />
repair data set.<br />
Index Terms - Neural Network; FCBP; PCBP; pruning<br />
I. INTRODUCTION<br />
Artificial neural networks have been proved to be a<br />
useful tool in pattern recognition and classification tasks<br />
in diverse areas like data mining, millions <strong>of</strong> databases<br />
are being used in business data management, scientific<br />
and engineering data management and other applications<br />
[1], and the most-widely used network is the standard<br />
Back Propagation (SBP) algorithm [2]. Indeed, the SBP<br />
learning algorithm has emerged as the standard algorithm<br />
for the training <strong>of</strong> multiplayer networks, and hence the<br />
one against which other learning algorithms are usually<br />
benchmarked. Actually, the SBP is fully connected, as<br />
called FCBP, and it has been commonly used as a matter<br />
<strong>of</strong> fact, since they usually do not need a priori<br />
information <strong>of</strong> data, <strong>of</strong> course, this is the feature <strong>of</strong> FCBP,<br />
but unfortunately, FCBP have several drawbacks, as<br />
reported by researchers [3]: it is extremely slow; training<br />
performance is sensitive to the initial conditions; it may<br />
become trapped in local minima before converging to a<br />
solution; oscillations may occur during learning (this<br />
usually happens when users increase the learning rate in<br />
an unfruitful attempt to speed up convergence); and, if<br />
the error function is shallow, the gradient is very small<br />
© 2011 ACADEMY PUBLISHER<br />
doi:10.4304/jcp.6.9.1949-1954<br />
Lin Zhao<br />
HP China, Shanghai, China<br />
Email: lemiozl@163.com<br />
leading to small weight changes.<br />
Also, as for FCBP, due to the learning style, the<br />
structure <strong>of</strong> the trained FCBP usually have unnecessary<br />
connections which induces the issue <strong>of</strong> the complexity <strong>of</strong><br />
the networks and causes the slow training time, especially<br />
for large networks. The complexity problem has attracted<br />
the interest <strong>of</strong> researchers because <strong>of</strong> the advantages that<br />
would be obtained by solving it. One critical advantage is<br />
that the simpler the system, the better it is [4]. So, if these<br />
unnecessary connections can be removed from the<br />
network, then training times would be greatly reduced, it<br />
is especially important for data mining, where database<br />
usually contains large number <strong>of</strong> data records ranging<br />
from millions to even billions, without faster training<br />
time, data mining using neural network is mission<br />
impossible.<br />
One way to reduce the complexity <strong>of</strong> the networks is<br />
to reduce the number <strong>of</strong> redundant connections, nodes [5],<br />
or input features. The reduction <strong>of</strong> the connections or<br />
nodes can be achieved by removing the weights that<br />
contribute the least to the network outputs. To our best<br />
knowledge, most reduction methods have been done<br />
during training networks. And one important thing is to<br />
determine what kinds <strong>of</strong> connections are redundant?<br />
II. BUILDING A PCBP NETWORK<br />
As mentioned earlier, FCBP requires more training<br />
time than PCBP (see Fig1).<br />
Fig.1 Example <strong>of</strong> PCBPs
1950 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
Generally, there are two ways to build a PCBP, one<br />
is manually; the other is automatic generated by starting<br />
from a FCBP and then prune FCBP to remove the<br />
unnecessary connections. The previous way is mandated<br />
and requires a deep insight into the data patterns involved,<br />
or else the network structure is not properly set, it may<br />
need more training time than FCBP; the latter way does<br />
not require user participated, and could determined the<br />
would-be removed connections automatically, the process<br />
is illustrated in Figure 2.<br />
Construct a FCBP<br />
Tr ai n FCBP wi t h<br />
sampl e dat a<br />
Pr une t he FCBP<br />
Sat i sf y<br />
accur acy?<br />
Yes<br />
The f i nal PCBP<br />
No<br />
Fig.2. Build a PCBP by pruning FCBP<br />
III. TRAINING FCBP<br />
Before training network, several things should be<br />
pre-defined, that is network structure including number <strong>of</strong><br />
input and hidden and output nodes, generally speaking,<br />
numbers <strong>of</strong> input and output nodes depend on sample<br />
data, as for hidden nodes, it is usually determined by<br />
experience, some researchers have reported that a few<br />
number <strong>of</strong> hidden nodes is just enough. As for error<br />
function, the typical SBP employs Mean Squared Error<br />
(MSE) as follows:<br />
k o 1<br />
i i 2<br />
Error = ∑∑(<br />
S p − t p )<br />
(1)<br />
2 i=<br />
1 p=<br />
1<br />
i<br />
Where S p stands for the actual output <strong>of</strong> output<br />
i<br />
node i, and tp for expected corresponding output value,<br />
while k is number <strong>of</strong> output nodes.<br />
Although MSE is the most widely used error function,<br />
it requires more training time and may become trapped in<br />
local minima before converging to a solution. It has been<br />
suggested by several authors, for example Lang [6] and<br />
Ooyen [7], that the cross-entropy error function improves<br />
the convergence <strong>of</strong> the training process, and can<br />
significantly reduce training time, the cross-entropy error<br />
function is as follows:<br />
© 2011 ACADEMY PUBLISHER<br />
Error=<br />
−<br />
k<br />
o<br />
∑∑<br />
i=<br />
1 p= 1<br />
i i<br />
i<br />
(t log S + ( 1−t<br />
) log(<br />
1−S<br />
)) (2)<br />
i<br />
p<br />
p<br />
During our experiments, we also found that the cross<br />
entropy error function in equation 2 is pretty good, so we<br />
employ cross entropy as error function for both FCBP<br />
and PCBP.<br />
Before deriving the equations, we need to introduce<br />
the notations used in deriving the mathematical<br />
expressions <strong>of</strong> FCBP and PCBP training.<br />
Notations:<br />
n Numbers <strong>of</strong> input nodes<br />
h Numbers <strong>of</strong> hidden nodes<br />
o Numbers <strong>of</strong> output nodes<br />
x Sample input vector<br />
t Sample output vector<br />
j<br />
w l Weights between input node l to hidden node<br />
j<br />
j<br />
v p Weights between hidden node j to output<br />
node p<br />
σ () Activation function in hidden and output layer<br />
(here, we suppose it is sigmoid)<br />
j<br />
wc l Connection status between input node l to<br />
hidden node j<br />
j<br />
vc p Connection status between hidden node j to<br />
output node p (see definition 1)<br />
For FCBP, the components <strong>of</strong> the gradient <strong>of</strong><br />
cross-entropy error function are given by equation 3 to 4:<br />
i<br />
i i<br />
i<br />
∂Error<br />
∂Error<br />
∂S<br />
p S p − t p ∂S<br />
p<br />
= × =<br />
×<br />
j<br />
i<br />
j i<br />
i j<br />
∂v<br />
∂S<br />
∂v<br />
S × ( 1−<br />
S ) ∂v<br />
i i<br />
S p − t p i<br />
i<br />
i j<br />
=<br />
× S p × ( 1−<br />
S p ) × σ ( ∑ x pwl<br />
) (3)<br />
i<br />
i<br />
S × ( 1−<br />
S )<br />
p<br />
= ( S<br />
p<br />
i<br />
p<br />
− t<br />
(( S<br />
i<br />
p<br />
i<br />
p<br />
p<br />
p<br />
) × σ (<br />
i<br />
p<br />
∑<br />
x<br />
j<br />
p<br />
i<br />
p<br />
w<br />
i ∂σ<br />
( x<br />
∂Error<br />
∂Error<br />
∂S<br />
p<br />
= ×<br />
×<br />
j<br />
i<br />
∂wi<br />
∂S<br />
p<br />
∂w<br />
∂σ<br />
( x w )<br />
=<br />
∑<br />
p<br />
j<br />
l<br />
)<br />
∑<br />
− t ) × v ) × σ (<br />
∑<br />
p<br />
i<br />
p<br />
x<br />
i<br />
p<br />
p<br />
j<br />
l<br />
w<br />
j<br />
l<br />
p<br />
p<br />
∑<br />
) × ( 1−<br />
σ (<br />
j<br />
i<br />
p<br />
i<br />
p<br />
∑<br />
(4)<br />
So, adjustment items <strong>of</strong> input to hidden weights and<br />
hidden to output weights can be calculated by equation 3<br />
and 4 plus learning rate.<br />
IV. PRUNING FCBP<br />
Before going on, we have to introduce a new<br />
definition:<br />
Definition 1: Connection Status: a vector that<br />
w<br />
j<br />
l<br />
)<br />
x w )) × x<br />
i<br />
p<br />
j<br />
l<br />
i<br />
p
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1951<br />
represents how one node connects with its adjacent nodes<br />
in the following layer.<br />
From a macro perspective view, connection status<br />
represents network structure; while from a micro point <strong>of</strong><br />
view, it is just a vector consisting <strong>of</strong> binary elements, that<br />
3<br />
is, zeros or ones. For example, if wc 2 = 1 , then we can<br />
say that the connection between second node in input<br />
layer and third node in hidden layer exists, while if<br />
wc 0 , then there is no connection between second<br />
3<br />
2 =<br />
node in input layer and third node in hidden layer.<br />
Actually, FCBP can be viewed as a particular PCBP<br />
whose connection status vectors are just ones. Hence, it<br />
can be concluded that by creating and maintaining such<br />
connection status vectors, FCBP can be easily defined<br />
and implemented.<br />
After training FCBP achieved predetermined accuracy,<br />
take 0.98 for example, unnecessary connections should be<br />
removed from the network in order to get a simple but<br />
efficiency PCBP. The pruning process consists <strong>of</strong> five<br />
steps:<br />
Step1: clustering weights between adjacent layers, it<br />
starts from the first node to the last one in the hidden or<br />
output layer, respectively.<br />
Step2: automatically determine a pruning bias that<br />
satisfied, if absolute clustered weights <strong>of</strong> connections that<br />
below this bias are all deleted, the pruning ratio can be<br />
met.<br />
Step3: removed all the connections that below the bias<br />
Step4: if network accuracy falls far below expected,<br />
then roll back pruning, set another pruning ratio, and go<br />
to step 2<br />
Step5: update connection status<br />
A. Algorithm for clustering weight<br />
The notations used in the algorithm are:<br />
β Clustering distance between connections<br />
w Weight vector <strong>of</strong> a node that contains all the<br />
connections that connect to it from its previous layer,<br />
either input to hidden layer or hidden to output layer<br />
num Length <strong>of</strong> w<br />
ClusterTyp e Represents which cluster type each<br />
w element belongs to<br />
ClusterVal ue A vector that represents clustered<br />
value<br />
ClusterSum Sum <strong>of</strong> weights that belong to a<br />
specific cluster type<br />
ClusterCou nt A vector that contains number <strong>of</strong><br />
each cluster type<br />
Count Number <strong>of</strong> clustered type, that is the length<br />
<strong>of</strong> ClusterVal ue<br />
For each node <strong>of</strong> hidden or output layer, do the<br />
following:<br />
Step1: Initially, set<br />
ClusterSum ( 1)<br />
= w(<br />
1)<br />
ClusterVal ue(<br />
1)<br />
= w(<br />
1)<br />
Count = 1<br />
ClusterCou nt ( 1)<br />
= 1<br />
© 2011 ACADEMY PUBLISHER<br />
ClusterTyp e(<br />
1)<br />
= 1<br />
Step2: for each i = 2 to num , if there exists an<br />
index j that satisfies:<br />
max w(<br />
i)<br />
− ClusterVal ue(<br />
j)<br />
< β ,<br />
j=<br />
1:<br />
Count<br />
Then it means that weight w( j)<br />
should be clustered<br />
with ClustedVal ue(<br />
j)<br />
, so set<br />
ClusterSum ( j)<br />
= ClusterSum ( j)<br />
+ w(<br />
i)<br />
ClusterCou nt ( j)<br />
= ClusterCou nt ( j)<br />
+ 1<br />
ClusterTyp e(<br />
i)<br />
= j<br />
Else, it means w( j)<br />
should be another cluster, then<br />
set<br />
Count = Count + 1<br />
ClusterSum ( count ) = w(<br />
i)<br />
ClustedVal ue(<br />
Count ) = w(<br />
i)<br />
ClusterVal ue(<br />
i)<br />
= w(<br />
i)<br />
ClusterTyp e(<br />
i)<br />
= Count<br />
Step3: calculate the average clustered value for each<br />
cluster value;<br />
For i = 1 to Count<br />
Set<br />
ClusterVal ue(<br />
i)<br />
= ClusterSum ( i)<br />
/ ClusterCou nt ( i)<br />
Step4: update w to relevant cluster value:<br />
For i = 1 to num<br />
Set<br />
w ( i)<br />
= ClusterVal ue(<br />
ClusterTyp e(<br />
i))<br />
B. Deleting unnecessary connections<br />
We define the following criterion that evaluates which<br />
kinds <strong>of</strong> connections are unnecessary: connections that<br />
have relatively small weights. Small is an obscure word<br />
that it is difficult to determine exactly, especially facing<br />
the fact that the distribution <strong>of</strong> weights after training is<br />
unpredictable, as initial weights are random numbers<br />
usually ranging from zero to one. It is not practical to set<br />
pruning bias manually. In order to solve this problem, we<br />
propose a heuristic method to automatically generate such<br />
a pruning bias that depends on the distribution <strong>of</strong> weights.<br />
The algorithm is based on pruning ratio, which is defined<br />
as:<br />
Definition 2: Pruning Ratio: numbers <strong>of</strong> pruned<br />
connections divided by total connections <strong>of</strong> the previous<br />
FCBP, it ranges from zero to one.<br />
The algorithm is as follows:<br />
Step1: let µ ( 0 < µ < 1)<br />
be a predetermined<br />
pruning ratio that indicates how much connections should<br />
be pruned; let ϖ be the best pruning bias; num is the<br />
accumulated number <strong>of</strong> connections;<br />
Step2: sort ClusterVal ue in ascending order, at<br />
the same time, ClusterCou nt changes accordingly in<br />
order to make the two vectors still consistent with each<br />
other; set<br />
index = 1, num = ClusterCount(<br />
1)<br />
, ϖ = ClusterValue(<br />
1)<br />
Step3: do the following loop
1952 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
WHILE( index < Count) do<br />
num<br />
IF( < µ ) THEN<br />
Count<br />
index = index + 1;<br />
num = num + ClusterCount( index);<br />
Continue<br />
ELSE<br />
ϖ = ClusterValue( index); Exit<br />
end<br />
After step 3, we can get the best pruning bias. When<br />
deleting unnecessary connections, we just set the<br />
corresponding elements in vector wc and vc to be<br />
zeros, and by doing this, a FCBP becomes a PCBP<br />
network.<br />
When pruning FCBP, one thing should be paid special<br />
attention to which may ignored by other researchers. To<br />
illustrate it, look at figure 3, on the left part is a pruned<br />
network which has three hidden nodes marked as A, B<br />
and C respectively. Notice that, node A has two<br />
connections with nodes in output layer (in bold line),<br />
while none with input layer, as they have been deleted as<br />
unnecessary connections in the above pruning process,<br />
though the probability <strong>of</strong> happening is small, but if it<br />
does happen, it must be handled properly, or else some<br />
error would happen, for example, for any kind <strong>of</strong> input<br />
pattern, the output <strong>of</strong> node A is always one if activation<br />
function <strong>of</strong> node A is sigmoid, because sum <strong>of</strong> input plus<br />
weight is zero, if node A has a bias, then the output <strong>of</strong><br />
node A can be any float number as long as it varies with<br />
the value <strong>of</strong> bias, thus the actual output <strong>of</strong> output node is<br />
affected. For this kind <strong>of</strong> conditions, we propose that<br />
node A be deleted too, that is to delete connections<br />
between A and the output layer, then A is totally removed<br />
from the network as illustrated at the right part <strong>of</strong> figure 2.<br />
If node A has connections with its previous layer but none<br />
with following layer, then delete its connections with its<br />
previous too. By doing this, pruning connections and<br />
nodes can be handled together.<br />
Fig.3 Example <strong>of</strong> inconsistent connections<br />
© 2011 ACADEMY PUBLISHER<br />
V. TRAINING PCBP<br />
When apply PCBP in a specific domain, like data<br />
mining, usually, PCBP should be trained first just like<br />
FCBP does. The training process <strong>of</strong> PCBP is a slightly<br />
different comparing with FCBP, take a hidden node for<br />
example, not all the input nodes connect with it, so the<br />
actual input for hidden node is calculated by:<br />
n<br />
∑<br />
i=<br />
1<br />
i i i<br />
( w × x × wc ) , notice that an additional items<br />
l<br />
l<br />
(connection status between input layer and hidden layer)<br />
is added, similarly, equation 3 and 4 change to equation 5<br />
and 6:<br />
∂Error i i<br />
i j j<br />
= ( S p − t p ) × σ ( ∑ x pwl<br />
wcl<br />
) (5)<br />
j<br />
∂v<br />
∂Error<br />
=<br />
∂w<br />
j<br />
i<br />
p<br />
∑<br />
(( S − t ) × v ) × σ ( x w wc ) × ( 1−<br />
σ (<br />
i<br />
p<br />
i<br />
p<br />
j<br />
p<br />
∑<br />
i<br />
p<br />
j<br />
l<br />
VI. EXPERIMENTS<br />
j<br />
l<br />
∑<br />
x w wc )) × x<br />
Enterprise data warehouse<br />
Sources Users<br />
Operation<br />
al<br />
database<br />
Operation<br />
al<br />
database<br />
Data files<br />
Sourcing Area<br />
Staging<br />
History<br />
Area<br />
Integrated Area<br />
Meta data and Security<br />
Finance<br />
Sales<br />
Marketing<br />
MPP (Massively Parallel Processing)<br />
Figure 4: data warehouse architecture<br />
In the data warehouse architecture, from Staging Area<br />
to Enterprise Report are considered as Enterprise Data<br />
Warehouse, because every one <strong>of</strong> them is integral part <strong>of</strong><br />
a warehouse, and they can satisfy the current and future<br />
needs for all the business users across the enterprise.<br />
It is common understanding that data warehouse is<br />
basis for BI and DSS application, and implementing a<br />
successful data warehouse requires not only technologies<br />
but also methodology as well as culture and cooperation<br />
across the enterprise.<br />
The experiment data set which recorded submersible<br />
pump repair history contains four attributes classification<br />
codes, these attributes are separately: Single rotor electric<br />
power (kW Per Rotor), Cable Temperature Level(℃),<br />
Casing size(inch), Protector Length(m). In the following<br />
experiments, we use the data set to train FCBP and PCBP.<br />
The experiment data set which recorded submersible<br />
pump repair history contains four attributes classification<br />
codes, these attributes are separately: Single rotor electric<br />
power (kW Per Rotor), Cable Temperature Level(℃),<br />
Casing size(inch), Protector Length(m). In the following<br />
experiments, we use the data set to train FCBP and PCBP.<br />
i<br />
p<br />
j<br />
l<br />
(6)<br />
Enterprise Report<br />
j<br />
l<br />
i<br />
p<br />
`
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1953<br />
First, we need to encode the numeric data into binary,<br />
as illustrated in table 1:<br />
Table 1 The data attribute encoding table<br />
Attributes Range Encoded<br />
Input<br />
Single rotor Single rotor electric power >8 1 1 1<br />
electric<br />
power 4
1954 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
Fig 5: A PCBP after pruning<br />
VII. CONCLUSION<br />
We have presented here that how to construct a PCBP<br />
with cross-entropy as error function, and pruning<br />
algorithm was discussed in detail. Also the experiments<br />
showed that a PCBP had fewer connections, but still<br />
remained accuracy. When PCBP is applied in data<br />
recognition or other fields, like data mining, it learns<br />
faster than FCBP does, especially trained with a huge<br />
amount <strong>of</strong> sample data.<br />
ACKNOWLEDGMENTS<br />
The research work presented in this paper has been<br />
partially supported by science research foundation project<br />
<strong>of</strong> Shanghai bureau <strong>of</strong> education (NO.08YS103), and<br />
Shanghai important education subject project<br />
(NO.S20504). The authors would like to express their<br />
appreciation to the agency.<br />
REFERENCES<br />
[1] [1] Changchien, S.W.&Lu T. Mining association rules<br />
procedure to support on-line recommendation by<br />
customer and products fragmentation. Expert Systems<br />
with Applications, 20, 2001,pp:325-335.<br />
[2] [2] D. Rumelhart, G. Hinton, and R. Williams, Parallel<br />
Distributed Processing. MIT Press,Cambridge, MA, 1986.<br />
[3] [3] D. Sarkar, “Methods to speed up error back<br />
propagation learning algorithm,” ACM Computing<br />
Surveys, vol. 27, no. 4, pp. 519–542, 1995.<br />
[4] [4] Sanggil Kang, Can Isik. Patially Connected<br />
Feedforward Neural Network Structured by Input Types.<br />
IEEE Transactions on Neural Networks, Vol.16, No.1,<br />
January 2005.<br />
[5] [5] D. E. Duckro, D. W. Quinn, and S. J. Gardner, “Neural<br />
network pruning with Tuckey Kramer multiple<br />
comparison procedure,” Neural Computat., vol. 14, pp.<br />
1149–1168, 2002.<br />
© 2011 ACADEMY PUBLISHER<br />
[6] [6] Lang K.J and Witbrock M.J. Learning to tell two<br />
spirals apart. In proc. <strong>of</strong> the 1988 Connectionist Summer<br />
School, pp:52-59. Morgan Kaufmann, San Mateo, CA.<br />
[7] [7] A.Van Ooyen. Improving the Covergence <strong>of</strong> the<br />
Back-Propagation Algorithm. Neural Networks, Vol.5,<br />
1992, pp:465-471.<br />
Gang Li, Ph.D., born in Dong Yin<br />
city, Shan Dong Province , on 1970-08-04.<br />
He has a MS in computer science from<br />
Xi’an Jiaotong University and PhD in<br />
Information Management System from<br />
Donghua University.<br />
He is an associate pr<strong>of</strong>essor in the<br />
Management Science and Technology at<br />
Management Department, Shanghai<br />
University for Science and Technology, China. In 2005/2009,<br />
he was an PhD candidate in management information system at<br />
Donghua University. His industrial career includes Information<br />
Management Department, Power Machinery Factory <strong>of</strong> Shengli<br />
oilfield. (1993-2004). Then he has been a teacher <strong>of</strong> Dongying<br />
Vocational College(2004-2005).<br />
Xingsan Qian ,Pr<strong>of</strong>essor, Director <strong>of</strong> Shanghai branch <strong>of</strong><br />
industrial engineering in Mechanical Engineering Society <strong>of</strong><br />
China, president <strong>of</strong> industrial engineering teaching seminars in<br />
east China ,Shanghai registered consultants, Committee <strong>of</strong><br />
Shanghai Institute <strong>of</strong> Electronics, Microelectronics.Pr<strong>of</strong>essor<br />
Qian Xingsan is a famous expert in industrial engineering,<br />
logistics and engineering. He has undertake more than 40<br />
research subject about development, innovation and<br />
development, strategic over 20 years,has twice served as<br />
national information technology (or IC) policy drafting group<br />
members.<br />
His research areas:high (IC) industry development (strategy,<br />
reform, planning, industrial zone location); management science<br />
and engineering; Technology Management; regional innovation;<br />
Industrial Engineering .<br />
Awards: Has won third prize <strong>of</strong> Shanghai Science and<br />
Technology Progress Award 2 times, won second prize <strong>of</strong><br />
Shanghai Science and Technology Progress Award.won second<br />
prize <strong>of</strong> National Science and Technology Progress Award. He<br />
has take charge <strong>of</strong> and participated more than 40 scientific<br />
research projects,has published 3 Book.<br />
Chunming Ye, Ph.D., Pr<strong>of</strong>essor, Industrial Engineering<br />
expert <strong>of</strong> China, Secretary <strong>of</strong> the Shanghai Institute <strong>of</strong> Industrial<br />
Engineers, winner <strong>of</strong> Baosteel teachers in 2008. His research<br />
areas include: industrial engineering, intelligent algorithms,<br />
enterprise resource planning, supply chain management and<br />
evaluation <strong>of</strong> intellectual property, production planning and<br />
scheduling . He is the earlier <strong>of</strong> using cultural evolution <strong>of</strong><br />
swarm algorithm applied to the field <strong>of</strong> production planning and<br />
scheduling researchers.He han won the first prize <strong>of</strong> scientific<br />
and technological progress <strong>of</strong> the State Administration <strong>of</strong><br />
Machinery Industry (1999) ; third prize in 2005 in Shanghai<br />
teaching achievements; second Prize <strong>of</strong> 2009 State Education<br />
Commission Science and Technology Progress Award,and has<br />
published over one hundred and sixty papers at home and<br />
abroad.<br />
Lin Zhao Master, he received the degree from Dong Hua<br />
University in 2006. His research interests are data mining,<br />
artificial intelligence, knowledge management, etc.
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1955<br />
A Cooperative Co-evolution PSO for Flow Shop<br />
Scheduling Problem with Uncertainty<br />
Bin Jiao<br />
Electric School, Shanghai DianJi University<br />
No.690, Jiang Chuan Rd., Min Hang District, Shanghai, 200240, China<br />
Email: abinjiaocn@163.com<br />
Qunxian Chen<br />
Electronic Information School, Shanghai DianJi University<br />
No.690, Jiang Chuan Rd., Min Hang District,Shanghai,200240, China<br />
Email: bchenqx@sdju.edu.cn<br />
Shaobin Yan<br />
School <strong>of</strong> Information Science and Engineering, East China University <strong>of</strong> Science and Technology<br />
No.130, Mei Long Rd., 200370, Shanghai, China<br />
Email: cyshaobin123@sina.com<br />
Abstract—Considering current situation <strong>of</strong> production<br />
scheduling with uncertainties in modern manufacturing<br />
enviroments, flow shop production scheduling model is<br />
established based on the theory <strong>of</strong> fuzzy programming, in<br />
which fuzzy processing time is considered and the duration<br />
time <strong>of</strong> intermediate is unlimited. The maximum<br />
membership function <strong>of</strong> mean value has been applied to<br />
solve the non-linear fuzzy scheduling model in order to<br />
convert the fuzzy optimization problem to the general<br />
optimization problem. Finally, a cooperative<br />
co-evolutionary particle swarm optimization algorithm<br />
based on catastrophe added to improve the diversity <strong>of</strong> the<br />
swarm (CCPSO) is adopted to solve flow shop production<br />
scheduling with uncertainty within infinite intermediate<br />
storage and the simulation results obtained are effective and<br />
satisfactory.<br />
Index Terms—uncertainty; fuzzy programming; Flow Shop<br />
scheduling problem; cooperative co-evolutionary particle<br />
swarm optimization algorithm;<br />
I. INTRODUCTION<br />
Production scheduling tackles effective allocation <strong>of</strong><br />
production resources over time. Flow shop scheduling<br />
problem(FSSP), which represents nearly a quarter <strong>of</strong><br />
manufacturing systems and information service facilities<br />
in use nowadays, is one <strong>of</strong> the most important issues in<br />
shop floor control <strong>of</strong> a manufacturing firm,. The first<br />
research conducted on the flow shop scheduling problem<br />
was proposed by Johnson (1954) [1], who developed an<br />
optimization algorithm to achive a minimum makespan<br />
for the n-jobs and 2-machines flow shop scheduling<br />
problem. Previously, researchers mainly studied flow<br />
shop scheduling problem based on ideal conditions such<br />
as processing time assigned or estimated as a fixed value<br />
etc.In many real world applications, however, there exist<br />
many uncertain factors including human intervention,<br />
incomplete information, and uncertain environment.<br />
Recently, considering lots <strong>of</strong> uncertain factors that appear<br />
© 2011 ACADEMY PUBLISHER<br />
doi:10.4304/jcp.6.9.1955-1961<br />
in operations, planning and other processes, researchers<br />
mostly conduct researches on uncertain processing times<br />
and due dates in the real world applications and use fuzzy<br />
number theory to describe this problem [2,3].<br />
As to scheduling problems with fuzzy processing time,<br />
a few approaches have been developed. In fact, large<br />
quantities <strong>of</strong> uncertainties including fuzzy processing<br />
time and fuzzy due dates are always considered [4].<br />
Fortemps [5], developing a fuzzy approach in job shop<br />
scheduling problem with imprecise durations, enrolled<br />
the important application <strong>of</strong> the uncertainty in time<br />
parameters. Chanas studied minimization <strong>of</strong> maximum<br />
lateness <strong>of</strong> jobs in a single machine scheduling problem<br />
[6]. Litoiu and Tadei [7] proposed some novel models for<br />
real-time task scheduling with fuzzy processing times and<br />
deadlines. Hong [8] applied triangular membership<br />
functions for flexible flow shop problem with two<br />
machine centers to examine uncertain processing times.<br />
Wu Chaochao [9] used an efficient genetic algorithm to<br />
solve single machine scheduling problems with fuzzy<br />
processing time and multiple objectives. Niu Qun [10]<br />
proposed a novel particle swarm optimization for flow<br />
shop scheduling problem with fuzzy processing time. Xu<br />
Zhenhao [11] established a scheduling model for flow<br />
shop problems with finite intermediate and adopted a<br />
fuzzy immune algorithm to optimize this problem. In this<br />
paper, flow shop scheduling problem with fuzzy<br />
processing time is considered and the time duration <strong>of</strong><br />
intermediate is unlimited. Besides, a cooperative<br />
co-evolutionary particle swarm optimization algorithm<br />
based on catastrophe (CCPSO) is adopted to verify the<br />
model and to solve the fuzzy scheduling problem.<br />
Ecological models and co-evolutionary architectures are<br />
effective methods to improve the performance <strong>of</strong> original<br />
particle swarm optimizer [12, 13]. And co-evolutionary<br />
scheme, which is inspired by the reciprocal evolutionary<br />
change driven by the cooperative [14] or competitive
1956 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
interaction [15] between different species, can avoid the<br />
exponential increase in difficulty by dividing the search<br />
space into several smaller subspaces, and then conducting<br />
the overall optimization process over smaller regions.<br />
The remainder <strong>of</strong> the paper is designed as follows. In<br />
section 2, fuzzy scheduling problem is depicted briefly.<br />
The following section 3 introduces cooperative<br />
co-evolutionary particle swarm optimization algorithm<br />
based on catastrophe (CCPSO) algorithm. Experiment is<br />
undertaken in section 4. Finally, we draw a conclusion in<br />
section 5.<br />
II. PROBLEM DESCRIPTION<br />
Flow shop scheduling problem (FSSP) is <strong>of</strong>ten<br />
expressed by the symbols<br />
n/ m/ P/ Obj .<br />
, in which n<br />
jobs J = {1, 2, . . . , n} have to be processed on m<br />
machines M = {1, 2, . . . , m}, P shows that only<br />
permutation schedules are considered and<br />
Obj .<br />
, the<br />
objective function, describes the performance measure by<br />
which the schedule is to be evaluated. Also, all machines<br />
should process all jobs according to the sequence <strong>of</strong><br />
pre-defined permutation schedule. Hence a schedule is<br />
uniquely represented by a permutation <strong>of</strong> jobs. At any<br />
time, each machine can only process one job and each job<br />
can only be processed by one machine.<br />
A triangular fuzzy number is given to describe the<br />
uncertain processing time <strong>of</strong> products in this paper. The<br />
maximum membership function is defined as follows:<br />
⎧0,<br />
⎪ L<br />
⎪ c − x<br />
,<br />
⎪ M L<br />
x − x<br />
µ x(<br />
c)<br />
= ⎨ U<br />
⎪ x − c<br />
,<br />
⎪ U M<br />
x − x<br />
⎪<br />
⎩0,<br />
A. Problem Definition<br />
x<br />
x<br />
c ≤ x<br />
L<br />
M<br />
< c ≤ x<br />
< c ≤ x<br />
c > x<br />
The following notation has been introduced to<br />
describe the problem more precisely.<br />
N ——a set <strong>of</strong> n products which must be processed,<br />
N = { 1,<br />
2,<br />
,<br />
i,<br />
n}<br />
;<br />
M ——a set <strong>of</strong> m processing units which are<br />
M = 1,<br />
2,<br />
,<br />
j,<br />
m<br />
;<br />
available for our purpose, { }<br />
T ~<br />
ij<br />
——The processing time <strong>of</strong> products i on unit j,<br />
which includes the transfer time, the set-up time, the<br />
clean-up time, and so on. Because it is mutative and<br />
uncertain, it is represented by the triangular fuzzy<br />
number;<br />
S ~<br />
ij<br />
——the starting time <strong>of</strong> product i processed on<br />
unit j, the parameter also is uncertain;<br />
© 2011 ACADEMY PUBLISHER<br />
L<br />
U<br />
M<br />
U<br />
(1)<br />
C ij<br />
~<br />
——the completing time <strong>of</strong> product i processed<br />
on unit j , and it is represented by the triangular fuzzy<br />
number;<br />
S ie<br />
~<br />
——the starting time <strong>of</strong> the last operation <strong>of</strong><br />
product i; and it is represented by the triangular fuzzy<br />
number;<br />
T ie<br />
~<br />
——the processing time <strong>of</strong> the last operation <strong>of</strong><br />
product i, and it is represented by the triangular fuzzy<br />
number;<br />
In flow shop scheduling problem, every job has the<br />
same sequence <strong>of</strong> operating on all machines. All jobs are<br />
processed at time-zero. But the following constraints<br />
must be taken into account:<br />
Sequence Constraints<br />
T ~<br />
S ~<br />
S ~<br />
ij ≥ i(<br />
j−1)<br />
+ i(<br />
j−1)<br />
, i ∈ N , j ∈ M<br />
(2)<br />
Equation (2) indicates that the operation <strong>of</strong> product i<br />
on unit j can start after completing its previous processing<br />
procedure, that is the starting time <strong>of</strong> each operation <strong>of</strong><br />
product i can be more than or equal to the finishing time<br />
<strong>of</strong> the last operation. And the different procedure <strong>of</strong> the<br />
same product can not be operated at the same time.<br />
Resource Constraints<br />
S ~<br />
ij<br />
≥ −<br />
S ~<br />
( i−1)<br />
j + ( i 1)<br />
j i ∈ N , j ∈ M<br />
T ~<br />
Equation (3) means that the product i on unit j can<br />
start after the completion <strong>of</strong> the previous product i-1, that<br />
is the same unit can’t process two or more different<br />
products at a time.<br />
Time Constraints<br />
(3)<br />
S i N , j M<br />
~ ij ≥ 0 ∈ ∈<br />
(4)<br />
Equation (4) represents each product can be available<br />
at time zero.<br />
Moreover, we make the following assumptions<br />
regarding the process: there is no priority between<br />
products; once an operation has started, it can’t be<br />
interrupted unless having been finished; a unit can not<br />
process different products at one time, and a product<br />
can’t be processed by more than one unit simultaneously.<br />
There are many different optimal objectives, i.e. the<br />
maximum or average tardiness, the average flow time, the<br />
lateness and earliness and so on. In this paper, the<br />
scheduling goal is to find a feasible schedule which<br />
minimizes the maximum completion time, which is<br />
makespan:<br />
( ) ( ie ie ) T~ S ~<br />
min makespan = min max +<br />
(5)<br />
In order to calculate the completion time <strong>of</strong> products<br />
with fuzzy durations, the addition and maximum<br />
operations are needed.
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1957<br />
x~ = ( x )<br />
Defining 1,<br />
x2<br />
, x3<br />
y~ = ( y )<br />
and 1,<br />
y2<br />
, y3<br />
be<br />
the triangular fuzzy numbers, the addition and maximum<br />
operations are given in the form as follows:<br />
Fuzzy<br />
x~ + y~ = ( x<br />
)<br />
Addition: 1+<br />
y1,<br />
x2<br />
+ y2,<br />
x3<br />
+ y3<br />
:<br />
x~ ∨ y~ = ( x<br />
)<br />
Fuzzy maximum: 1 ∨ y1,<br />
x2<br />
∨ y2<br />
, x3<br />
∨ y3<br />
B. Fuzzy Scheduling Model based on Triangular<br />
Fuzzy Number<br />
x~ = x , x , x<br />
( )<br />
Triangular fuzzy number 1 2 3<br />
adopted to express uncertain processing time. Due to the<br />
resolvability <strong>of</strong> the fuzzy addition and maximum<br />
operations, the detail <strong>of</strong> solution can be described as<br />
follows:<br />
1) Ifi = 1 ,<br />
ij<br />
ij<br />
j = 1:<br />
ij ij T ij<br />
~<br />
T ~<br />
S ~<br />
C ~<br />
S ,<br />
~<br />
= 0<br />
(6)<br />
= + =<br />
2) Ifi = 1 , j > 1:<br />
ij i(<br />
j ) ij ij ij i(<br />
j ) T ij<br />
~<br />
C ~<br />
T ~<br />
S ~<br />
C ~<br />
C ,<br />
~<br />
S ~<br />
(7)<br />
= −1 = + = −1<br />
+<br />
3) If i > 1 , j = 1:<br />
ij ( i ) j<br />
ij ij ij ( i ) j T ij<br />
~<br />
C ~<br />
T ~<br />
S ~<br />
C ~<br />
C ,<br />
~<br />
S ~<br />
(8)<br />
= −1 = + = −1<br />
+<br />
4) If i > 1 , j > 1:<br />
ij<br />
( i ) j C i(<br />
j ) ,<br />
~<br />
C ,<br />
~<br />
S max<br />
~<br />
= −1 −1<br />
(9)<br />
5) Objective function is:<br />
( ) ( ie ie )<br />
C ~<br />
T<br />
min<br />
~<br />
S ~<br />
min makespan = min max +<br />
=<br />
( ) ij ij T ij<br />
~<br />
S ~<br />
C ~<br />
= +<br />
( ie )<br />
L M U ( C , C , C )<br />
= min ie ie ie (10)<br />
As is shown above, the fuzzy programming problem is<br />
transformed into multi-objective programming model.<br />
C<br />
L<br />
ie<br />
, C<br />
M<br />
ie<br />
, C<br />
U<br />
ie<br />
Owing to<br />
being related to fuzzy<br />
L M U<br />
Tij<br />
, Tij<br />
, Tij<br />
processing time<br />
respectively, the solutions<br />
get by multi-objective programming model are also the<br />
worst solution, the most possible solution and the best<br />
solution <strong>of</strong> fuzzy programming model. So the following<br />
task is to apply the maximum membership function <strong>of</strong><br />
mean value to manage to obtain a single objective model.<br />
C. Model Transformation<br />
Here, Zimmerman method is applied to transform ie C~<br />
into two solutions including positive ideal solution (PIS)<br />
PIS<br />
NIS<br />
C ie<br />
C<br />
and negative ideal solution (NIS) ie , that is<br />
C ( k 1 , 2,<br />
3)<br />
PIS<br />
k =<br />
C ( k 1 , 2,<br />
3)<br />
and NIS<br />
k =<br />
respectively<br />
© 2011 ACADEMY PUBLISHER<br />
is<br />
formulated as follows:<br />
C<br />
C<br />
PIS<br />
1<br />
NIS<br />
1<br />
where,<br />
= min<br />
C<br />
= max C<br />
L<br />
ie<br />
L<br />
ie<br />
,<br />
,<br />
PIS<br />
C ie and<br />
C<br />
C<br />
PIS<br />
2<br />
NIS<br />
2<br />
NIS<br />
Cie<br />
= min C<br />
= max C<br />
M<br />
ie<br />
M<br />
ie<br />
,<br />
,<br />
C<br />
C<br />
PIS<br />
3<br />
NIS<br />
3<br />
= min C<br />
= max C<br />
(11)<br />
U<br />
ie<br />
U<br />
ie<br />
also represent the optimistic<br />
solution and the pessimistic solution, by which we can<br />
define another kind <strong>of</strong> membership function like this:<br />
µ<br />
Ck<br />
( x)<br />
⎧0,<br />
⎪<br />
⎪ x − Cie<br />
= ⎨ NIS<br />
⎪Cie<br />
− C<br />
⎪<br />
⎩1,<br />
PIS<br />
PIS<br />
ie<br />
,<br />
C<br />
PIS<br />
ie<br />
x > C<br />
≤ x ≤ C<br />
x < C<br />
NIS<br />
ie<br />
PIS<br />
ie<br />
NIS<br />
ie<br />
k = 1,<br />
2,<br />
3<br />
Then, the fuzzy scheduling model above can be<br />
transformed into the singular objective nonlinear<br />
objective model :<br />
L<br />
U<br />
max { Γα + ( 1 − Γ)<br />
α }<br />
L<br />
s.t. α ≤ µ<br />
U<br />
≤ α , k = 1, 3<br />
µ<br />
C<br />
k<br />
U<br />
L U<br />
α ≤ µ C , α ∈[<br />
0, 1]<br />
C<br />
( x)(<br />
k = 1 , 2,<br />
3)<br />
2<br />
α (13)<br />
where k<br />
is satisfactory membership<br />
function <strong>of</strong> ie C~ L<br />
. α depends on the minimum value <strong>of</strong><br />
µ C ( x)(<br />
k = 1 , 2,<br />
3)<br />
U<br />
k<br />
. α is determined by the<br />
µ C ( x)(<br />
k = 1 , 2,<br />
3)<br />
maximum value <strong>of</strong> k<br />
. During the<br />
actual decision-making process, the highest level <strong>of</strong><br />
satisfaction <strong>of</strong> objective value is expected to gain in the<br />
most possible situation not in the worst or optimal cases.<br />
µ C ( x)<br />
Therefore, in the model above, let 2 be the<br />
maximum membership value while the minimum one is<br />
produced in the worst or the best circumstances. And the<br />
operator Γ is used to reflect the tendency degree <strong>of</strong><br />
decision-maker choosing positive side and negative side.<br />
The smaller is the value Γ , the more positive is<br />
decision-making, on the contrary, the more negative is<br />
decision.<br />
III. COOPERATIVE PARTICLE SWARM OPTIMIZATION<br />
WITH CATASTROPHE (CCPSO)<br />
A. Review <strong>of</strong> Particle Swarm Optimization.<br />
Particle Swarm Optimization (PSO) is an evolutionary<br />
calculation technique proposed by Kennedy and<br />
Eberhart[16] in the mid 1990s. Different from other<br />
algorithm, PSO is simple and easily implemented due to<br />
having no operators such as crossover and mutation. It<br />
was inspired by the natural biologic phenomenon seen in<br />
a flock <strong>of</strong> birds attempting to find food through its own<br />
position as well as experience gained from others. The<br />
population <strong>of</strong> PSO is called a swarm and each individulal<br />
(12)
1958 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
in the population is called a particle. PSO is an<br />
evolutionary computation technique through individual<br />
improvement plus population cooperation and<br />
competition. A particle’s status among the search space is<br />
characteristic with its position and velocity. Then, the<br />
position and the velocity are adjusted according to its<br />
own and its companions’ flying experience.<br />
Suppose that there is an d-dimensional search space for<br />
a swarm with m particles , and the i th particle is<br />
denoted by an d-dimensional vector<br />
X i = ( xi1, xi2, ,<br />
xid)<br />
while its velocity is represented<br />
V<br />
by i = ( vi1, vi2, , vid)<br />
.Also, two key points directing<br />
particles moving to the best solution are i P and g P , <strong>of</strong><br />
P<br />
which i = ( pi1, pi2, ,<br />
pid)<br />
means the best<br />
previously visited position <strong>of</strong> the particle i and<br />
P = ( p , p , ,<br />
p )<br />
g g1 g2 gd<br />
means the position <strong>of</strong> the<br />
best individual <strong>of</strong> the whole swarm. The fitness value <strong>of</strong><br />
each particle is evaluated by the objective function.<br />
During all the iteration, the velocity and position are<br />
updated according to the following equations:<br />
( k + 1)<br />
= vid<br />
( k)<br />
+ c1r1<br />
( pid<br />
( k)<br />
− xid<br />
( k))<br />
+ c2r2<br />
( pgd<br />
( k)<br />
− x ( k))<br />
(14)<br />
( k + 1)<br />
= x ( k)<br />
+ v ( k + 1)<br />
vid id<br />
xid id id<br />
( i= 1,2, , m; d = 1,2, , d)<br />
(15)<br />
where k is the iterative number, the variables<br />
c1,<br />
c2<br />
are<br />
learning factors, usually<br />
c1<br />
= c2<br />
= 2<br />
, which assign a<br />
fixed range for a particle’s moving and 1 r<br />
,<br />
r2 are<br />
elements from two uniform random sequences in the<br />
range (0, 1):<br />
r 1 ~U (0,1);<br />
r 2 ~U (0,1).<br />
B. Principle <strong>of</strong> Cooperative Co-evolution Algorithm<br />
Co-evolution mechanism, obviously a biologic process<br />
where population <strong>of</strong> interacting individuals challenge<br />
eachother in an ongoing <strong>of</strong> adaptation, can be classified<br />
into two main categories, cooperative co-evolution and<br />
competitive co-evolution. For cooperative co-evolution,<br />
in natural ecosystems, almost all species own appetence<br />
to interact with other species to improve the survival<br />
cooperatively. We can name the cooperative co-evolution<br />
as symbiosis, firstly introduced by German mycologist,<br />
Anton de Bary in 1879 [17].<br />
As mentioned above, symbiosis is made up <strong>of</strong> three main<br />
categories including mutualism (both species benefit by<br />
the relationship), commensalism (one species benefits<br />
while the other species is not affected), and parasitism<br />
(one species benefits and the other is harmed) [18]. In this<br />
paper, we choose the mutualism and incorporate it into<br />
QPSO.<br />
(1) Form sub-swarms<br />
It is important to make up several sub-swarms for<br />
co-evolution algorithm. In this text, an initial main<br />
© 2011 ACADEMY PUBLISHER<br />
population is randomly generated, and each particle in the<br />
population has an initial main vector. Then according to a<br />
divide-up parameter set at the beginning, we separate the<br />
initial main vector into several sub-vectors. Through a<br />
kind <strong>of</strong> cooperative method introduced below, a<br />
newly-combined main vector is reached. Finally, the<br />
number <strong>of</strong> sub-swarm depends on the value <strong>of</strong> the<br />
divide-up parameter.<br />
(2) Design cooperative method<br />
Besides, how to design a cooperative method for<br />
sub-swarms is an important part <strong>of</strong> co-evolution<br />
algorithm. Generally speaking, cooperative method can<br />
be classified into three main categories namely, greedy,<br />
conservative and meta-heuristic methods [19]. Taking<br />
into account the advantage <strong>of</strong> fast convergence velocity,<br />
greedy method is applied to the co-evolution. In other<br />
words, the best particle <strong>of</strong> every sub-swarm is taken as<br />
the representative. Then we can gain a novel complete<br />
vector on condition that own current sub-vector and<br />
others’ representative are combined correctly.<br />
Fig. 1 An example for the cooperative method<br />
C. Catastrophe Operation in CCPSO<br />
In the process <strong>of</strong> searching best solution, velocity <strong>of</strong> a<br />
particle may be zero soon, which leads population to trap<br />
into local optimal. Therefore, on the basis <strong>of</strong><br />
co-evolutionary particle swarm optimizer, catastrophe<br />
operation is brought into getting a novel algorithm. The<br />
catastrophe operation plays a part <strong>of</strong> judging every<br />
sub-swarm’s solution whether in a local convergence<br />
region and carries out some measures to ensure<br />
population’s global search.<br />
M<br />
Assuming that 1 = ( x1, x2,..., xn)<br />
and<br />
M 2 = ( y1, y2,..., yn)<br />
are two random individuals in<br />
n dimensional space, each variable in 1 M and 2 M is<br />
encoded as a m system, that is<br />
xi = ( xx i1 i2... xil<br />
) y<br />
i i = ( yi1yi2... yil<br />
)<br />
, i . So the<br />
Hamming Distance between 1 M and 2 M is defined<br />
as follows:<br />
n li<br />
H( M1, M2) = ∑∑ | xij − yij<br />
| (16)<br />
i= 1 j=<br />
1
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1959<br />
xij = yij ( ∀i,<br />
j)<br />
From equation (3), when<br />
( , ) 0<br />
H M1 M 2 =<br />
, and the maximum value <strong>of</strong> the<br />
Hamming Distance between<br />
H( M1, M2) = ( m−1) ∑ li<br />
(17)<br />
n<br />
i=<br />
1<br />
1 M and<br />
Assuming that 1 1 2<br />
M = ( y , y ,..., y )<br />
2 M is:<br />
M = ( x , x ,..., x )<br />
2 1 2 n are two random individuals in<br />
n dimensional space, each variable in 1 M and 2 M is<br />
x = ( xx... x)<br />
encoded as a m i i1 i2 il<br />
system, that is i<br />
y = ( y y ... y )<br />
i i1 i2 ili<br />
. So the dissimilarity factor is<br />
expressed as follows:<br />
( , )<br />
H M1 M2<br />
µ =<br />
n<br />
( m−1) ∑l<br />
i<br />
i=<br />
1<br />
n<br />
(18)<br />
Obviously, the value <strong>of</strong> dissimilarity factor ranges from<br />
0 to 1.When two individuals are the same, the<br />
dissimilarity factor µ = 0<br />
. And the bigger µ is, more<br />
diversiform the population is. So it’s important to ensure<br />
the great difference between two different individuals.<br />
In the co-evolutional particle swarm optimizer, M<br />
sub-swarms are randomly separated into M/2 pairs <strong>of</strong><br />
individuals. Then average dissimilarity factor µ is<br />
computed:<br />
M /2<br />
∑ µ i<br />
i=<br />
1<br />
(19)<br />
µ =<br />
M /2<br />
When average dissimilarity factor µ is very small,<br />
the local convergence appears. And catastrophe factor<br />
Ca is set. When µ < Ca<br />
, the catastrophe operation is<br />
adopted. In this paper, to reach a more ideal effect, the<br />
catastrophe operation, keeping best solution and<br />
reinitializing other particles, is chosen to increase<br />
diversity <strong>of</strong> population.<br />
IV. EXPERIMENTAL RESULTS<br />
To illustrate the effectiveness and performance <strong>of</strong><br />
CCPSO for flow shop scheduling with fuzzy processing<br />
times to minimize makespan proposed in this paper, the<br />
scheduling problem <strong>of</strong> ten jobs on five machines has been<br />
selected to test. The fuzzy operating time <strong>of</strong> jobs on<br />
machines represented with triangular fuzzy number are<br />
listed in TABLE I.<br />
© 2011 ACADEMY PUBLISHER<br />
,<br />
,<br />
,<br />
During experiments, every run is repeated for 10 times<br />
and the population size is 60. The maximum iterative<br />
generation is 150. Also, the learning factors 1 c , 2 c is 2<br />
and weight parameter w is 0.3 in CCPSO. Catastrophe<br />
factor is 0.35 and the allowed catastrophe happen is 3<br />
times.<br />
In algorithm, two sub-swarms are given to conduct<br />
parallel evolutions, and then several experiments with<br />
different values <strong>of</strong> Γ are undertaken.Fig.2 and 3 are<br />
results for two sub-swarms when Γ= 0.3 .<br />
View from Fig. 2 and Fig. 3, two curve lines are<br />
depicted, in which the real-line represents the best<br />
solution is got by each generation and the broken one is<br />
the average objective solution <strong>of</strong> each generation. With<br />
the evolution <strong>of</strong> algorithm, the two lines travel towards<br />
the optimum point. It indicates that the novel algorithm<br />
CCPSO has a good convergence and the strong robust<br />
performance.<br />
objective value<br />
objective value<br />
Fig. 2 Evolutionary curve <strong>of</strong> species one<br />
Fig. 3 Evolutionary curve <strong>of</strong> species two<br />
To find how Γ effects on CCPSO, five experiments<br />
under different value <strong>of</strong> Γ are finished in the same<br />
condition and the results are taken down in TABLE Ⅱ.<br />
From TABLE Ⅱ, the smaller Γ is, the bigger<br />
objective fitness value is. Also, when Γ is 0.1, the
1960 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
medium makespan is smaller than other cases.<br />
Meanwhile, a better scheduling scheme is obtained. On<br />
the contrary, when Γ is 0.9, the makesapn is the biggest<br />
among all results and the scheduling method is most<br />
negative. Therefore, Γ should be equaled a value<br />
ranging from 0 to 1 as small as possible. Besides, in any<br />
case, CCPSO algorithm can reach up to good effect and<br />
has a favorable convergence.<br />
The Fig. 4 is drawing for Γ = 0.3 .<br />
TABLE I.<br />
Uncertain processing time <strong>of</strong> productions<br />
Unit 1 Unit 2 Unit 3 Unit 4 Unit 5<br />
Job 1 (23 25 31) (11 15 21) (10 12 14) (34 40 46) (6 10 12)<br />
Job 2 (6 17 11) (37 41 47) (21 22 24) (28 36 40) (6 8 10)<br />
Job 3 (38 41 45) (137 155 167) (27 33 37) (111 121 141) (145 160 188)<br />
Job 4 (64 74 90) (8 12 16) (16 24 30) (40 48 58) (66 78 86)<br />
Job 5 (6 7 9) (69 95 107) (62 72 84) (51 52 56) (148 153 179)<br />
Job 6 (10 12 16) (8 14 16) (58 62 74) (26 32 38) (140 162 190)<br />
Job 7 (9 11 17) (5 7 12) (23 31 35) (20 26 30) (26 32 38)<br />
Job 8 (25 31 39) (35 39 43) (135 141 175) (4 6 10) (15 19 23)<br />
Job 9 (24 32 34) (84 92 98) (10 12 14) (8 14 18) (84 102 122)<br />
Job 10 (19 27 31) (109 114 128) (17 21 23) (78 90 102) (44 52 66)<br />
© 2011 ACADEMY PUBLISHER<br />
TABLE II<br />
Scheduling results under different Γ conditions<br />
Γ Job List Target Value Makespan L Makespan M Makespan U<br />
0.1 7 6 8 4 5 9 1 3 2 10 0.9950 792 887 1064<br />
0.3 7 1 6 5 4 3 8 9 2 10 0.9794 785 893 1061<br />
0.5 7 1 6 5 8 4 9 3 10 2 0.9769 785 893 1061<br />
0.7 7 2 6 4 5 8 3 10 9 1 0.9178 805 911 1082<br />
0.9 1 6 7 4 2 9 5 8 3 10 0.8438 828 941 1113<br />
Fig. 4 Gantt <strong>of</strong> the scheduling with optimal solution
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1961<br />
V. CONCLUSION<br />
In this paper, Flow Shop production scheduling<br />
problem with uncertain processing time and infinite<br />
intermediate storage is researched on the basis <strong>of</strong> actual<br />
scheduling problem. The scheduling model is set up<br />
based on the theory <strong>of</strong> fuzzy programming, in which<br />
fuzzy processing is considered. The maximum<br />
membership function <strong>of</strong> mean value has been applied to<br />
solve the non-linear fuzzy scheduling model in order to<br />
transform the fuzzy optimization problem to the general<br />
optimization problem. Also, by simulating the<br />
phenomenon <strong>of</strong> the nature, Cooperative Particle Swarm<br />
Optimization with Catastrophe (CCPSO) is proposed, to<br />
which a cooperative and catastrophe operation is added.<br />
Finally, the novel algorithm is adopted to verify the<br />
model and satisfactory results are obtained.<br />
In my future work, this algorithm can be used to<br />
optimize more complex scheduling problem like<br />
multi-objectives and scheduling with uncertainty.<br />
Meanwhile, some other strategies can be taken into<br />
consideration to advance the algorithm so that<br />
performance can be enhanced.<br />
ACKNOWLEDGMENT<br />
This work was supported in part by Shanghai Municipal<br />
Science and Technology Commission. (Grant No.<br />
10JC1405800), Project <strong>of</strong> Science and Technology<br />
Commission <strong>of</strong> Shanghai Municipality(08DZ1200505),<br />
Project <strong>of</strong> Shanghai Municipal Economic and Information<br />
Commission(09A118) ,and Key Discipline <strong>of</strong> Shanghai<br />
Municipal Education Commission(J51901).<br />
REFERENCES<br />
[1]Johnson SM. Optimal two- and three-stage production<br />
schedules with setup times included. Novel Research<br />
Logistics Quarterly 1954;1:61–8.<br />
[2] B.J.Lagewag, J.K.LEnstra and A.H.G.Rinnooy<br />
Kan, Job shop scheduling by implicit enumeration.<br />
Management Science, 1977, 24:441-450.<br />
[3] Prade H.Usmg Fuzzy Set Theory in a Scheduling Problem<br />
:a Case Study [J] .Fuzzy Sets and Systems, 1979,<br />
2(2):153-165.<br />
[4] Chanas S,Kasperski A.Minimizing Maximum Lateness in<br />
a Single Machine Sched uling Problem with Fuzzy<br />
Processing Times and Fuzzy Due Dates [J] .<br />
Engineering Applications <strong>of</strong> Artificial<br />
© 2011 ACADEMY PUBLISHER<br />
Intelligence,2001,14(3):377-386.<br />
[5] Fortemps P.Job shop Scheduling with Imprecise Duration:<br />
Fuzzy Approach [J] . IEEE, Transactions on Fuzzy<br />
Systems, 1997, 5(4): 557-569.<br />
[6] McCahon S, Lee E S. Job Sequencing with Fuzzy<br />
Processing Times [J]. <strong>Computers</strong> and Mathematics with<br />
Applications, 1990, 19(7):31-41.<br />
[7] Litoiu M, Tadei R. Real-time Task Scheduling with Fuzzy<br />
Deadlines and Processing Times [J]. Fuzzy Sets and<br />
Systems, 2001, 117(1):35-45.<br />
[8] TzungPei Hong,TzuTing Wang. Fuzzy Flexible Flow Shops<br />
at Two M achine Centers for Continuous Fuzzy Domains<br />
[J]. Information Sciences, 2000, 129(1-4):227-237.<br />
[9] Wu Chaochao, Gu Xingsheng. A Genetic Algorithm for<br />
Single Machine Scheduling with Fuzzy Processing Time<br />
and Multiple Objectives [J]. <strong>Journal</strong> <strong>of</strong> Donghua<br />
University, 2004, 21(3):185-189.<br />
[10] NIU Qun , GU Xing-sheng. A Novel Particle Swarm<br />
Optimization for Flow Shop Scheduling with Fuzzy<br />
Processing Time [J]. <strong>Journal</strong> <strong>of</strong> Donghua University,<br />
2008, 25(2):115-122<br />
[11] XU Zhen-hao, GU Xing-sheng, Earliness and tardiness<br />
flow shop scheduling problems under uncertainty with<br />
finite intermediate storage[J]. Control Theory &<br />
Applications, 2006, 23(3): 480-486.<br />
[12] M.A. Potter, The design and analysis <strong>of</strong> a computational<br />
model <strong>of</strong> cooperative co-evolution, Ph.D. Thesis, George<br />
Mason University, 1997.<br />
[13] G. Venter, R.T. Haftka, A two species genetic algorithm<br />
for designing composite laminates subjected to<br />
uncertainty, in: Proceedings <strong>of</strong> 37th<br />
AIAA/ASME/ASCE/AHS/ASC Structures, Structural<br />
Dynamics, and Materials Conference, 1996,<br />
pp.1848–1857.<br />
[14] M.A. Potter, K.A. De Jong, Cooperative co-evolution: An<br />
architecture for evolving coadapted subcomponents,<br />
Evolutionary Computation 8 (1) (2000) 1–29.<br />
[15] C.D. Rosin, R.K. Belew, New methods for competitive<br />
co-evolution, Evolutionary Computation 5 (1) (1997)<br />
1–29.<br />
[16] J Kennedy, R Eberhart.: Particle Swarm<br />
Optimization[C].In: Proc IEEE Int Conf on Neural<br />
Network (1995):1942-1948.<br />
[17] V. Ahmadjian, S. Paracer, Symbiosis: An Introduction to<br />
Biological Associations, Oxford University Press, New<br />
York, 2000.<br />
[18] A.E. Douglas, Symbiotic Interactions, Oxford University<br />
Press, Oxford, 1994.<br />
[19] Portter M A.: The design and analysis <strong>of</strong> a computational<br />
model <strong>of</strong> cooperative co-evolution [D]. Washington DC:<br />
George Mason University (1997).
1962 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
A Double Margin Based Fuzzy Support Vector<br />
Machine Algorithm<br />
Kai Li<br />
School <strong>of</strong> Mathematics and Computer Science, Hebei University, Baoding, China<br />
Email: likai_njtu@163.com<br />
Xiaoxia Lu<br />
School <strong>of</strong> Mathematics and Computer Science, Hebei University, Baoding, China<br />
Email: yingli453@sina.com.cn<br />
Abstract—Although fuzzy support vector machine<br />
introduces the fuzzy membership degree in maximizing the<br />
margin and improves performance <strong>of</strong> classifier, it has not<br />
fully considered the position <strong>of</strong> training samples in the<br />
margin. In this paper, a double margin (rough margin)<br />
based fuzzy support vector machine (RFSVM) algorithm is<br />
presented by introducing rough set into fuzzy support<br />
vector machine. Firstly, we compute the degree <strong>of</strong> fuzzy<br />
membership <strong>of</strong> each training sample. Secondly, the data<br />
with fuzzy memberships are trained to obtain the decision<br />
hyperplane that maximizing rough margin method which<br />
contains the lower margin and the upper margin. In this<br />
algorithm, points in the lower margin have major penalty<br />
than those in the boundary in the rough margin. Finally,<br />
experiments on several benchmark datasets show that the<br />
RFSVM algorithm is very effective and feasible relative to<br />
the existing support vector machines.<br />
Index Terms—fuzzy support vector machine, double margin,<br />
classification, accuracy<br />
I. INTRODUCTION<br />
Support vector machine is firstly proposed by Vapnik<br />
et al for binary-class classification problem in 1995[1] [2]<br />
[3]. It has superior performance than traditional learning<br />
algorithms and has drawn the concern <strong>of</strong> many scholars<br />
in recent years. Support vector machine is based on<br />
Statistical Learning Theory (SLT) on (VC) dimension [4]<br />
deciding a confidence interval term and structural risk<br />
minimization (SRM) principle minimizing the upper<br />
bound <strong>of</strong> the generalization error. Support vector machine<br />
introduces a kernel tick to deal with non-separable<br />
problem. It maps points in the input space into a higherdimensional<br />
feature space such that the binary-class<br />
classification problem are indeed linearly separable or<br />
linearly approximately separable through a nonlinear map,<br />
and then finds an optimal separating hyperplane that<br />
maximizes the margin between two classes in the highdimensional<br />
feature space. However, there are still have<br />
Manuscript received October 1, 2010; revised December 1, 2010;<br />
accepted January 1, 2011.<br />
Corresponding author. Tel.:+86 0312 5079660<br />
Email: likai_njtu@163.com, yingli453@sina.com.cn<br />
© 2011 ACADEMY PUBLISHER<br />
doi:10.4304/jcp.6.9.1962-1970<br />
two questions needed to be further study which are how<br />
to effectively expand the binary-class classification<br />
problem to multiclass classification problem and how to<br />
overcome sensitivity or overfitting due to noises and<br />
outliers in optimal hyperplane.<br />
About the first problem, many scholars expand binaryclass<br />
classification to multiclass classification problem,<br />
wherein one-against-one (1-a-1) and one-against-all (1-ar)<br />
are common methods which transform multiclass<br />
classification problem into binary-class classification<br />
problem. Hsu and Lin studied a comparison <strong>of</strong> methods<br />
for multiclass support vector machines such as oneagainst-all,<br />
one-against-one, directed acyclic graph SVM<br />
(DAGSVM) [5]. To deal with unclassifiable region,<br />
Inoue and Abe proposed fuzzy support vector machine<br />
for multiclass problem [6]. This method uses fuzzy<br />
membership to resolve unclassifiable regions. In Ref. [7],<br />
the authors proposed a new fuzzy membership function in<br />
the nonlinear fuzzy support vector machine. Moreover,<br />
Yan and He propose a new method—multiclass fuzzy<br />
support vector machine <strong>of</strong> dismissing margin (DFSVM)<br />
based on class-center [8].<br />
To the second problem, many scholars put forward a<br />
lot <strong>of</strong> variant SVM. In traditional support vector machine,<br />
each input point is fully assigned to one <strong>of</strong> two classes<br />
wherein some noises and outliers are ignored in training<br />
set. Therefore, it results in overfitting problem to some<br />
extent. In fact, only few input points can decide the<br />
hyperplane. In more and more real-world applications,<br />
the effects <strong>of</strong> the training points, especially noises and<br />
outliers, are different. Aimed at these problems, Lin and<br />
Wang introduced fuzzy set theory into support vector<br />
machine to overcome the sensitivity <strong>of</strong> noises and outliers<br />
to optimal hyperplane, called fuzzy support vector<br />
machine (FSVM) [9]. Fuzzy support vector machine<br />
associates a fuzzy membership with each input point such<br />
that different examples make different contributions to<br />
the learning <strong>of</strong> optimal surface. Other scholars combined<br />
FSVM with genetic algorithms (GA) [10] to improve the<br />
generalization performance <strong>of</strong> SVM. However, these<br />
need a prior knowledge <strong>of</strong> the distribution <strong>of</strong> training set.<br />
Wu and Law proposed fuzzy support vector regression
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1963<br />
machine with Gaussian noises on triangular fuzzy number<br />
space to forecast fuzzy nonlinear system [11].<br />
The rough set theory [12] is a powerful preprocessing<br />
tool to find out knowledge from an amount <strong>of</strong> uncertain<br />
and incomplete data and is applied to the support vector<br />
machines to reduce the features <strong>of</strong> data to process and<br />
eliminate redundancy. At the same time, it also improves<br />
performance <strong>of</strong> the classical support vector machines. To<br />
deal with the overfitting problem <strong>of</strong> the traditional<br />
support vector machine, Zhang and Wang proposed a<br />
rough margin based support vector machine [13]. In this<br />
paper, we propose a double margin based fuzzy support<br />
vector machine by combination <strong>of</strong> rough theory and<br />
fuzzy support vector machine, namely a double margin<br />
(rough margin) based on fuzzy support vector machine<br />
(RFSVM). The proposed method not only inherits the<br />
characteristic <strong>of</strong> the FSVM method, but also considers the<br />
effects <strong>of</strong> decision hyperplane depending on the position<br />
<strong>of</strong> training samples in the rough margin. So presented<br />
method further reduce overfitting due to noises or outliers.<br />
This paper is organized as follows. In section 2, a brief<br />
review <strong>of</strong> support vector machine is described. In Section<br />
3, we describe the proposed RFSVM in detail which<br />
contain both binary classification and multiple<br />
classification RFSVM. In the following section, we<br />
evaluate our method on benchmark data sets and compare<br />
it with the existing support vector machine. Some<br />
conclusions are given in the final section.<br />
II. SUPPORT VECTOR MACHINES ALGORITHM<br />
In this section, we briefly describe the support vector<br />
machines in binary classification problems.<br />
Given a dataset <strong>of</strong> labeled training points (x1, y1), (x2,<br />
y2),…, (xl, yl), where N<br />
( xi, yi) ⊆ R × { + 1, − 1} , i=1, 2…l.<br />
Supposed training data are linearly separable. That is to<br />
say, there is some hyperplane which correctly separates<br />
the positive examples and negative examples. The point x<br />
lying on the hyperplane satisfies +b = 0, where w<br />
is normal to the hyperplane. In this case, support vector<br />
machine algorithm finds the optimal separating<br />
hyperplane with the maximal margin. When the training<br />
data are linearly non-separable or approximately<br />
separable, it is needed to introduce the trade-<strong>of</strong>f<br />
parameter. When the training data is not linearly<br />
separable, support vector machine learning algorithm<br />
introduces kernel strategy that maps the input data to a<br />
higher-dimension feature space z by using a nonlinearly<br />
mapping function ϕ () x and then the data in feature space<br />
z is indeed linearly or approximately separable. All<br />
training data satisfy the following decision function<br />
⎧+<br />
1, if yi<br />
=+ 1<br />
f( xi) = sign( < w, x >+ b)<br />
= ⎨<br />
. (1)<br />
⎩ − 1, if yi<br />
= -1<br />
All training points satisfy the following inequalities:<br />
⎧<<br />
wx , i > + b≥ + 1, ifyi = + 1 . (2)<br />
⎨<br />
⎩ < wx , i > + b≤ − 1, ifyi = -1<br />
In fact, it can be written as yi( < w, xi > + b)<br />
≥ 1,<br />
i=1,2,…,l. above inequalities. It is seen that finding the<br />
hyperplane is equivalent to obtain the maximizing margin<br />
© 2011 ACADEMY PUBLISHER<br />
2<br />
by minimizing || w || subject to constraints (2). So the<br />
primal optimal problem is given as<br />
1 2<br />
min || w ||<br />
wb , 2<br />
st .. y( < w, x >+ b)<br />
≥1.<br />
(3)<br />
i i<br />
i = 1, 2,..., l<br />
To solve optimal problem, we introduce Lagrange<br />
multiplier to transform the primal problem (3) into its<br />
dual problem that becomes the following quadratic<br />
programming (QP) problem:<br />
l l l<br />
1<br />
min ∑∑αiα jyy i j( xi⋅xj) −∑αi<br />
α 2 i= 1 j= 1 i=<br />
1 . (4)<br />
l<br />
∑<br />
s. t. α y = 0, 0 ≤ α , i= 1,2,..., l.<br />
i=<br />
1<br />
i i i<br />
In classifier, the solution in feature space using a<br />
linearly mapping function ϕ ( x)<br />
only replaces the dot<br />
product x ⋅ x j by inner product vectors ϕ( x) ⋅ ϕ(<br />
x j ) . The<br />
mapping function ϕ( x)<br />
and ϕ ( xi<br />
) satisfy<br />
< ϕ( x), ϕ(<br />
xj) >= K( x, xi)<br />
, where K( x, xi) is called kernel<br />
function. In real world application, we would never need<br />
to explicitly know what ϕ is. A decision function with<br />
SVM is obtained by computing dot products <strong>of</strong> a given<br />
test point x, or more specifically by computing following<br />
sign:<br />
Ns<br />
*<br />
f( x) = α y ( s ⋅ x) + b<br />
∑<br />
i=<br />
1<br />
i i i<br />
Ns<br />
*<br />
∑α<br />
i<br />
i=<br />
1<br />
iϕ i ϕ<br />
Ns<br />
*<br />
= ∑α<br />
i<br />
i=<br />
1<br />
i ( i,<br />
) +<br />
. (5)<br />
= y ( s ) ⋅ ( x) + b<br />
yK s x b<br />
Where the coefficient α is positive, i<br />
i<br />
s is support vector<br />
and Ns is the number <strong>of</strong> support vectors.<br />
In most cases, as the learning <strong>of</strong> a suitable hyperplane<br />
is too restrictive to be <strong>of</strong> practical use and causes a large<br />
overlap <strong>of</strong> classes, there is nonexistent some separable<br />
hyperplane. To deal with linearly non-separable data, it<br />
<strong>of</strong>ten allows that some points are misclassified, and<br />
introduces nonnegative slack variables ξ > 0 measuring<br />
the number <strong>of</strong> misclassifications and a punishment<br />
parameter C which is a cost trade-<strong>of</strong>f between<br />
maximizing the margin and minimizing the classification<br />
error <strong>of</strong> training data. The sum <strong>of</strong> the slacks ∑ ξ is an<br />
i<br />
upper bound on the number <strong>of</strong> training errors. And, the<br />
original constraints (2) are relaxed to<br />
yi( < w, xi>+ b) ≥1 − ξi,<br />
i = 1,2,..., l.<br />
(6)<br />
Thus, constructing optimal hyperplane is equivalent to<br />
solve the following optimization problem:<br />
l<br />
1 2<br />
min || w|| + C∑ξi<br />
wb , , ξ 2<br />
i=<br />
1<br />
st .. yi( < w, ϕ( xi) >+ b)<br />
≥1−ξ (7)<br />
i<br />
ξ ≥ 0, i = 1, 2,..., l.<br />
i
1964 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
The corresponding dual problem is as following:<br />
l l l<br />
1<br />
min ∑∑αiαjyyK i j ( xi⋅xj) −∑αi<br />
α 2 i= 1 j= 1 i=<br />
1 . (8)<br />
l<br />
∑<br />
s. t. α y = 0, 0 ≤α ≤ C, i= 1, 2,..., l<br />
i=<br />
1<br />
i i i<br />
To overcome the difficulty <strong>of</strong> the value <strong>of</strong> chosen<br />
parameter C, an alternative classifier model, called ν-<br />
SVM was proposed and developed [14] [15] [16]. In this<br />
model, C is replaced by a more meaningful parameter<br />
ν ∈ (0,1) , which is the lower and upper bound on the<br />
number <strong>of</strong> training points that are support vectors and that<br />
lies on the wrong side <strong>of</strong> the hyperplane, respectively.<br />
The primal optimization problem <strong>of</strong> this algorithm<br />
becomes<br />
l<br />
1 2 1<br />
min || w || − νρ + ∑ ξi<br />
wb , , ξ , ρ 2<br />
l i=<br />
1<br />
st .. yi( < w, ϕ ( xi) > + b)<br />
≥ ρ −ξi<br />
, (9)<br />
ξi≥ 0, ρ ≥ 0, i = 1, 2,..., l<br />
where variables wbξ ,,, ρ are optimized.<br />
The dual problem <strong>of</strong> this primal optimal problem can<br />
be solved by the following quadratic optimization<br />
problem:<br />
l l 1<br />
min ∑∑αα<br />
i jyyK i j ( xi, xj)<br />
α 2 i= 1 j=<br />
1<br />
l<br />
1 . (10)<br />
st .. ∑αiyi<br />
= 0, 0 ≤αi ≤ ,<br />
l<br />
i=<br />
1<br />
l<br />
∑<br />
i=<br />
1<br />
α ≥ vi , =1,2,..., l.<br />
i<br />
In many practices, training points are not fully<br />
assigned to one class <strong>of</strong> two classes, so Lin and Wang<br />
proposed fuzzy support vector machine (FSVM) [8].<br />
Given labeled each training point associate with a fuzzy<br />
membership, namely (x1,y1,s1),(x2,y2,s2),…,(xl,yl,sl), where<br />
N<br />
( x, y) ⊆ R × { + 1, − 1} , i = 1, 2,..., l and si ( 0 < s ≤ 1 ) is<br />
i i<br />
fuzzy membership corresponding to each training point,<br />
the parameter ξ is the measure <strong>of</strong> misclassification, the<br />
i<br />
term siξ is a measure <strong>of</strong> error with different weight. It is<br />
i<br />
equivalent to solve the following optimal problem:<br />
l<br />
1 2<br />
min || w|| + C∑siξi wb , , ξ 2<br />
i=<br />
1<br />
st .. yi( < w, ϕ ( xi) > + b)<br />
≥1 −ξi<br />
, (11)<br />
ξi<br />
≥ 0, i = 1, 2,..., l<br />
where C is a constant.<br />
Finding the optimal hyperplane can be solved by<br />
constructing a Lagrange function and transformed the<br />
primal problem into the following dual problem:<br />
l l l 1<br />
max ∑αi − ∑∑αα<br />
i jyyK i j ( xi, xj)<br />
α<br />
i= 1 2 i= 1 j=<br />
1<br />
, (12)<br />
l<br />
∑<br />
s. t. α y = 0, 0 ≤α ≤ sC, i= 1,2,..., l<br />
i=<br />
1<br />
i i i i<br />
© 2011 ACADEMY PUBLISHER<br />
i<br />
where α is the nonnegative Lagrange multiplier<br />
i<br />
associated with the inequality constraint. The points<br />
corresponding to α i > 0 are called support vectors.<br />
From the Kuhn-Tucker conditions, we can obtain<br />
αi( yi( < w, xi >+ b)<br />
− 1 + ξi)<br />
= 0<br />
. (13)<br />
(siC- αi) ξi<br />
= 0, i = 1, 2,..., l<br />
To deal with overfitting problem due to noises or<br />
outliers in support vector machine, Zhang and Wang<br />
proposed a rough margin based support vector machine<br />
(RMSVM) [13]. In this paper, they considered the<br />
training points with different effects on the learning <strong>of</strong> the<br />
separating hyperplane depending on their positions in the<br />
rough margin. It searches the optimal separating<br />
hyperplane that maximizes the rough margin which<br />
contains the lower and upper margin. The primal<br />
optimization problem <strong>of</strong> RMSVM can be defined as<br />
l l<br />
1 2 1 δ '<br />
min || w||<br />
−νρl − νρu + ∑ξi + ∑ξ<br />
wb , , ξξ , ', ρ, i<br />
l ρu<br />
2<br />
l i= 1 l i=<br />
1<br />
'<br />
st .. y( < w, ϕ( x) >+ b)<br />
≥ρ −ξ −ξ<br />
, (14)<br />
i i u i i<br />
ξi ρu '<br />
ρl ξi ρl ρu<br />
0 ≤ ≤ − , ≥0, ≥0, ≥0.<br />
ρ ρ<br />
whereσ > 1 , l<br />
u and ( ρu > ρ ) are the width <strong>of</strong><br />
l<br />
|| w || || w ||<br />
the lower margin and upper margin, respectively.<br />
This primal optimal problem can be solved by its dual<br />
problem as follows<br />
l l 1<br />
min ∑∑αα<br />
i jyyK i j ( xi, x j)<br />
α 2 i= 1 j=<br />
1<br />
l<br />
δ . (15)<br />
st .. ∑ αiyi = 0,0 ≤αi ≤<br />
l<br />
i=<br />
1<br />
l<br />
∑<br />
i = 1<br />
α ≥ 2v<br />
i<br />
Ⅲ. DOUBLE MARGIN BASED FUZZY SUPPORT MACHINE<br />
A. Binary Classification Case<br />
Aimed at fuzzy support vector machine, to further<br />
overcome the overfitting problem and to reduce the<br />
effects <strong>of</strong> outliers or noises, we propose a double margin<br />
(rough margin) based fuzzy support vector machine<br />
(RFSVM), in which it not only associates a fuzzy<br />
membership with each training point, but also considers<br />
each training example’s position in the rough margin.<br />
According to rough theory, rough margin contains lower<br />
margin 2ρ<br />
l and upper margin 2ρ<br />
u ( ρu > ρ ). l<br />
|| w ||<br />
|| w ||<br />
The region <strong>of</strong> training points within the lower margins is<br />
equivalent to positive region in rough theory; the data in<br />
this region are noises or outliers. The regions <strong>of</strong> the<br />
training examples within the upper margins and outside<br />
the lower margins are equivalent to boundary regions. In<br />
addition, the data outside the upper margins<br />
corresponding to the negative regions are correctly<br />
classified and are non-support vectors that are not noises<br />
and outliers. Training examples in the lower margin have
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1965<br />
major penalty but those in the boundary region <strong>of</strong> the<br />
rough margin have minor penalty to hyperplane. RFSVM<br />
finds the maximum rough margin in some highdimensional<br />
feature space.<br />
Similar to classical support vector machine, the primal<br />
problem <strong>of</strong> RFSVM can be described as:<br />
l l<br />
1 2 1 δ '<br />
min || w|| −νρl − νρu + ∑siξi + ∑siξ<br />
wb , , ξξ , ', ρ, i<br />
l ρu<br />
2<br />
l i= 1 l i=<br />
1<br />
'<br />
st .. y( < w, ϕ( x) >+ b)<br />
≥ρ −ξ −ξ<br />
. (20)<br />
i i u i i<br />
'<br />
0 ≤ξi ≤ρu −ρl, ξ ≥0, ρ 0, 0<br />
i l ≥ ρu<br />
≥<br />
Here, we set δ ≥ 1 and ν ∈ (0,1) . ρ l and ρ u construct<br />
the inner and the outer wall <strong>of</strong> the boundary regions,<br />
respectively. When δ = 1 , RFSVM is equivalent to<br />
contain the parameter ν <strong>of</strong> fuzzy support vector machine<br />
without ρ u andδ . When the data points locate on the<br />
positive region, those are regarded as outliers and noises.<br />
When training points lie on the negative regions, they are<br />
regard as non-support vector and are not noises and<br />
outliers. The slack variable '<br />
ξ and ξ is introduced by<br />
datum locating in the positive region and boundary<br />
regions <strong>of</strong> rough margin, respectively.<br />
To solve this optimization problem, we construct the<br />
Lagrange function:<br />
l l<br />
1 2 1 δ '<br />
L= || w|| −νρl− νρu+ ∑siξi+ ∑siξi<br />
2<br />
l l<br />
i= 1 i=<br />
1<br />
l<br />
'<br />
− ∑αi(<br />
yi( < w, ϕ( xi) >+ b)<br />
− ρu+ ξi+ ξ)<br />
i<br />
i=<br />
1<br />
l l l<br />
'<br />
−∑βξ i i−∑λi( ρu−ρl−ξi) −∑ηξ i −θρ i 1 l−θ2ρu i= 1 i= 1 i=<br />
1<br />
where i, i, i, i,<br />
1, 2 0<br />
, (21)<br />
α β λ η θ θ ≥ are Lagrange multipliers.<br />
According to KKT conditions, the parameters satisfy the<br />
following conditions:<br />
l ∂ L<br />
= w− ∑ αiyiϕ( xi)<br />
= 0<br />
∂w<br />
i=<br />
1<br />
l ∂ L<br />
=− ∑ α iyi = 0<br />
∂b<br />
i=<br />
1<br />
∂ L si<br />
= −αi − βi + λi<br />
= 0<br />
∂ξi<br />
l<br />
∂L<br />
δ si<br />
= −α 0<br />
'<br />
i − ηi<br />
=<br />
∂ξ<br />
l<br />
i<br />
l<br />
∂L<br />
=− v + ∑ λi− θ1=<br />
0<br />
∂ρ<br />
l<br />
i = 1<br />
l l<br />
∂L<br />
=− v + ∑α i −∑ λi − θ2=<br />
0<br />
∂ρ<br />
u<br />
i= 1 i=<br />
1<br />
'<br />
α i( yi( < w, ϕ ( xi) > + b)<br />
− ρ u + ξ i + ξ ) = 0<br />
i<br />
βξ i i = 0<br />
. (22)<br />
λi( ρ u − ρ l − ξ i)<br />
= 0<br />
'<br />
ηξ i = 0, θ1ρ0, 2 0<br />
i<br />
l = θ ρ u =<br />
Applying these equations into the Lagrange function<br />
(21), the primal problem (20) can be transformed into the<br />
Wolf dual problem:<br />
© 2011 ACADEMY PUBLISHER<br />
l l 1<br />
min ∑∑αα<br />
i jyyK i j ( xi, xj)<br />
α 2 i= 1 j=<br />
1<br />
st ..<br />
l<br />
∑αiyi<br />
= 0, 0 ≤αi δ si<br />
≤<br />
l<br />
. (23)<br />
i=<br />
1<br />
l<br />
∑<br />
i=<br />
1<br />
α ≥ 2v<br />
i<br />
From the Kuhn-Tucker conditions, we obtain<br />
'<br />
αi( yi( < w, ϕ( xi) >+ b)<br />
− ρu + ξi + ξ ) = 0<br />
i<br />
si<br />
( − αi + λi) ξi<br />
= 0<br />
l<br />
δ si<br />
'<br />
( − αi) ξi<br />
= 0<br />
l<br />
. (24)<br />
The point xi with the corresponding to α i = 0 satisfies<br />
yi( < w, ϕ( xi) >+ b)<br />
> ρ lying outside the upper margin<br />
u<br />
<strong>of</strong> rough margin. The point xi with the corresponding to<br />
α i > 0 is called support vector. When<br />
si<br />
0 < α i <<br />
l<br />
, the<br />
one lying on the border <strong>of</strong> the upper margin <strong>of</strong> the<br />
hyperplane satisfies yi( < w, ϕ( xi) >+ b)<br />
= ρu.<br />
When<br />
si<br />
α i = , the one lying within the boundary <strong>of</strong> the rough<br />
l<br />
margin,<br />
satisfies yi( < w, ϕ( xi) >+ b)<br />
= ρu−ξi, where ξ i > 0.<br />
When si δ si<br />
< α i < , the one lying on the boundary <strong>of</strong> the<br />
l l<br />
lower margin satisfies yi( < w, ϕ( xi) >+ b)<br />
= ρl.<br />
When<br />
δ si<br />
α i =<br />
l<br />
, the one lying within the lower margin, is<br />
'<br />
misclassified and satisfies yi( < w, ϕ( xi) >+ b)<br />
= ρu − ξ .<br />
i<br />
From the optimal values α i ( i = 1, 2,..., l)<br />
<strong>of</strong> (21), we<br />
can obtain the decision function <strong>of</strong> RFSVM:<br />
f ( x) = sgn( α K( x, x ) + b)<br />
, (25)<br />
∑<br />
i= RSV<br />
i i<br />
where RSV denotes the index set <strong>of</strong> data with<br />
α > 0 , 1<br />
b =− α iyi( K( xi, x j) + K( xi, xk))<br />
,where<br />
∑ 2 i= RSV<br />
s j<br />
i∈{ j| α j ∈ (0, ), y j = 1} and<br />
l<br />
s j<br />
k ∈ { j | α j ∈ (0, ), y j = − 1} , or<br />
l<br />
sj δ sj<br />
i∈{ j| α j ∈ ( , ), y j = 1} and<br />
l l<br />
sj δ sj<br />
k∈{ j| α j ∈ ( , ), yj<br />
= − 1} .<br />
l l<br />
The design <strong>of</strong> fuzzy membership function is the key to<br />
the fuzzy algorithm using fuzzy technology. In this paper,<br />
we use class center method to generate fuzzy membership.<br />
Firstly, we denote the mean <strong>of</strong> class +1 as classcenter<br />
x + , and the mean <strong>of</strong> class -1 as class center x − .<br />
The farthest distance between the each class training
1966 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
points and its class-center, the radius <strong>of</strong> class are<br />
r = max || x − x || and r = max || x − x || ,<br />
+ +<br />
{ xi: yi=<br />
1}<br />
i<br />
− −<br />
{ xi: yi=−1}<br />
respectively.<br />
Fuzzy membership s i is a function <strong>of</strong> the mean and<br />
radius <strong>of</strong> each class<br />
⎧1<br />
−|| x+ − xi ||/( r+ + σ ), if yi<br />
=+ 1<br />
si<br />
= ⎨<br />
(26)<br />
⎩ 1 −|| x- − xi ||/( r- + σ ), if yi<br />
= -1<br />
where σ > 0 is used to avoid the case s = 0 .<br />
B. Multiple Classification Case<br />
In this section, we extend binary-class classification <strong>of</strong><br />
rough margin to multi-class classification and implement<br />
it on one-against-all and one-against-one methods in<br />
detail.<br />
The one-against-all method constructs p RFSVM,<br />
where p is the number <strong>of</strong> classes. The ith RFSVM is<br />
trained with all training point in the ith class with positive<br />
class, and all other training points are considered as<br />
negative class; at the same time, we computed fuzzy<br />
membership <strong>of</strong> each training point and the position <strong>of</strong> the<br />
rough margin <strong>of</strong> training set. p - RFSVM algorithm<br />
obtains linear decision function<br />
i i<br />
f ( x, y) =sgn(g ( x, y))<br />
wb ,<br />
i i<br />
=sgn(< w , ϕ(<br />
x, y)>+ b )<br />
i<br />
class <strong>of</strong> x = max ( g ( x, y)).<br />
i= 1,2,, p<br />
i<br />
When the value <strong>of</strong> g ( xy , ), i= 1, 2,..., l is equivalent<br />
or very little different, unclassifiable region exists. We<br />
use rough margin and fuzzy membership functions to<br />
resolve unclassifiable regions to realize the same<br />
classification results with that <strong>of</strong> traditional multiclass<br />
support vector machine and multiclass fuzzy support<br />
vector machine.<br />
The one-against-one method constructs p(p-1)/2<br />
RFSVM where p is the number <strong>of</strong> classes. The ith<br />
RFSVM is trained with the training point in the ith class<br />
with positive class and the jth class with negative class<br />
( i≠ j ). At the same time, we computed fuzzy<br />
membership <strong>of</strong> each training point and according to each<br />
example’s position <strong>of</strong> training set in the rough margin to<br />
further reduce the effects <strong>of</strong> the outliers and noises. The<br />
decision function for class i against class j, with the<br />
ij<br />
maximum margin, is fw, b () x =< wij , ϕ()>+ x b where ij<br />
ij w<br />
is the l-dimensional vector, bij is a scalar and<br />
ij ji<br />
f ()=- x f (). x For test vector x, we calculate<br />
wb , wb ,<br />
p<br />
g ()= x sgn( f ()), x class <strong>of</strong> x =<br />
∑<br />
i ij<br />
wb ,<br />
j= 1, j≠i i<br />
i<br />
max ( g ( x)).<br />
i= 1,2,, p<br />
Ⅳ. EXPERIMENTAL RESULTS AND ANALYSIS<br />
We conduct some experiments on benchmark datasets<br />
to test performance <strong>of</strong> RFSVM algorithm and compare it<br />
with other related approaches which include rough<br />
margin based support vector machine (RMSVM), ν-SVM,<br />
© 2011 ACADEMY PUBLISHER<br />
i<br />
fuzzy support vector machine (FSVM) and Standard<br />
SVM. Experiments are conducted on 16 different data<br />
sets from UCI [17], Statlog [18] and TKH96a [19].<br />
Details about these data sets are given in Table I. In<br />
selected data sets, the number <strong>of</strong> features has a large<br />
range. In experiments, we use randomly selected<br />
techniques to evaluate the performance <strong>of</strong> an algorithm.<br />
In random selecting approach, dataset is divided into two<br />
parts: training and testing set.<br />
For each dataset <strong>of</strong> all experiments, the experiments<br />
are repeated ten times using randomly selected training<br />
and testing sets (70% <strong>of</strong> the examples for training and<br />
30% for testing) from each dataset. At the same time, we<br />
compute the predicted accuracy <strong>of</strong> each testing set every<br />
time. The parameter C is fixed on 100 and 10. For each<br />
dataset, the best parameter value ν is used for training. In<br />
most cases, the selected optimal parameter ν <strong>of</strong> ν-SVM<br />
was between 0.3 and 0.6. For RFSVM, the ν value is also<br />
within the range <strong>of</strong> 0.3-0.6 and δ is within 3.0-15. In our<br />
experiments, RFSVM and RMSVM use exactly the same<br />
parameter values on each dataset. In the experiments, we<br />
2<br />
use Gaussian kernel, Kxx (, ') = exp( −γ|| x− x'||),<br />
where<br />
γ = 1.0 .<br />
The experimental results are shown in Table II and Fig.<br />
1. The average classification accuracies <strong>of</strong> each algorithm<br />
are presented in Table II. The best result in each dataset<br />
using different algorithm is shown in boldface. It is seen<br />
by Table II that RFSVM outperforms the other support<br />
vector machine learning algorithm in most cases. In<br />
addition, RFSVM usually improves the classification<br />
results <strong>of</strong> the fuzzy support vector machine. In some<br />
cases, the improvement is very large such as Fourclass,<br />
German and Heart. Especially Australian, Diabetes,<br />
Fourclass, German, and Liver-disorders have smaller<br />
standard deviation.<br />
The experimental results in Table II demonstrate that<br />
in most cases RFSVM beats ν-SVM. Similarly, RFSVM<br />
usually outperforms RMSVM and standard SVM. These<br />
conclusions are further validated by Fig. 1. This means<br />
that for given dataset, introducing the rough margin and<br />
fuzzy membership is a good choice.<br />
TABLE I.<br />
DATASETS AND THEIR CHARACTERISTICS<br />
Dataset Data<br />
items<br />
Features Class Source<br />
Australian 695 14 2 Statlog<br />
Breast-cancer 683 10 2 UCI<br />
Bupa 345 6 2 UCI<br />
Cancer 683 7 2 UCI<br />
Diabetes 768 8 2 UCI<br />
Fourclass 862 2 2 TKH96a<br />
German 1000 24 2 Statlog<br />
Glass 214 9 6 UCI<br />
Heart 270 13 2 Statlog<br />
Iris 150 4 3 UCI<br />
Liver-disorders 345 6 2 UCI<br />
Sonar 208 60 2 UCI<br />
Splice 1000 60 2 UCI<br />
Vowel 528 10 11 UCI<br />
Wdbc 569 30 2 UCI<br />
Wine 178 13 3 UCI
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1967<br />
The classification accuracies <strong>of</strong> RFSVM, RMSVM, ν-<br />
SVM, FSVM-100 (C=100), SVM-100(C=100), FSVM-<br />
10(C=10) and SVM-10(C=10) on each dataset with 10<br />
times are shown in Fig. 2, where the digit 1-7 in the xaxis<br />
represents different classifiers respectively and the yaxis<br />
presents the classification accuracy <strong>of</strong> repeated 10<br />
times.<br />
To test whether the new proposed algorithm is superior to<br />
current algorithm, two-pairs t-test is performed among<br />
RMSVM and other algorithms which contain RMSVM,<br />
ν-SVM, FSVM-100, SVM-100, FSVM-10, and SVM-10.<br />
Results are presented in Table III. It was shown that for<br />
all datasets the differences between the results obtained<br />
by two compared classifiers were statistically significant<br />
(significance p < 0.05). From the point <strong>of</strong> view <strong>of</strong><br />
statistics, win means that RFSVM algorithm is<br />
Accuracy<br />
Accuracy<br />
significantly better than any other algorithm; when tie<br />
appear, it shows that there is no obvious difference<br />
between two algorithms; but when significance is loss,<br />
the performance <strong>of</strong> RFSVM algorithm is inferior other<br />
binary classification support vector machine algorithm.<br />
0.86<br />
0.84<br />
0.82<br />
0.8<br />
0.78<br />
0.76<br />
TABLE II.<br />
EXPERIMENTAL RESULTS WITH DIFFERENT METHOD<br />
RFSVM RMSVM ν-SVM FSVM-100 SVM-100 FSVM-10 SVM-10<br />
Figure 1. Average value <strong>of</strong> classification accuracy for all datasets.<br />
Dataset RFSVM RMSVM ν-SVM FSVM-100 SVM-100 FSVM-10 SVM-10<br />
Australian 0.8606 0.7675 0.8575 0.8282 0.8466 0.8575 0.8593<br />
Breast-cancer 0.9606 0.9446 0.9619 0.9543 0.9494 0.9559 0.9641<br />
Bupa 0.7168 0.5918 0.6897 0.6696 0.7063 0.5559 0.6678<br />
Cancer 0.9659 0.9428 0.9623 0.9636 0.9619 0.9685 0.9685<br />
Diabetes 0.7623 0.6757 0.7458 0.7528 0.7584 0.7670 0.7690<br />
Fourclass 0.9912 0.9902 0.9414 0.7996 0.8125 0.7989 0.8080<br />
German 0.7400 0.7036 0.7291 0.6894 0.7009 0.7158 0.7127<br />
Heart 0.8272 0.7778 0.8092 0.7789 0.7767 0.8081 0.7957<br />
Liver-disorders 0.7386 0.6128 0.6538 0.6976 0.7343 0.5621 0.6906<br />
Sonar 0.8773 0.8846 0.8889 0.8773 0.8846 0.8889 0.8846<br />
Splice 0.7945 0.7688 0.7688 0.7585 0.7688 0.7115 0.7688<br />
Wdbc 0.9793 0.9756 0.9596 0.9734 0.9708 0.9788 0.9777<br />
Different algorithm<br />
(1) Australian<br />
Different algorithm<br />
(5) Diabetes<br />
© 2011 ACADEMY PUBLISHER<br />
Accuracy<br />
Accuracy<br />
Different algorithm<br />
(2) Breast-cancer<br />
Different algorithm<br />
(6) Fourclass<br />
Accuracy<br />
Accuracy<br />
Different algorithm<br />
(3) Bupa<br />
Different algorithm<br />
(7) German<br />
Accuracy<br />
Accuracy<br />
Different algorithm<br />
(4) Cancer<br />
Different algorithm<br />
(8) Heart
1968 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
Accuracy<br />
Figure 2. Accuracy for different classifier with 10times for all datasets (1-RFSVM, 2-RMSVM, 3-ν-SVM, 4-FSVM-100, 5-SVM-100, 6-FSVM-10,<br />
7-SVM-10).<br />
TABLE III.<br />
EXPERIMENTAL RESULTS OF TWO-PAIRS T-TEST FOR ALL DATASETS<br />
Accuracy (70%) Win Tie Loss<br />
RFSVM RMSVM 10 2 0<br />
RFSVM ν-SVM 5 7 0<br />
RFSVM FSVM-100 8 4 0<br />
RFSVM SVM-100 2 7 0<br />
RFSVM FSVM-10 4 8 0<br />
RFSVM SVM-10 5 7 0<br />
Besides above experiments with binary classification,<br />
we also perform some experiments on multi-class<br />
datasets. RFSVM is also compared with multiclass fuzzy<br />
support vector machine (FSVM) and multiclass support<br />
vector machine (SVM) on one-against-all method and<br />
Accuracy<br />
Different algorithm<br />
(9) Liver-disorders<br />
Different algorithm<br />
(1) Glass<br />
Accuracy<br />
Different algorithm<br />
(10) Sonar<br />
Accuracy<br />
Different algorithm<br />
(11) Splice<br />
TABLE IV.<br />
EXPERIMENTAL RESULTS OF MULTI-CLASS CLASSIFICATION PROBLEM<br />
one-against-one method. We set C=100 in default. For<br />
each algorithm, we estimate the generalized accuracy<br />
using same kernel function, kernel parameters γ and cost<br />
parameters C in multiclass FSVM and multiclass SVM.<br />
Experimental results are shown in the Table IV. It can be<br />
seen that the accuracy obtained by RFSVM is same or<br />
even better compared with FSVM and SVM aimed at<br />
both the one-against-one method and one-against-all<br />
method.<br />
Similarly, we give the average accuracy and standard<br />
deviation as shown In Fig. 3 and Fig. 4. The x-axis<br />
represents the classifiers, namely RFSVM, FSVM and<br />
SVM. The y-axis represents the average accuracy and<br />
standard deviation <strong>of</strong> ten times on random selecting<br />
method. Multi-class RFSVM improves the generalization<br />
ability compared with multi-class support vector machine<br />
and fuzzy support vector machine, although it has larger<br />
standard deviation than the others.<br />
Dataset<br />
One-against-All method<br />
RFSVM FSVM SVM<br />
One-against-One method<br />
RFSVM FSVM SVM<br />
Glass 0.6358 0.6006 0.6104 0.6431 0.6155 0.6286<br />
Iris 0.9667 0.9600 0.9536 0.9697 0.9688 0.9536<br />
Vowel 0.9748 0.9740 0.9731 0.9887 0.9868 0.9851<br />
Wine 0.9668 0.9649 0.9652 0.9812 0.9706 0.9689<br />
Different algorithm<br />
(2) Iris<br />
Different algorithm<br />
(3) Vowel<br />
Figure 3. Experimental results with different algorithms on one-against-all method (1-RFSVM, 2-FSVM, 3-SVM).<br />
© 2011 ACADEMY PUBLISHER<br />
Accuracy<br />
Accuracy<br />
Accuracy<br />
Accuracy<br />
Different algorithm<br />
(12) Wdbc<br />
Different algorithm<br />
(4) Wine
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1969<br />
Accuracy<br />
Different algorithm<br />
(1) Glass<br />
Figure 4. Experimental results with different algorithms on one-against-one method (1-RFSVM, 2-FSVM,3-SVM).<br />
Ⅴ. CONCLUSIONS<br />
The support vector machine is a powerful tool for<br />
classification. However, the final decision function<br />
obtained by the support vector machine depends on few<br />
extreme value points, which makes the support vector<br />
machine sensitive to outliers or noises in the training set.<br />
In this paper, following the rough theory, we propose a<br />
double margin (rough margin) based fuzzy support vector<br />
machine that combines the notion <strong>of</strong> rough set with the<br />
fuzzy support vector machine to deal with the outlier<br />
sensitivity problem <strong>of</strong> fuzzy support vector machine, and<br />
then we design a classifier building method based on<br />
fuzzy support vector machine. The key idea <strong>of</strong> building<br />
the classifier is to find suitable fuzzy membership<br />
function and controlled parameter. This combination<br />
allows us adaptively consider more data information in<br />
the construction <strong>of</strong> the optimal hyperplane. The double<br />
margin (rough margin) based fuzzy support vector<br />
machine depends on the number <strong>of</strong> training set and the<br />
position <strong>of</strong> training data in rough margin. In this RFSVM,<br />
it consists <strong>of</strong> three regions: positive region, negative<br />
region and boundary region. It makes the original crisp<br />
margin become rough margin, the lower margin and the<br />
upper margin. The user can control the parameter<br />
ν and δ . One advantage <strong>of</strong> this method is that the<br />
classifier RFSVM is effective and robust with respect to<br />
misclassification and it considers the position <strong>of</strong> rough<br />
margin in fuzzy support vector machine. The<br />
experimental results on 16 datasets demonstrate that the<br />
generalization performance <strong>of</strong> RFSVM is better than the<br />
other SVM classifiers.<br />
ACKNOWLEDGMENT<br />
This work is supported by Natural Science Foundation<br />
<strong>of</strong> China (No. 60773062, No. 61073121) and Nature<br />
Science Foundation <strong>of</strong> Hebei Province (No.<br />
F2009000236).<br />
REFERENCES<br />
Different algorithm<br />
(2) Iris<br />
[1] V. N. Vapnik, “The Nature <strong>of</strong> Statistical Learning<br />
Theory.” New York: Springer-Verlag New York. 1995.<br />
“ISBN:0-387-94559-8”<br />
[2] C. Cortes, and V. N. Vapnik, “Support-Vector<br />
Networks.” Machine Learning, vol. 20, pp. 273-297,<br />
1995. “doi:10.1023/A:1022627411411”<br />
© 2011 ACADEMY PUBLISHER<br />
Accuracy<br />
Accuracy<br />
Different algorithm<br />
(3) Vowel<br />
Accuracy<br />
Different algorithm<br />
(4) Wine<br />
[3] C. J. C. Burges, “A tutorial on support vector machines<br />
for pattern recognition.” Data Mining and Knowledge<br />
Discovery, vol. 2, no. 2, pp. 121-167, June 1998.<br />
“doi:10.1023/A:1009715923555”<br />
[4] A. Blumer, A. Ehrenfeucht, D. Haussler, and M. K.<br />
Warmuth, “Learnability and the Vapnik-Chervonenkis<br />
Dimension.” <strong>Journal</strong> <strong>of</strong> the Association for Computing<br />
Machinery, vol. 36, no. 4, pp. 929-965, 1989.<br />
“doi:10.1145/76359.76371”<br />
[5] C. Hsu, and C. Lin, “A Comparison <strong>of</strong> Methods for<br />
Multiclass Support Vector Machines.” IEEE transactions<br />
on neural networks, vol. 13, no. 2, pp. 415-425, March<br />
2002. “doi:10.1109/72.991427”<br />
[6] T. Inoue S., and Abe, “Fuzzy Support Vector Machines<br />
for Pattern Classification.” International Joint<br />
Conference on Neural Networks, pp.1449-1454, July<br />
2001. “doi:101109/IJCNN.2001.939575”<br />
[7] X. F. Jiang, Z. Yi, and J. C. Lv, “Fuzzy SVM with a new<br />
fuzzy membership function.” Neural Computing and<br />
Applications, vol. 15, no. 3-4, pp. 268-276, 2006.<br />
“doi:10.1007/s00521-006-0028-z”<br />
[8] W. Yan, and Q. He, “Multi-class Fuzzy Support Vector<br />
Machine Based on Dismissing Margin.” Proceedings <strong>of</strong><br />
the Eighth International Conference on Machine<br />
Learning and Cybernetics, vol. 2, pp. 1139-1144, July<br />
2009. “doi:10.1109/ICMLC.2009.5212368”<br />
[9] C. F. Lin, and S. D. Wang, “Fuzzy Support Vector<br />
Machines.” IEEE transactions on neural works, vol. 13,<br />
no. 2, pp. 464-471, March 2002.<br />
“doi:10.1109/72.991432”<br />
[10] B. Jin, Y. C. Tang, and Y. Q. Zhang, “Support vector<br />
machines with genetic fuzzy feature transformation for<br />
biomedical data classification.” Information Sciences, vol.<br />
177, pp. 476-489, 2007. “doi:10.1016/j.ins.2006.03015”<br />
[11] Q. Wu, and R. Law, “Fuzzy support vector regression<br />
machine with penalizing Gaussian noises on triangular<br />
fuzzy number space.” Expert Systems with Applications,<br />
vol.37, no. 12, 2010. “doi:10.1016/j.eswa.2010.04061”<br />
[12] Z. Pawlak, “Rough sets.” International <strong>Journal</strong> <strong>of</strong><br />
Parallel Programming, vol. 11, no. 5, pp. 341-356, 1982.<br />
“doi:10.1007/BF01001956”<br />
[13] J. H. Zhang, and Y. Y. Wang, “A Rough Margin based<br />
Support Vector Machine.” Information Sciences, vol. 178,<br />
pp. 2204-2214, 2008. “doi:10.1016/j.ins.2007.12.012”<br />
[14] B. Scholkopf, A. J.Smola, R. C. Williamson, and P. L.<br />
Bartlett, “New Support Vector Algorithms.” Neural<br />
Computation, vol. 12, no. 5, pp. 1207–1245, 2000.<br />
“doi:10.1162/089976600300015565”<br />
[15] C. C. Chang, and C. J. Lin, “Training v-Support Vector<br />
Classifiers: Theory and Algorithms.” Neural<br />
Computation, vol. 13, pp. 2119–2147, 2001.<br />
“doi:10.1162/089976601750399335”
1970 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
[16] P. H. Chen, C. J. Lin, and B. Scholkopf, “A Tutorial on<br />
[nu]-Support Vector Machines.” Applied Stochastic<br />
Models in Business and Industry, vol. 21, no. 2, pp. 111-<br />
136, 2002. “doi:10.1002/asm.537”<br />
[17] C. L. Blake, and C. J. Merz, UCI Repository <strong>of</strong> machine<br />
learning databases<br />
[http://www.ics.uci.edu/~mlearn/MLRepository.html].<br />
Irvine, CA: University <strong>of</strong> California. Department <strong>of</strong><br />
Information and Computer Science, 1998.<br />
[18] R. D. King, C. Feng, and A. Sutherland, “Statlog :<br />
Comparison <strong>of</strong> classification algorithms on large realworld<br />
problems.” Applied Artificial Intelligence, vol. 9,<br />
no. 3, pp. 289-333, 1995.<br />
“doi:10.1080/08839519508945477”<br />
[19] T. K. Ho, and E. M. Klernberg, “Building projectable<br />
classifiers <strong>of</strong> arbitrary complexity.” Proceeding <strong>of</strong> the<br />
13th International Conference on Pattern Recognition,<br />
vol. 2, pp. 880-885, 1996.<br />
“doi:10.1109/ICPR.1996.547202”<br />
© 2011 ACADEMY PUBLISHER<br />
Kai Li, born in Baoding, China, 1963. He received Bachelor<br />
Degree and Master Degree in mathematics and education<br />
technology from Hebei University, Baoding, China in 1986 and<br />
1995, respectively. In 2005, he received PhD degree in<br />
computer from Beijing Jiaotong University, Beijing, China. His<br />
research interests include machine learning, neural network,<br />
pattern recognition, data mining, etc.<br />
Currently, he is a Pr<strong>of</strong>essor at school <strong>of</strong> mathematics and<br />
computer science, the Hebei University. He has published over<br />
fifty papers on machine learning, clustering, ensemble learning,<br />
support vector machine, and pattern recognition.<br />
Xiaoxia Lu, Born in Shijiazhuang, China, 1984. She received<br />
Bachelor Degree in computer science from Hebei University,<br />
Baoding, China in 2009.<br />
Currently, she is a Master student in the school <strong>of</strong><br />
mathematics and computer science at Hebei University. Her<br />
research interests are in the areas <strong>of</strong> support vector machine,<br />
fuzzy sets and rough sets theory.
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1971<br />
A Modified Technique for Analysis <strong>of</strong><br />
Synchronous Counters Constructed with<br />
Flip-flops<br />
Dangui Yan 1<br />
1 College <strong>of</strong> Mathematics and Physics<br />
Chongqing University <strong>of</strong> Post and Telecom<br />
Chongqing, P.R. China<br />
Email: yandg@cqupt.edu.cn<br />
Ruijun Tong 2 , Chengchang Zhang 3 , Changyong Li 4<br />
2 Deptartment <strong>of</strong> Electronic Engineering, Chongqing College <strong>of</strong> Electronics Engineering, P.R. China<br />
tongrj@163.com<br />
3 College <strong>of</strong> Communication Engineering, Chongqing University, Chongqing, P.R. China<br />
Email:zcc_918@163.com<br />
4 Chongqing Communication Acadimic <strong>of</strong> P.L.A. ,Chongqing, P.R. China<br />
Email: lll_ccc_yyy@163.com<br />
Abstract—Some methods <strong>of</strong> fabrication make it<br />
economically attractive to construct counters (and other<br />
devices) by connecting sets <strong>of</strong> identical flip-flops(FFs), if the<br />
FFs have a common clock input, the state transitions <strong>of</strong> the<br />
whole counters are as rapid as the state transitions <strong>of</strong> each<br />
FFs, so that the counter is further restricted to be<br />
synchronous. In order to simplify the process for analysis <strong>of</strong><br />
synchronous counters constructed with flip-flops, a simple<br />
and successful method is proposed. Using this method, the<br />
state transition equations obtained from logic diagram <strong>of</strong><br />
counter are converted to standard sum-<strong>of</strong>-products<br />
forms(SOPs). By finding out the logic principle for<br />
achieving the value <strong>of</strong> logic function based on the standard<br />
SOPs, the values <strong>of</strong> next state can be directly obtained<br />
without any Boolean calculation. Analysis for a 3-bits<br />
counter shows that this method eliminates complex<br />
calculations, and makes the process <strong>of</strong> obtaining next state<br />
value and developing truth table more rapid and<br />
convenient.<br />
Index Terms— counter, synchronous, flip-flops,<br />
equation, calculation, sum-<strong>of</strong>-products forms<br />
I. INTRODUCTION<br />
We call a device that accepts clock pulses as input and<br />
that exhibits periodic behavior as output a counter. Some<br />
methods <strong>of</strong> fabrication make it economically attractive to<br />
construct counters (and other devices) by connecting sets<br />
<strong>of</strong> identical flip-flops(FFs), such as D FFs, J-k FFs, and<br />
so on[1-6]. If the FFs have a common clock input, the<br />
state transitions <strong>of</strong> the whole counters are as rapid as the<br />
state transitions <strong>of</strong> each FFs, so that the counter is further<br />
restricted to be synchronous[7-10]. These counters may<br />
be clocked at a maximum rate counter for they have no<br />
gates or ripple effects to introduce delays. Thus, a<br />
© 2011 ACADEMY PUBLISHER<br />
doi:10.4304/jcp.6.9.1971-1975<br />
synchronous counter can operate at a much higher input<br />
frequency and have numerous well-known uses in digital<br />
apparatus[11,12].<br />
For a given synchronous counter constructed with FFs,<br />
in order to know its logic function, there are two basic<br />
messages should be obtained by analyzing the logic<br />
diagram, one is to obtain modulus value <strong>of</strong> the counter,<br />
the other is to know whether the counter can self-starting.<br />
The first step for analysis is to get the state transition<br />
equations from counter’s logic diagram, the following<br />
step is to develop truth table and state transition table,<br />
from which one can obtain information about modulus<br />
value and self-starting. The key to develop truth table is<br />
how to obtain the next state values. Present method uses<br />
substitution method[13-15], in which the values <strong>of</strong><br />
present state are substituted into the state transition<br />
equations and the next state values are obtained by<br />
Boolean calculation. The disadvantage <strong>of</strong> this method is<br />
that it needs a large number <strong>of</strong> calculations with timeconsuming<br />
and error-prone.<br />
In this paper, we propose a modified method, which<br />
converts the state transition equations to standard SOPs,<br />
by finding out the principle for the values <strong>of</strong> next state,<br />
the next state values should be directly obtained based on<br />
the standard SOPs without the need for any calculation.<br />
Analysis for a 3-bit synchronous counter constructed with<br />
three J-k FFs flip-flops shows that this method eliminates<br />
the complex calculations <strong>of</strong> current method, and makes<br />
the process for analyzing synchronous counters more<br />
rapid and convenient.<br />
II. PRESENT METHOD<br />
Consider a 3-bits synchronous counter whose logic<br />
diagram is shown in Fig.1. It uses three J-K FFs-FF0,<br />
FF1and FF2-and each one has a J and a K input.
1972 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
&<br />
&<br />
CLK<br />
J<br />
K<br />
Q<br />
Q<br />
Q0<br />
&<br />
K<br />
J<br />
Q<br />
Q<br />
Q1<br />
Fig 1. The logic diagram<br />
&<br />
K<br />
J<br />
Q<br />
Q<br />
Q2<br />
In order to obtain the modulus value <strong>of</strong> the counter and<br />
to know whether the counter can self-starting, a general<br />
procedure is applied in three steps when analyzing this<br />
counter.<br />
Step 1: Obtaining the state transition equations<br />
From the logic diagram <strong>of</strong> Fig .1, we can obtain the<br />
following expressions for the J and K inputs <strong>of</strong> each FFs:<br />
J = Q + Q, K = Q + Q<br />
0 2 1 0 2 1<br />
J = QQ , K = Q<br />
1 2 0 1 2<br />
J = QQ , K = Q<br />
2 1 0 2 1<br />
According to the characteristic equation <strong>of</strong> J-K FFs<br />
[16,17], which is shown as equation (2).<br />
(1)<br />
Q′ = J Q + k Q 0 ≤ i ≤ n−<br />
1 (2)<br />
i i i i i<br />
where i Q′ denotes next state, Q denotes present state,<br />
i<br />
n is the number <strong>of</strong> FFs. The expressions <strong>of</strong> J and i<br />
K ( 0≤i≤ 2)<br />
are separately substituted into equation(2),<br />
i<br />
we can obtain the state transition equations as shown in<br />
equations(3):<br />
⎧Q′<br />
= QQQ + Q Q<br />
2 2 1 0 2 1<br />
⎪<br />
⎨Q′<br />
= QQQ + QQ<br />
1 2 1 0 2 1<br />
⎪<br />
⎩ Q′ = QQQ + QQ + QQ<br />
0 2 1 0 2 0 1 0<br />
Step 2: The state truth table<br />
We develop a truth table for the equations(3) as shown<br />
in Tab.1, the present state includes three variables<br />
Q2、 Q and 1 Q in the domain, so there are eight<br />
0<br />
possible combinations <strong>of</strong> binary values <strong>of</strong> the variables as<br />
listed in medium columns, which are 000, 001, 010, 011,<br />
100, 101, 110 and 111, the decimal digit corresponding to<br />
each binary value is listed in left column, which are from<br />
0 to 7, the next state is listed in right columns.<br />
After a CLK pulse input, the FFs enter a new<br />
state(from present state to next state), assuming initial<br />
present state is 000, that is,<br />
Q = 0, Q = 0, Q = 0 ( 2 1 0 000<br />
QQQ = ), substituting it<br />
2 1 0<br />
into equation 2 Q′ <strong>of</strong> equations(3), we can get<br />
corresponding value <strong>of</strong> next state, the process is as follow:<br />
© 2011 ACADEMY PUBLISHER<br />
(3)<br />
Tab. 1 The truth table<br />
Q2 Q1 Q0<br />
' ' '<br />
Q 2 Q 1 Q 0<br />
Q′ = 0⋅0⋅ 0+ 0⋅ 0= 0<br />
(4)<br />
2<br />
Here, we define such process <strong>of</strong> obtaining value <strong>of</strong> i Q′<br />
as one time next state equation calculation, obviously, it<br />
perhaps includes NOT, AND and OR operations. Next,<br />
we can calculate the values <strong>of</strong> 1 Q′ and 0 Q′.<br />
Q′<br />
= 0⋅0⋅ 0+ 0⋅ 0= 0<br />
1<br />
Q′<br />
= 0⋅0⋅ 0+ 0⋅ 0+ 0⋅ 0= 1<br />
(5)<br />
0<br />
So, after three times calculations, we obtain the next<br />
state QQQ ′ ′ ′ = 001,<br />
and fill the values in correspondence<br />
2 1 0<br />
positions in right-hand column <strong>of</strong> the truth table, which is<br />
shown in Tab.1.<br />
Regarding a new present state is 001 and substituting it<br />
into equations(3), the new next state can be obtained,<br />
which is 011. Regarding the third present state is 010 and<br />
substituting it into equations(3) again, the new next state<br />
can be obtained, which is 110. This process is continued,<br />
when the next state corresponding to the present state 111<br />
is calculated, the process is ended.<br />
It is clearly to know, in order to complete the truth<br />
table(Tab.1), we need accomplish 24(8*3) times next<br />
state equation calculations.<br />
Step 3: The state transition table and the state<br />
transition diagram<br />
After obtaining the truth table, the next step is to<br />
construct the state transition table and the state transition<br />
diagram.<br />
Firstly, we construct the state transition table. The table<br />
is constructed with three columns, which are comments,<br />
present state and nest state, which is shown in Tab.2.<br />
Regarding initial present state is 000, observing Tab.1,<br />
the next state is 001, filling 000 and 001 to corresponging<br />
column <strong>of</strong> Tab.2. Regarding the next state 001 is as a new<br />
present state, the new next state is 011. Regarding the<br />
third present state is 011 and the third next state is 010.<br />
This process is continued, in the end, we can observe 101<br />
is as a present state, which next state is 000, the counting<br />
cycle is finished, there are six counting states, which are<br />
000, 001, 011, 010, 110 and 101, we call these states as<br />
counting states. For next two states, which are 100 and
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1973<br />
111, for they are not in counting cycle, we call them as<br />
<strong>of</strong>fset states.<br />
Comments<br />
Counting<br />
states<br />
Offset<br />
states<br />
Tab.2 The state transition table<br />
Present State Next State<br />
Q2 Q1 Q0<br />
' ' '<br />
Q 2 Q 1 Q 0<br />
0<br />
0<br />
0<br />
0<br />
1<br />
1<br />
0<br />
0<br />
1<br />
1<br />
1<br />
0<br />
0<br />
1<br />
1<br />
0<br />
0<br />
1<br />
0<br />
0<br />
0<br />
1<br />
1<br />
0<br />
0<br />
1<br />
1<br />
1<br />
0<br />
0<br />
1<br />
1<br />
0<br />
0<br />
1<br />
0<br />
1<br />
1<br />
0<br />
1<br />
0<br />
1<br />
0<br />
1<br />
0<br />
0<br />
1<br />
0<br />
Secondly, we can construct the state transition diagram<br />
based on Tab.2 easily, which is shown in Fig.2.<br />
Fig.2. The state transition diagram<br />
Observing Tab.2 and Fig.2, there are six counting<br />
states, which are 000, 001, 011, 010, 110 and 101, and<br />
two <strong>of</strong>fset states, they are 111 and 100. These <strong>of</strong>fset<br />
states can return to counting states after one or two clock<br />
cycles, obviously, it is a Mod-6 counter and has selfstarting<br />
performance.<br />
Summarizing above steps, we can see that the key step<br />
<strong>of</strong> this method is to construct the state truth table. It is<br />
clear that the disadvantage <strong>of</strong> this method is its large<br />
calculations during constructing the state truth table . If n<br />
is the number <strong>of</strong> state variable, there are n state transition<br />
equations,which are Q′ , Q′ ,…, n−1<br />
n−2<br />
0 Q′ , and 2 n present<br />
states (from 00…0 to 11…1). In order to create the truth<br />
table, one must calculate 2 n n times next state equation<br />
calculations, and every calculation includes NOT, AND<br />
and OR operations. This procedure seems to be simple<br />
but in fact it is time-consuming and easy to make<br />
mistakes.<br />
III. PROPOSED METHOD<br />
A. Logic principle<br />
Any n-variable state transition equation can be<br />
expressed canonically by the standard SOPs as follows<br />
[18].<br />
© 2011 ACADEMY PUBLISHER<br />
n<br />
j=2 −1<br />
Q′ i ( Q ... Q ) = a Q Q ∑ ... Q<br />
n−1 0 j n−1 n−2<br />
0<br />
j=0<br />
=<br />
=<br />
where { 0,1}<br />
n<br />
j=2 −1<br />
∑<br />
j=0<br />
n<br />
j=2 −1<br />
∑<br />
j=0, a j =1<br />
a m<br />
j j<br />
<br />
j<br />
, 0≤ p ≤ n − 1,<br />
p p p<br />
Σ<br />
is the OR operator, m j is a minterm, j is the minterm<br />
number.<br />
m<br />
a ∈ , Q ∈{<br />
Q , Q }<br />
n−1<br />
k<br />
2 ( Qk<br />
)<br />
k=0<br />
j = ∑ d <br />
where if Q = Q that dQ ( ) = 1,<br />
if Q = Q that<br />
k k<br />
k<br />
k k<br />
d ( Q<br />
) = 0.<br />
k<br />
For any present state Qn−1... Q0 = x n− 1... x k...<br />
x0<br />
,<br />
x ∈ 0,1 , 0≤k ≤ n−<br />
1.<br />
Definiting q,<br />
k<br />
{ }<br />
q =2 x + ... 2 x<br />
n -1<br />
0<br />
n−<br />
1 + 0<br />
j<br />
(6)<br />
The value <strong>of</strong> next state i Q′ is calculated as follows:<br />
Q′ ( Q ... Q )= Q′ ( x ... x )<br />
i n−1 0 i n−1<br />
0<br />
n<br />
j=2 −1<br />
∑<br />
= a m + a m<br />
q q j j<br />
j=0, j≠ q<br />
n<br />
j=2 −1<br />
∑<br />
= a .1 + a .0 = a<br />
q j q<br />
j=0, j≠ q<br />
Where if the minterm m is included in the standard<br />
q<br />
SOPs that a q is equal to 1, otherwise aq is equal to 0.<br />
That is, if the minterm corresponding to a present state is<br />
included in the state transition equation that the next state<br />
value is equal to 1, otherwise equal to 0.<br />
As an example, taking into account a 2-bits state<br />
transition equations as equations(8):<br />
⎧<br />
⎨<br />
⎩<br />
0 1 0 1 0<br />
(7)<br />
Q′ = QQ = ∑m( 00 ) = ∑m(<br />
0)<br />
1 1 0<br />
Q′ = QQ + Q Q = ∑m( 01,10 ) = ∑m(<br />
1, 2)<br />
(8)<br />
For expression Q′ = QQ , which includes one<br />
1 1 0<br />
minterm QQ (m0), when present state values 00, Q′ is<br />
1 0<br />
1<br />
equal to 1, for any other present state, such as 01,10,11,<br />
Q′ = QQ + Q Q , it<br />
Q′ is equal to 0. For expression 1<br />
0 1 0 1 0<br />
includes two minterms QQ (m1) and QQ (m2), when<br />
1 0<br />
1 0<br />
present state values 01 or 10, the value <strong>of</strong> next state 0 Q′
1974 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
equals to 1, otherwise equals to 0. All values are showm<br />
in Fig.3.<br />
B. Analysis for Fig.1<br />
Q1 Q0<br />
' '<br />
Q1 Q0<br />
Fig.3 All values for equations(8)<br />
By the proposed method, the analysing process for the<br />
3-bit synchronous counter shown in Fig.1 includes main<br />
three steps.<br />
Step 1: The standard SOPs<br />
Each state transition equation is converted to standard<br />
SOPs, taking 2 Q′ as an example, the converting processes<br />
are as shown in Equation.9, this processes can also be<br />
realized by Karnaugh Maps[19,20].<br />
Q′ = Q Q Q + Q Q<br />
2 2 1 0 2 1<br />
= QQQ + QQ( Q + Q)<br />
2 1 0 2 1 0 0<br />
= QQQ + QQQ + QQQ<br />
= ∑ m (010,110,111)<br />
= ∑ m (2,6,7)<br />
2 1 0 2 1 0 2 1 0<br />
The conversion to obtain standard SOPs By similar<br />
processes, the equations Q′ and 1<br />
0 Q′ can be obtained as<br />
shown in equations(10).<br />
Q′ = QQQ + Q QQ + Q QQ<br />
1 2 1 0 2 1 0 2 1 0<br />
= ∑ m(<br />
001,010,011)<br />
= ∑ m(<br />
1,2,3)<br />
Q′ = Q QQ + Q QQ + Q QQ + Q Q Q<br />
0 2 1 0 2 1 0 2 1 0 2 1 0<br />
= ∑ m(<br />
000, 001,100, 110)<br />
= ∑ m(<br />
01,4,6 , )<br />
Step 2: The truth table<br />
(9)<br />
(10)<br />
(11)<br />
We develop a truth table for the equations(9), (10) and<br />
(11), eight possible combinations <strong>of</strong> binary values are<br />
listed in the medium columns, the decimal digit<br />
corresponding to each binary value is listed in the left<br />
column. According to the principle for the value <strong>of</strong> next<br />
state, the present state values that make the next state 2 Q′<br />
equal to 1 are 010(2), 110(6), and 111(7). For each <strong>of</strong><br />
these values, a 1 is filled in each corresponding position<br />
in the first column <strong>of</strong> three right columns.<br />
© 2011 ACADEMY PUBLISHER<br />
Using the same principle, when the present state values<br />
are 001(1), 010(2), and 011(3), the next state Q′ equal<br />
1<br />
to 1, a 1 is filled in each corresponding position in the<br />
second column <strong>of</strong> right columns. When the present state<br />
values are 000(0); 001(1); 100(4); and 110(6), the next<br />
state 0 Q′ equals to 1, a 1 is filled in each corresponding<br />
position in the third column <strong>of</strong> right columns. After all 1<br />
are filled, the view <strong>of</strong> truth table is shown as Tab.3.<br />
Tab. 3 The truth table<br />
Q2 Q1 Q0<br />
' ' '<br />
Q 2 Q 1 Q 0<br />
All the remaining positions in right columns are placed<br />
by a 0, we can get the same truth table as Tab. 1.<br />
The third step is same as the present method, detailed<br />
analysis is no longer given here.<br />
IV. CONCLUSIONS AND DISCUSSIONS<br />
This paper deduces the principle for obtaining next<br />
state values in state transition equations. Based on the<br />
principle, a modified method has been proposed to<br />
analyse synchronous counters constructed with flip-flops.<br />
The method can develop the truth table directly from state<br />
transition equations with SOPs. This method facilitates<br />
analysis <strong>of</strong> synchronous counters constructed with Flip-<br />
Flops by eliminating large number <strong>of</strong> Boolean<br />
calculations.<br />
The number <strong>of</strong> state variables n is three in the example,<br />
if there have more state variables, the calculation will be<br />
increased, such as n=4, the number <strong>of</strong> minterm is 16 and<br />
the state transition equation is 4, it is needed 64 times<br />
next state equation calculations to accomplish the truth<br />
table. Clearly, the advantages are more obvious with the<br />
increase <strong>of</strong> n. Certainly, if n is enough large, for example,<br />
n=10, although the proposed method is still more rapid<br />
and convenient than present one, the obtaining for<br />
minterm is very complex, so we advise that one should<br />
analyze with computer.<br />
Obviously, the method is also suitable for synchronous<br />
counters constructed with other FFs, such as D FFs, and<br />
so on.<br />
ACKNOWLEDGMENT<br />
The authors thank their colleagues at College <strong>of</strong><br />
Communication Engineering for fruitful discussions. This<br />
work was supported by Natural Science Foundation<br />
Project <strong>of</strong> CQ CSTC under contract no: 2010BB2240.
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1975<br />
REFERENCES<br />
[1] Stan, M.R.; Tenca, A.F.; Ercegovac, M.D. Long and fast<br />
up/down counters. IEEE Trans on computers, Vol. 47,<br />
No. 7, July 1998, pp. 722-735.<br />
[2] ZENG Xiao-pang; WANG Peng-jun. Design <strong>of</strong> four valued<br />
synchronous reversible counter based on the theory <strong>of</strong><br />
three essential circuit elements. <strong>Journal</strong> <strong>of</strong> Zhejiang<br />
University(Science Edition), 2009, vol.36(5), pp. 553-<br />
556.<br />
[3] YE Xi-en, TAO Wei-jiong , WANG Lun-yao. A low power<br />
terminary D type flip-flop design based on clock gating<br />
technique . <strong>Journal</strong> <strong>of</strong> Circuits and Systems, 2006, 11(3),<br />
pp. 106-109.<br />
[4] Vaquero, A.R.; Aguilo, J.. Gateless synchronous counters<br />
with D flip-flops. Electronics Letters Vol.14 (16) , 1978 ,<br />
pp. 496-498.<br />
[5] F. B. Manning. Autonomous, Synchronous Counters<br />
Constructed Only <strong>of</strong> J-K Flip-Flops. S.M. thesis,<br />
available in micr<strong>of</strong>iche from MIT Barker Engineering<br />
Library or in paperback as Project MAC, Massachusetts<br />
Inst. Technol.,Cambridge, MA, Tech. Rep. Period<br />
Counlter Comment TR-96, 1972.<br />
[6] YE Xi-en, TAO Wei-jiong, WANG Lun-yao. A low power<br />
ter nary D type flip-flop design based on clock gating<br />
technique. <strong>Journal</strong> <strong>of</strong> Circuits and Systems, 2006, 11(3),<br />
pp.106-109.<br />
[7] WU Zhong guang 1; YANG Yu zhi. The Method <strong>of</strong> Real<br />
Time and Synchronous Counting for High Speed Multi-<br />
Event by CPLD. <strong>Journal</strong> <strong>of</strong> Sichuan University (Natural<br />
Science Edition). 2002,vol.39(1), pp. 62-64.<br />
[8] WU Zhong guang; YANG Yu zhi. The Method <strong>of</strong> Real<br />
Time and Synchronous Counting for High Speed Multi-<br />
Event by CPLD. <strong>Journal</strong> <strong>of</strong> Sichuan University (Natural<br />
Science Edition). 2002,vol.39(1), pp. 62-64.<br />
[9] Misra, S.K.; Kolagotia, R.K.; Srinivas, H.R.; Mo, J.C.;<br />
Diamondstein, M.S. VLSI implementation <strong>of</strong> a 300-MHz<br />
0.35-µm CMOS 32-bit auto-reloadable binary<br />
synchronous counter with optimal test overhead delay .<br />
VLSI Design, 1998. Proceedings. Eleventh International<br />
Conference on , 1998, pp. 326-329.<br />
[10] Aguirre-Hernandez, M.; Linares-Aranda, M. A Clock-<br />
Gated Pulse-Triggered D Flip-Flop for Low-Power High-<br />
Performance VLSI Synchronous Systems. Devices,<br />
Circuits and Systems, Proceedings <strong>of</strong> the 6th International<br />
Caribbean Conference on, 2006, pp. 293-297.<br />
[11] Misra, S.K.; Kolagotia, R.K.; Srinivas, H.R.; Mo, J.C.;<br />
Diamondstein, M.S. VLSI implementation <strong>of</strong> a 300-MHz<br />
0.35-µm CMOS 32-bit auto-reloadable binary<br />
synchronous counter with optimal test overhead delay .<br />
VLSI Design, 1998. Proceedings., 1998 Eleventh<br />
International Conference on , pp. 326-329.<br />
[12] Aguirre-Hernandez, M.; Linares-Aranda, M. A Clock-<br />
Gated Pulse-Triggered D Flip-Flop for Low-Power High-<br />
Performance VLSI Synchronous Systems. Devices,<br />
Circuits and Systems, Proceedings <strong>of</strong> the 6th International<br />
Caribbean Conference on , 2006 , pp. 293-297.<br />
© 2011 ACADEMY PUBLISHER<br />
[13] Thomas L.Floyd. Digital Fundamentals. 9th ed. P.<br />
cm.2004, pp. 398-403.<br />
[14] Wang Yuyin. Digital circuit and logic design. 3rd ed.<br />
Higher education press, 2008, pp.181.<br />
[15] WANG Shi-yuan, XIE Kai-ming, SHI Ya-wei, CHEN<br />
Meng-gang, LONG Zheng-ji. Implementation <strong>of</strong> a New<br />
FPGA-Based Controllable Frequency Divider. <strong>Journal</strong> <strong>of</strong><br />
Southwest University (Natural Science Edition). 2007,<br />
vol. 29(1), pp. 89-93.<br />
[16] Frank B.Manning AND Rober R. Fenichel. Synchronous<br />
Counters Constructed Entirely <strong>of</strong> J-K Flip-Flops. IEEE<br />
Trans on computers, March 1976, pp. 300-306.<br />
[17] Manning, Frank B.; Fenichel, Robert R. Synchronous<br />
Counters Constructed Entirely <strong>of</strong> J-K Flip-Flops . IEEE<br />
Transactions on <strong>Computers</strong>, Vol: C-25 (3), 1976, pp. 300-<br />
306.<br />
[18] L. Wang and A.E.A. Almani. Fast conversion algorithm<br />
for very large Boolean functions. Electronics letters, Vol.<br />
36, No. 16, August 2000, pp. 1370-1371.<br />
[19] Michel E. Holder. A Modified Karnaugh Map Technique.<br />
IEEE Trans on education, Vol. 48, NO. 1, February 2005.<br />
pp. 206-207.<br />
[20] Dean, K.J.. An extension <strong>of</strong> the use <strong>of</strong> karnaugh maps in<br />
the minimization <strong>of</strong> logical functions. Radio and<br />
Electronic Engineer. Vol.35(5) ,1968 , pp.294-296.<br />
Dangui Yan was born in Luotian, China, 1975. She received<br />
the BS degree in Department <strong>of</strong> Mathematics from Hubei<br />
institutes for nationalities in 1997. She received the MS degree<br />
in Department <strong>of</strong> Mathematics from Hubei University in 2000.<br />
She is currently lecturer <strong>of</strong> Chongqing University <strong>of</strong> Post and<br />
Telecom. Her research interest is logic algebra.<br />
Ruijun Tong was born in Shanxi, China, 1976. She received<br />
the MS degree in College <strong>of</strong> Communications Engineering from<br />
Chongqing University in 2005. She is currently lecturer <strong>of</strong><br />
Chongqing College <strong>of</strong> Electronics Engineering. Her research<br />
interests are digital system and FPGA design.<br />
Chengchang Zhang was born in Lichuan, China, 1975. He<br />
received the BS degree in automation engineering from the<br />
Wuhan Institute <strong>of</strong> Chemical Technology in 1997. He received<br />
the MS degree in College <strong>of</strong> Communications Engineering from<br />
Chongqing University in 2005. He is currently PhD candidate <strong>of</strong><br />
Chongqing University majoring in Communication and<br />
information systems. His research interests are s<strong>of</strong>tware radio<br />
and FPGA design.<br />
Changyong Li was born in Chongqing, China, 1971. His<br />
research interests are s<strong>of</strong>tware radio and ultra-wide band radar.
1976 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
A New Method <strong>of</strong> Detecting Multi-component<br />
LFM Signals Based on Blind Signal Processing<br />
1 Qiang Guo<br />
1 Space Control and Inertial Technology Research Center,Harbin Institute <strong>of</strong> Technology, P.R. China<br />
Email: guoqiang292004@163.com<br />
2 Yajun Li and 1 Changhong Wang<br />
2 College <strong>of</strong> Information and Communication Engineering, Harbin Engineering University,P.R.China<br />
Email: liyajun1985happy@163.com<br />
Abstract—To effectively detect and recognize multicomponent<br />
Linear Frequency-Modulated (LFM) emitter<br />
signals, a multi-component LFM emitter signal analysis<br />
method based on the complex Independent Component<br />
Analysis(ICA) which was combined with the Fractional<br />
Fourier Transform(FRFT) was proposed. The idea which<br />
was adopted to this method was the time-domain separation<br />
and then time-frequency analysis, and in the low SNR cases,<br />
the problem which is generally plagued by noised <strong>of</strong> feature<br />
extraction <strong>of</strong> multi-component LFM signal based on FRFT<br />
is overcame. Compared to the traditional method <strong>of</strong> timefrequency<br />
analysis, the computer simulation results show<br />
that the proposed method for the multi-component LFM<br />
signals separation and feature extraction was better.<br />
Index Terms—multi-component LFM emitter signals, timefrequency<br />
analysis, feature extraction,ICA<br />
I. INTRODUCTION<br />
Radar emitter signal detection is a key problem which<br />
is demanded to be resolved in modern electronic<br />
reconnaissance system. With large new complex radar<br />
systems in practice, a large number <strong>of</strong> pulses overlap and<br />
form the multi-component emitter signals(MCES)[1].<br />
MCES analysis is a prerequisite and primary task for<br />
detecting and identifying emitter signals. Multicomponent<br />
Linear Frequency Modulated (LFM) emitter<br />
signals is a non-stationary signal which is commonly<br />
used in active sonar, radar imaging, fuse <strong>of</strong> missile and so<br />
on. As a new time-frequency analysis tool, FRFT is a<br />
generalization <strong>of</strong> the Fourier transform (FT). It not only<br />
has a natural link in classical FT, but also provides some<br />
characteristics which FT do not have. So FRFT is<br />
specially suitable for processing LFM class (chirp-like)<br />
signal. At present, regardless <strong>of</strong> the traditional parameter<br />
estimation or detection methods <strong>of</strong> multi-component<br />
LFM signal, most <strong>of</strong> them are based on time-frequency<br />
analysis or all finds <strong>of</strong> FT method[2]. The parameter<br />
estimation methods mainly through two-dimensional<br />
Manuscript received January 2, 2011; revised February 1, 2011;<br />
accepted February 28, 2011.<br />
Qiang Guo, Yajun Li,Changhong Wang.<br />
© 2011 ACADEMY PUBLISHER<br />
doi:10.4304/jcp.6.9.1976-1982<br />
object function, and combined with two-dimensional<br />
search to estimate, such as the maximum likelihood<br />
method, the time-frequency analysis methods, FRFT,<br />
match fourier transform(DCFT)[3,4] ,S-method and so<br />
on. But when the low SNR, together with the existence <strong>of</strong><br />
weak signal, the traditional detection method is <strong>of</strong>ten<br />
difficult to effectively detect the MCES, even lead to<br />
misjudgment <strong>of</strong> signal and noise ,so the result <strong>of</strong> crossterm<br />
suppression is not good. Above-mentioned issue has<br />
been a relatively difficult problem. In this paper, a multicomponent<br />
LFM emitter signal analysis method based on<br />
complex FastICA which was combined with FRFT was<br />
proposed. Firstly, complex ICA algorithm was used as<br />
time-domain separation for multi-component LFM<br />
emitter signals with noise. Secondly, determine the signal<br />
and noise by the automatic identification method <strong>of</strong><br />
second central moment <strong>of</strong> FRFT. Lastly, the noise was<br />
removed and the LFM signals were detected by FRFT.<br />
The effect <strong>of</strong> noise can be greatly reduced.<br />
Simultaneously, the cross-terms are effectively deduced<br />
with higher time-frequency resolution. It is a good<br />
method for the multi-component LFM signals.<br />
Simulation results verify the effectiveness <strong>of</strong> this new<br />
method.<br />
II. MODEL OF MULTI-COMPONENT LFM EMITTER SIGNALS<br />
In modern electronic reconnaissance system, receiver<br />
<strong>of</strong>ten intercepted to pulses which emitted by multiple<br />
sources at the same time. A stream <strong>of</strong> pulses was formed<br />
through these pulses interleaved together. As the pulse<br />
signals density increases, the pulse formed MCES<br />
x() t ,the signal model is expressed as follows.<br />
k −1<br />
∑<br />
i=<br />
0<br />
i<br />
2<br />
j2 π ( fit+ ( µ it<br />
/2)) , −∆t/2 ≤ t ≤ ∆ t/2<br />
x() t = Ae + n() t<br />
(1)<br />
Where Ai is the amplitude <strong>of</strong> each signal, fi is initial<br />
frequency and µ i is chirp rate. nG() t is White Gauss<br />
Noise with zero mean and variance 2<br />
σ .<br />
III. MULTI-COMPONENT LFM EMITTER SIGNALS<br />
ANALYSIS
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1977<br />
In this paper, the flowchart <strong>of</strong> proposed multicomponent<br />
LFM signals analysis method is shown in<br />
Fig.1. Firstly, complex FastICA algorithm is used as the<br />
pretreatment <strong>of</strong> detection <strong>of</strong> multi-component LFM<br />
emitter signals. The emitter signals and noise were<br />
separated by it. Secondly, discriminate LFM signal and<br />
noise based on the automatic identification method <strong>of</strong><br />
second-order central moment <strong>of</strong> FRFT. Then detect the<br />
separated LFM signals through FRFT after removing the<br />
noise. Through the above steps, the impact <strong>of</strong> noise was<br />
effectively reduced and the cross terms were effectively<br />
restrained. Last, the LFM signal parameter estimation can<br />
be completed through the conversion formula [1,5].<br />
Figure 1. Multi-component LFM emitter signals analysis with low SNR<br />
A. Pretreatment<br />
1)Problem Description<br />
As the number <strong>of</strong> mixed-signal (decided by the number<br />
<strong>of</strong> channel) received by the radar signal receiver and do<br />
not necessarily match the number <strong>of</strong> radar emitter signals,<br />
and LFM signal is non-stationary signal, so the complex<br />
ICA <strong>of</strong> the blind signal processing techniques was used in<br />
multi-component LFM signals with low SNR to do timedomain<br />
separation pretreatment.<br />
At present, most <strong>of</strong> the FastICA algorithm is mainly<br />
concentrated in the real domain in blind separation <strong>of</strong><br />
mixed signals[7], but in practice, many real signal model<br />
is represented by linear mixed model <strong>of</strong> complex signals.<br />
Blind separation algorithm for complex signals is<br />
summarized into the methods based on higher order<br />
statistics and the methods based on second order statistics.<br />
It is actually an optimization problem, namely, how to<br />
make the separated independent component to maximum<br />
approach the source signal. Here we will extend the real<br />
domain variable to the complex domain[8,9].<br />
Taking into account the observed mixed-signal is<br />
instantaneous linear mixing <strong>of</strong> the each source signal, the<br />
standard ICA linear model with noise as follows [3]<br />
n<br />
xi() t = ∑ aijsj() t + ni() t ( i = 1,2, ⋅⋅⋅ , m)<br />
(2)<br />
j=<br />
1<br />
© 2011 ACADEMY PUBLISHER<br />
iii<br />
iii<br />
iii<br />
Expressed in matrix form, ie.<br />
X = A S+ n<br />
(3)<br />
where X = ( x1, x2,..., xm)<br />
is the vector <strong>of</strong> observed<br />
random variables, S = ( s1, s2,..., sn)<br />
is the vector <strong>of</strong><br />
statistically independent latent variables called the<br />
independent components, and A is an unknown constant<br />
complex mixing matrix. The above model is identifiable<br />
under the following fundamental restrictions:<br />
① At most one <strong>of</strong> the independent components s j may<br />
be Gaussian.<br />
② The matrix A must be <strong>of</strong> full column rank. The<br />
number <strong>of</strong> observing signals m is more than the number<br />
<strong>of</strong> source signals n . ( m≥ n)<br />
,here m=n.<br />
③ The various components <strong>of</strong> the source signals<br />
si ( i = 1,..., m)<br />
and observed signals x i are zero-mean and<br />
unit variance.<br />
In addition, the noise itself can be regarded as a source<br />
<strong>of</strong> signal to use BSS, and thus make the algorithm have a<br />
wider scope and greater robustness.<br />
Can be proved that we can find a matrix W by linear<br />
transformation to do m mixed-signal X , making between<br />
the each component <strong>of</strong> new vector Y obtained by<br />
transformating the X as independent as possible in the<br />
case m≥ n , that is<br />
H<br />
Y = W X<br />
(4)<br />
where, Y is the separated vector signal, that is, the<br />
estimated value <strong>of</strong> source signal vector S .<br />
Complex FastICA algorithm is a fixed-point algorithm<br />
using the batch processing. Compare with the ordinary<br />
ICA algorithm its convergence speed is more quickly. In<br />
this paper, we adopt complex FastICA algorithm in the<br />
pre-processing, in order to separate the multi-component<br />
LFM signals <strong>of</strong> low SNR in time-domain.<br />
2)Time-domain separation technique <strong>of</strong> multi-component<br />
LFM signals with low SNR based on complex FastICA<br />
Multi-component LFM time-domain separation<br />
technique <strong>of</strong> basic complex FastICA algorithm includes<br />
two steps: first, preprocess the chosen mixed-signal X<br />
which was composed <strong>of</strong> multi-component LFM signals<br />
and noise, that is, using whitening treatment. Whitening<br />
treatment can be used to remove the correlation between<br />
signals, which simplifies the process <strong>of</strong> follow-up to the<br />
extraction <strong>of</strong> independent component, second is the<br />
extraction <strong>of</strong> independent components, namely, the<br />
completion <strong>of</strong> the mixed signals separation. Complex<br />
FastICA algorithm flow chart is shown in Fig.2.<br />
Figure 2. Algorithm flow chart <strong>of</strong> complex FastICA<br />
Complex FastICA is a fast optimization iterative<br />
algorithm using the batch approach which has a large<br />
number <strong>of</strong> samples <strong>of</strong> data involved in computation in<br />
each iteration. According to the central limit theorem,<br />
linear sum <strong>of</strong> a number <strong>of</strong> independent random variables<br />
will tend to Gaussian distribution, so complex FastICA<br />
mainly achieve the purpose <strong>of</strong> separation by measuring
1978 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
the largest non-Gaussian. For complex random vector y ,<br />
its kurtosis is defined as<br />
4 * * * * * *<br />
kurt( y) = E{ y } −E{ yy } E{ yy } −E{ yy} E{ y y } − E{ yy } E{ y y}<br />
(5)<br />
As<br />
kurtosis can be converted to[2]<br />
*<br />
y is y conjugate transpose, the definition <strong>of</strong><br />
4 2 2 2<br />
kurt( y) = E{ y } −2( E{ y }) − E{ y }<br />
4<br />
= E{ y } − 2<br />
(6)<br />
where y is white, i.e., the real and imaginary parts <strong>of</strong><br />
y are uncorrelated and their variances are equal.<br />
We usually use some other suitable non-linear<br />
function Gy ( ) instead <strong>of</strong> y for (6), which makes the<br />
convergence calculation <strong>of</strong> kurtosis is more robust. The<br />
expectation function <strong>of</strong> separation matrix is expressed as<br />
2<br />
H<br />
2<br />
J ( W) = E{ G( Y )} = E{ G(<br />
W X )} (7)<br />
G<br />
1<br />
where G ( y) = log( a + y)<br />
, g ( y)<br />
= was choosen<br />
a+ y<br />
as non-linear function[3,15]. Where a is arbitrary constant<br />
for which values a ≈ 0.1 were chosen in this work.<br />
Because the above non-linear function give more robust<br />
estimators, g ( y) is derivative <strong>of</strong> G ( y ) .<br />
Now give the fixed-point algorithm for complex<br />
signals under the ICA data model(3). In this paper ,we<br />
obtain separation matrix W , which makes the separation<br />
H<br />
components Y = W X , so the estimation Y <strong>of</strong> source<br />
independent component were obtained. The need for<br />
preprocessing to reduce the difficulty <strong>of</strong> analysis due to<br />
less known information in blind separation. So the<br />
whitening process <strong>of</strong> the observation signal will greatly<br />
simplify the analysis difficulty. Firstly, centralize the<br />
mixed-signal X <strong>of</strong> multi-component LFM signal and noise,<br />
that is X = X−E{ X}<br />
so that the mean <strong>of</strong> X is 0. Then we<br />
can obtain zero mean vector X by observational data and<br />
whitening matrix Q ,i.e. X = Q Xold<br />
,<br />
H<br />
X = ( x + ix ,..., x + ix<br />
) ,and therefore E { XX } = I .<br />
1r1i nr ni<br />
Whitening can be accomplished by principal component<br />
analysis (PCA).<br />
The complex FastICA algorithm searches for the<br />
2<br />
extrema <strong>of</strong> EG { ( )}<br />
H w X .Details <strong>of</strong> the derivation are<br />
presented in the appendix[3] . Supposing the separation<br />
matrix W , first select an initial separation vector w<br />
(random). The fixed-point algorithm for one unit is<br />
+ H * H<br />
2<br />
H<br />
2<br />
w = E{( x w x) g( w x )} −E{(<br />
g w x )<br />
w<br />
H<br />
2<br />
H<br />
2<br />
'<br />
+ w x g ( w x )} w (8)<br />
new<br />
=<br />
w<br />
+<br />
+<br />
w (9)<br />
The one-unit algorithm can be extended to the<br />
H<br />
estimation <strong>of</strong> whole ICA transformation S = W X .To<br />
prevent different neurons from converging to the same<br />
© 2011 ACADEMY PUBLISHER<br />
2<br />
H H<br />
maxima, the outputs W X,..., X<br />
1 W are decorrelated after<br />
n<br />
every iteration. A simple way to accomplish this is a<br />
deflation scheme based on a Gram-Schmidt-like<br />
decorrelation: When we have estimated p independent<br />
components, or p vectors w ,..., w ,we run the one-unit<br />
1 p<br />
fixed-point algorithm for w new ,and after every iteration<br />
step subtract from wnew the projections <strong>of</strong> the previously<br />
estimated p vectors, and then renormalize wnew as follows<br />
=<br />
p−1<br />
−∑<br />
j = 1<br />
j j pnew<br />
H<br />
w w w w w (10)<br />
w<br />
pnew p<br />
pnew<br />
= w<br />
w<br />
pnew<br />
pnew<br />
(11)<br />
where w ( j = 1,..., p−1)<br />
is previous p − 1 separation<br />
j<br />
vector, wnew denotes the p new value <strong>of</strong> separation vector.<br />
Determining whether w is convergence. If not<br />
pnew<br />
convergence, the w obtained by (9) instead <strong>of</strong> the w in<br />
pnew<br />
(9) and instead <strong>of</strong> the w in (10) up to the time when<br />
p<br />
w convergence, therefore the p separation vector is<br />
pnew<br />
obtained. Sometimes it is preferable to estimate all the<br />
independent components simultaneously, and use a<br />
symmetric decorrelation. This can be accomplished e.g.,<br />
by<br />
1/ 2<br />
( ) − H<br />
w = w w w (12)<br />
where W = ( W,..., W ) is the matrix <strong>of</strong> the vectors.<br />
1<br />
n<br />
At this time we can get mixing matrix Α and separation<br />
matrixW . Then the separation signal y1, y2, ⋅⋅⋅,<br />
yncan be<br />
H<br />
calculated according to Y = W X . Until now, we have<br />
completed the mixed-signal time-domain separation<br />
process <strong>of</strong> multi -component LMF signals with<br />
noise[10,11].<br />
B.The automatic identification method based on second<br />
central moment <strong>of</strong> FRFT<br />
Multi-component LFM emitter signals after blind<br />
source separation, the signal has a random arrangement,<br />
in the case <strong>of</strong> unknown a priori information can not<br />
determine which way is the signal or noise. If use the<br />
FRFT <strong>of</strong> noise to estimate the parameter <strong>of</strong> LFM signal,<br />
the measuring results will be wrong. To solve the<br />
problem, we use the second central moment <strong>of</strong> FRFT<br />
method to achieve the above separated signal and noise<br />
auto-discrimination.<br />
FRFT is the promotion <strong>of</strong> the traditional Fourier<br />
transform. In recent years, it has attracted increasing<br />
attention in signal processing field. Setting one <strong>of</strong> the<br />
output signals after above-mentioned complex FastICA<br />
algorithm is yi() t . Its FRFT is defined as<br />
X ( u) = ∫ y ( t) K ( t, u) dt<br />
(13)<br />
α i α
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1979<br />
where K (, tu)<br />
is kernel function <strong>of</strong> FRFT, is defined as<br />
α<br />
2 2<br />
⎧ 1−jcotαt + u ut<br />
⎪ exp( j cot α − j ), α ≠nπ<br />
⎪ 2π 2 sinα<br />
⎪<br />
Kα(, t u) = δ( t− u), α = 2nπ<br />
(14)<br />
⎨<br />
⎪ δ( t+ u), α = (2n+ 1) π<br />
⎪<br />
⎪⎩<br />
where α is rotation angle. With a view <strong>of</strong> timefrequency<br />
plane rotation to explain, then the following<br />
equations are established<br />
0<br />
R ( u) = y ( t)<br />
yi π<br />
i<br />
yii R ( u) = y ( − t)<br />
π /2<br />
R ( u) = FT( y ( t))<br />
(15)<br />
yii π /2<br />
where Ry( u)<br />
corresponds to the FT <strong>of</strong> signal y ()<br />
i<br />
i t .<br />
The traditional FT can be seen as the time-frequency<br />
distribution <strong>of</strong> signal in the projection <strong>of</strong> frequency axis,<br />
while FRFT can be seen as the time-frequency<br />
distribution <strong>of</strong> signal in the projection <strong>of</strong> the rotated<br />
frequency axis. The representation <strong>of</strong> signal in the<br />
fractional Fourier domain includes both the time domain<br />
and frequency domain information, so FRFT is also<br />
considered a generalized time-frequency analysis[13,14].<br />
By the definition <strong>of</strong> FRFT, a LFM signal only at the<br />
appropriate fractional domain is an impulse function.<br />
Therefore, FRFT in a fractional Fourier domain has the<br />
best gathering characteristics for the given LFM signal. In<br />
the time-frequency plane, a limited length <strong>of</strong> the LFM<br />
signal appears as the distribution <strong>of</strong> dorsal fin shape <strong>of</strong><br />
diagonal line, but FRFT is essentially the "rotating" <strong>of</strong><br />
signal. If choose the appropriate rotation angle, it will<br />
show the energy aggregation and apparent peak in the<br />
fractional Fourier domain <strong>of</strong> signal. It was shown in Fig.3.<br />
| Xp( u)|<br />
Figure 3 .The distribution <strong>of</strong> time-frequency and in the projection <strong>of</strong><br />
fractional Fourier domain <strong>of</strong> LFM signals<br />
The bandwidth <strong>of</strong> signal in time domain and frequency<br />
domain can be estimated by the second-order central<br />
moments[15], and the bandwidth <strong>of</strong> signal in the<br />
fractional Fourier domain can be obtained by the secondorder<br />
central moments <strong>of</strong> FRFT[16]. The second-order<br />
central moments(SCM) <strong>of</strong> FRFT Pα is defined as<br />
∞<br />
2<br />
α<br />
2<br />
Pα = ∫ Ry( u) ( u−m ) du<br />
−∞ i<br />
α<br />
(16)<br />
© 2011 ACADEMY PUBLISHER<br />
where<br />
α<br />
Ry( u)<br />
i<br />
is FRFT <strong>of</strong> yi() t , mα is first-order<br />
moments <strong>of</strong> FRFT<br />
m<br />
∞<br />
= ∫<br />
2<br />
α<br />
R ( u) udu<br />
(17)<br />
α<br />
−∞<br />
yi<br />
As FRFT is a periodic function with the period <strong>of</strong><br />
α+ π α<br />
2π about α , and meet R ( u) = R ( − u)<br />
, so the<br />
yi yi<br />
second-order central moments <strong>of</strong> FRFT Pα has a<br />
maximum or minimum value in the range <strong>of</strong> α ∈ [0, π ) .<br />
As Pα represents the bandwidth <strong>of</strong> signal in the fractional<br />
Fourier domain, when the rotation angle <strong>of</strong> timefrequency<br />
planeα = αe<br />
, the bandwidth has a minimum.<br />
We can find spindle direction <strong>of</strong> time-frequency<br />
distribution α by searching the minimum point <strong>of</strong> Pα ,<br />
namely, the best fractional Fourier transform domain. The<br />
bandwidth <strong>of</strong> noise is wide in the fractional Fourier<br />
transform domain, α = αe<br />
corresponds to the minimum <strong>of</strong><br />
bandwidth (the minimum <strong>of</strong> FRFT) also very large,<br />
namely, Pα = α corresponds to the minimum. So we can<br />
e<br />
determine signal or noise by the bandwidth <strong>of</strong> fractional<br />
Fourier transform domain.<br />
The noise can be removed from the separated signals<br />
after the LFM signal and noise discrimination method<br />
based on second-order central moments <strong>of</strong> FRFT. Then<br />
only detect the remaining LFM signal.<br />
C. FRFT detection for LFM signal<br />
As shown in Fig.3, the observed signal was<br />
continuously proceed FRFT for rotation angle variable<br />
α , the two-dimensional distribution <strong>of</strong> signal energy was<br />
formed in the parameter ( α , u)<br />
plane [14]. And the<br />
detection and parameter estimation <strong>of</strong> LFM signals can<br />
be realized by two-dimensional search <strong>of</strong> peak point<br />
threshold in this plane. For type (1), the process <strong>of</strong> this<br />
model can be described as[16]<br />
∧ ∧<br />
2<br />
{ α0, u0} = arg max X ( u)<br />
α<br />
α , u<br />
(18)<br />
⎧<br />
⎪<br />
⎪<br />
∧ ∧<br />
⎪<br />
µ 0 =−cot<br />
α 0,<br />
⎪<br />
⎪ ∧ ∧ ∧<br />
⎨ f0 = u0<br />
csc α 0,<br />
⎪<br />
∧<br />
⎪ X ∧ ( u0<br />
)<br />
∧<br />
⎪<br />
α0<br />
⎪<br />
Ai<br />
=<br />
⎪ ∆tA<br />
∧<br />
⎩<br />
α0<br />
(19)<br />
IV. SIMULATION VERIFICATION<br />
Select a group <strong>of</strong> mixed-signal in order to verify the<br />
validity <strong>of</strong> the method proposed in this paper. Mixedsignal<br />
composed <strong>of</strong> two-component LMF signal and<br />
noise, the first LFM signal is<br />
2<br />
− j5π t<br />
x1= e (initial<br />
f = 0 and chirp rate k 10 =− 10 ), the second is<br />
frequency 10
1980 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
x2 2<br />
− j 2π<br />
t<br />
0.5* e<br />
f 20 = 0 and chirp rate 20 4<br />
weak LFM signal<br />
= (initial frequency<br />
k =− ), the noise is Gaussian<br />
white noise, the SNR <strong>of</strong> the weakest component LMF<br />
signal(the second LFM signal and noise ratio) is SNR=-<br />
15dB. Nonlinear function selected G ( y) = log( a + y)<br />
,<br />
here a ≈ 0.1 , sampling point N=1601, the order <strong>of</strong><br />
FRFT is 0< p < 2 and p = 2 α / π .<br />
Using the new method to simulate and analyse the<br />
above assumed mixed-signal. The multi-component<br />
signals with noise were separated by complex FastICA<br />
algorithm. The time-domain separation results were<br />
shown in Fig.4(a-b).<br />
(a) The results <strong>of</strong> mixed-signal by the complex FastICA (taking the real<br />
part <strong>of</strong> the signal)<br />
(b)The convergence <strong>of</strong> complex FastICA algorithm for Multicomponent<br />
LFM emitter signals<br />
Figure.4. Multi-component LFM emitter signals separated by complex<br />
FastICA<br />
As can be seen from Fig.4(a), the effect <strong>of</strong> signal<br />
separation was very good in the low SNR(-15dB) by the<br />
complex ICA algorithm. Fig.4(b) shows the convergence<br />
<strong>of</strong> the fixed-point algorithm using contrast function<br />
G( y) = log( a+ y)<br />
, average result over ten runs. About six<br />
iteration steps were needs for convergence. Fig.5 (a-b)<br />
shows the second-order central moments <strong>of</strong> FRFT <strong>of</strong> the<br />
each separated signal( 0< p < 2,<br />
α ∈ [0, π ) ).<br />
© 2011 ACADEMY PUBLISHER<br />
(a) Second-order central moments <strong>of</strong> time-domain separated output<br />
signal 1<br />
(b) Second-order central moments <strong>of</strong> time-domain separated output<br />
signal 2<br />
(c) Second-order central moments <strong>of</strong> time-domain separated output<br />
signal 3<br />
Figure.5. Second-order central moments <strong>of</strong> FRFT <strong>of</strong> the each separated<br />
signal<br />
According to (16) , the change results <strong>of</strong> second-order<br />
central moments P α with FRFT angle α were shown in<br />
Fig.5 (a-c) . The minimum <strong>of</strong> second-order central<br />
moments (ie, bandwidth <strong>of</strong> signal) <strong>of</strong> three separated<br />
5<br />
signals is respectively Pα 1 = 4.671*10 ,<br />
6<br />
6<br />
Pα 2 = 1.2001*10 and Pα 3 = 1.3201*10 by computer<br />
calculation. So the third signal is Gaussian white noise<br />
which to be filtered. Next only the remaining two LFM<br />
signals were detected by FRFT. Fig.6(a-b) shows the<br />
FRFT time-frequency map <strong>of</strong> remaining two signals.<br />
Fig.7 shows the time-frequency map <strong>of</strong> traditional FRFT
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1981<br />
<strong>of</strong> the same group <strong>of</strong> mixed-signal without the complex<br />
FastICA(SNR=-15dB).<br />
(a) The FRFT <strong>of</strong> one LFM signal after automatic identification method<br />
(b) The FRFT <strong>of</strong> the other LFM signal after automatic identification<br />
method<br />
Figure.6. The FRFT time-freqency map <strong>of</strong> remaining two signals after<br />
automatic identification method<br />
Figure.7. Distribution <strong>of</strong> traditional FRFT <strong>of</strong> the same group <strong>of</strong> mixedsignal<br />
without the complex FastICA algorithm(SNR=-15dB)<br />
From Fig.6 and Fig.7 can be seen in SNR=-15dB,<br />
when the mixed-signal was processed by the new method,<br />
the FRFT distribution <strong>of</strong> the separated signal was<br />
influenced by noise slightly and cross-term has also been<br />
suppressed. However, when the same mixed-signal was<br />
processed by traditional FRFT, the signals were<br />
influenced by noise and cross-term largely, weak signal<br />
has been drowned by the noise and the noise would cause<br />
great difficulties in the extraction and detection <strong>of</strong><br />
signals. Especially,the weak signal is more affected by<br />
it[13].<br />
© 2011 ACADEMY PUBLISHER<br />
Using the method described in section (C) to<br />
estimate the parameters. Parameter estimation results in<br />
the following Tab.Ⅰ:<br />
Table Ⅰ<br />
PARAMETER ESTIMATION RESULTS OF TWO METHODS<br />
SNR=-<br />
15dB<br />
Signal<br />
1<br />
Signal<br />
2<br />
Real Value<br />
10<br />
The new method<br />
(FastICA&SCM&<br />
FRFT)<br />
Test results<br />
Traditional<br />
method<br />
(FRFT)<br />
f =0 0.0012 0.0102<br />
10 =-5 µ<br />
-5.00757 -5.9894<br />
f 20 =0 0.0007 2.0691<br />
20 =-2 µ -2.00133 -3.311<br />
Above the table can be seen the new method can<br />
correctly estimate the parameters <strong>of</strong> the LFM signal in<br />
SNR=-15dB. For the signal 1, the relative error <strong>of</strong> the<br />
estimate value was η f = 0.24% and η 0.342%<br />
10<br />
µ = (the<br />
10<br />
relative error <strong>of</strong> the estimate value was expressed as<br />
η f and η µ ). For the signal 2, the relative error <strong>of</strong> the<br />
estimate value was η f = 0.35% and η 0.356%<br />
20<br />
µ = .<br />
20<br />
However, the traditional FRFT method is no longer<br />
correctly estimate the signal parameters in SNR=-15dB.<br />
V. CONCLUSION<br />
To effectively detect and recognize multi-component<br />
LFM signals in low SNR, a new multi-component LFM<br />
signals analysis method which was based on the complex<br />
FastICA&FRFT was proposed. First, multi-component<br />
LFM signals were processed by complex ICA to obtain<br />
the time domain separate signals in low SNR. Second, the<br />
time domain separated signals were respectively<br />
discriminated by automatic identification method based<br />
on second central moment <strong>of</strong> FRFT. Then the LFM<br />
signals were processed by the FRFT. In this paper, the<br />
new method was compared with the traditional FRFT to<br />
prove the validity <strong>of</strong> the new method. The simulation<br />
results show that the new method can effective analysis<br />
multi-component LFM emitter signals in low SNR.<br />
ACKNOWLEDGMENT<br />
We thank the National Natural Science Foundation<br />
Project (No.:60872108), China Postdoctoral Science<br />
Foundation Special Support Project (No.:200902411) ,<br />
the financial support from China Postdoctoral Science<br />
Foundation (No.:20080430903), Heilongjiang Postdoctor<br />
Financial Assistance (LBH-Z08129), the Scientific and<br />
Technological Creative Talents Special Research<br />
Foundation <strong>of</strong> Harbin Municipality (2008RFQXG030)<br />
and Central University Basic Research Pr<strong>of</strong>essional<br />
Expenses Special Fund Project (No.:HEUCFZ1015) for<br />
this paper support. We thank members <strong>of</strong> College <strong>of</strong><br />
Information and Communication Engineering, Harbin
1982 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
Engineering University and Space Control and Inertial<br />
Technology Research Center, Harbin Institute <strong>of</strong><br />
Technology for technical support.<br />
REFERENCES<br />
[1] Liu feng, Sun dapeng, Huang yu, Tao ran, Wang yue,in<br />
:“Multi-component LFM signal feature extraction based on<br />
improved Wigner-Hough transform,” <strong>Journal</strong> <strong>of</strong> Beijing<br />
Technology University.(2008.10)<br />
[2] Ashok Narayanan V, Prabhu K M M. The fractional<br />
fourier transform:theory, implementation and error<br />
analysis[J]. Elsevier Micorprocessors and Microsystems,<br />
Vol.27(2003),p.511-521<br />
[3] Ella Blngham and Aapo Hyvarinen. “A Fast Fixed-Point<br />
Algorithm For Independent Component Analysis Of<br />
Complex Valued Signals”. International <strong>Journal</strong> <strong>of</strong> Neural<br />
Systems, Vol.10, No.1(February,2000),p.1-8<br />
[4] J.Herault and C.Jutten.Blind separation <strong>of</strong> sources, part: an<br />
adaptive algorithm based on neuro mimetic. Signal<br />
Processing, Vol.24(1)(1991),p.1-10.<br />
[5] LIU Q S , LU H Q , MA S D, “A Non-parameter Bayesian<br />
Classifier for Face Recognition [J] ,”<strong>Journal</strong> <strong>of</strong> Electronics<br />
(China), Vol.20(5)(2003),p.362 -370.<br />
[6] Shimizu S., Hyvarinen A., Kano Y.. A generalized least<br />
squares approach to blind separation <strong>of</strong> sources which have<br />
variance dependencies[J].Statistical Signal Processing,<br />
IEEE/SP 13th Workshop on(2005),p.1080-1083<br />
[7] Tachibana K., Saruwatari H., Mori Y.. Efficient Blind<br />
Source Separation Combining Closed-Form Second-Order<br />
ICA and Nonclosed-Form Higher-Order ICA. IEEE<br />
International Conference on Acoustics,Speech and Signal<br />
Processing. ICASSP 2007. Vol. 1(2007), p.I-45-I-48<br />
[8] Chee-Ming Ting, Salleh S.-H., Zainuddin Z.Z.. Spectral<br />
Estimation <strong>of</strong> Nonstationary EEG Using Particle Filtering<br />
With Application to Event-Related Desynchronization<br />
(ERD) [J]. IEEE Transactions on Biomedical Engineering.<br />
Vol. 58(2011) p.321-331<br />
[9] Zou Hong-xing, LU Xu-guang, DAI Qiong-hai.<br />
Nonexistence <strong>of</strong> cross-term free time-frequency<br />
distribution with concentration <strong>of</strong> Wigner-ville<br />
distribution, Vol.3(2002)<br />
[10] Yuan junquan,Sun minqi,Sun xiaoxu, “LFM signal<br />
parameters estimation method based on Wigner Hough<br />
Transform”,Aerospace Electronic Countermeasures, Vol.6<br />
(2004).<br />
[11] Solvang H.K.,Nagahara Y.,Araki S.. Frequency-Domain<br />
Pearson Distribution Approach for Independent<br />
Component Analysis (FD-Pearson-ICA) in Blind Source<br />
Separation[J]. IEEE Transactions on Audio,Speech,and<br />
Language Processing . Vol.17,No.4(2009),p.:639-648<br />
[12] Liu ju, He zhenya, Zhang xianda. Blind Source Separation<br />
and Blind Deconvolution. Electronics <strong>Journal</strong>, Vol.30(4)<br />
(2002),p.570-576<br />
[13] Li xiaoju,Zhu xiaolong,Zhang xianda. Blind source<br />
separation classification and prospects. <strong>Journal</strong> <strong>of</strong> Xi'an<br />
University <strong>of</strong> Electronic Science and Technology,<br />
Vol.31(3) (2004),p.399–404<br />
[14] Zhang xianda, Bao zheng. Blind signal separation. E-<br />
<strong>Journal</strong>. Vol.29(12)( 2001),p.1766-1771.<br />
[15] Zou hong. Time-frequency analysis <strong>of</strong> multi-component<br />
LFM signals [D]. Xi'an Electronic Science and Technology<br />
University, 2000.<br />
[16] Liu Jiancheng, Wang Xuesong, Xiao Shunping, et a1.<br />
Radial acceleration estimation based on Wigner-Hough<br />
transform[J]. Acta Electronica Sinica, Vol.33(12)<br />
(2005),p.2236-2238.<br />
© 2011 ACADEMY PUBLISHER<br />
Qiang Guo was born in 1972. He<br />
received the B.S., M.S. and Ph.D. degree<br />
from Harbin Engineering University in<br />
information and communication<br />
engineering in 1994, 2003, and 2007,<br />
respectively. He is now an associate<br />
pr<strong>of</strong>essor <strong>of</strong> information and<br />
communication engineering at Harbin<br />
Engineering University and a postdoctoral<br />
fellow <strong>of</strong> control science and engineering<br />
at Harbin Institute <strong>of</strong> Technology (HIT), China. His current<br />
research interests include radar signals sorting and recognition.<br />
More complex and dense pulses environments in modern<br />
electronic warfare present a severe challenge to the problem <strong>of</strong><br />
radar signal sorting. Based on fractal theory and Hilbert-Huang<br />
Transform (HHT), he presented a new feature extraction<br />
method for radar pulse signal sorting. It used structure function<br />
and empirical mode decomposition to process 2-dimension<br />
feature information, which constituted carrier frequency and<br />
time-<strong>of</strong>-arrival. The same scheme also applied to the analysis<br />
and extraction <strong>of</strong> hidden periodically changing features—G<br />
features. Experiment results show that the method can<br />
effectively identify the agile frequency in periodically changing<br />
radio frequency signals <strong>of</strong> complex pulse environment,<br />
therefore provides a new feature for signal sorting <strong>of</strong> interleaved<br />
radar pulse sequences.<br />
He received the national 100 excellent doctor degree<br />
dissertation candidate nomination in 2009. He is now Academic<br />
Degree & Graduate Education Evaluating expert <strong>of</strong> MOE.<br />
Yajun Li was born in 1983. He<br />
respectively received the B.S., M.S. degree<br />
from YaTai University and Harbin<br />
Engineering University in information and<br />
communication engineering in 2008, 2011.<br />
He is now an Ph.D. at Harbin Institute <strong>of</strong><br />
Technology (HIT), China.<br />
His current research interests include<br />
radar signals sorting and detection. At<br />
present, he has already published seven articles(EI index).<br />
Changhong Wang was born in 1961. He<br />
received the B.S. the M.S. and Ph.D. degree<br />
from Harbin Institute <strong>of</strong> Technology in<br />
1983, 1986, and 1991, respectively. He is<br />
currently a pr<strong>of</strong>essor and Ph.D. student<br />
supervisor <strong>of</strong> Harbin Institute <strong>of</strong><br />
Technology.<br />
His research interests are mainly in<br />
inertial navigation, precise servo control<br />
system, and robust control.
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1983<br />
Research on Self-built Digital Resource Backup<br />
Systems<br />
Abstract – This paper discussed the characteristics <strong>of</strong> the<br />
self-built digital resources and the requirement for<br />
long-term preservation. A backup system for self-built<br />
digital resources has been proposed based on the s<strong>of</strong>tware<br />
and hardware features <strong>of</strong> the resources. Furthermore,<br />
simple analysis has been carried out on the proposed<br />
system.<br />
Index Terms – self-built digital resource, backup system,<br />
long-term preservation, data duplication elimination<br />
I. INTRODUCTION<br />
Digitalized preservation, organization and sharing are<br />
very important in digital resources construction.<br />
Self-built digital resources is an important part in digital<br />
resources construction. Self-built digital resources <strong>of</strong>ten<br />
appear in the form <strong>of</strong> feature databases self-created by<br />
libraries, for example, numerous well developed feature<br />
databases, dissertation databases and textbook databases<br />
are planned in library projects. Furthermore, different<br />
schools and universities also construct their own<br />
databases based on their discipline and regional<br />
characteristics, technical specialty, as well as financial<br />
budget [1].<br />
In order to preserve self-built digital resources for long<br />
time, two aspects have to be considered, namely, how to<br />
prevent unauthorized modification and breach to digital<br />
information, and how to maintain long-term readability<br />
and authenticity <strong>of</strong> digital information. Technology is<br />
readily available to tackle the first problem, as a number<br />
<strong>of</strong> mature techniques have been proposed world widely to<br />
prevent illegal modification and breach <strong>of</strong> digital<br />
information; therefore, it is possible to solve the first<br />
problem to some extent if technical measures can be<br />
scientifically integrated with management practice.<br />
However, how to effectively maintain long-term<br />
readability <strong>of</strong> the digital information is still an open<br />
research area, no perfect solution has been proposed so<br />
far. The major difficulties lie on the deep involvement <strong>of</strong><br />
numerous issues in which the most important one is the<br />
adoption <strong>of</strong> standards. Adopting standards can ease the<br />
conflicts between the technological update and<br />
readability <strong>of</strong> digital information. Nevertheless, problems<br />
still remain as some standards, particularly industrial<br />
standards are commonly outdated; and it is also difficult<br />
to completely comply with standards in practice.<br />
Currently, techniques used for long term preservation<br />
© 2011 ACADEMY PUBLISHER<br />
doi:10.4304/jcp.6.9.1983-1987<br />
Li-zhen Shen<br />
Wenzhou University, Wenzhou, Zhejiang, China<br />
a wzu-slz@wzu.edu.cn<br />
include migration technique, updating technique,<br />
conversion technique, simulation technology, and<br />
digitizing technique using graphic tablets, etc. [3].<br />
II. CHARACTERISTICS OF BACKUP SYSTEMS FOR<br />
SELF-BUILT DIGITAL RESOURCES<br />
Self-built digital resources are diverse, including such<br />
resource types as WEB resources, electronic publications,<br />
scientific data, multimedia resources and electronic<br />
dissertations, etc. Furthermore, all self-built digital<br />
resources use internet to provide resource services,<br />
therefore, in addition to fully back up the content <strong>of</strong><br />
databases, it is <strong>of</strong>ten required to backup server systems<br />
and data publishing environment (both will be referred as<br />
servers in the sequel). The servers which need to be<br />
backed up may run multiple operating systems such as<br />
SUN Solaris, LINUX, Windows NT, and Windows 2000,<br />
and some may have Micros<strong>of</strong>t SQL Server 2000 database<br />
and ORACLE running on them. Considering all these<br />
aspects, backup systems should have the following<br />
functions:<br />
Backup across operating systems. Backup systems<br />
should support data backup and recovery across different<br />
operating systems such as Micros<strong>of</strong>t Windows, Unix, and<br />
IBM Aix, that is, a backup server can back up data from<br />
multiple operating platforms, thus reduce operational<br />
complexity and lower total cost <strong>of</strong> the backup work.<br />
Automatic Backup. Backup uses system resources. In<br />
practice, a running backup job may take 60% <strong>of</strong> the CPU<br />
resource <strong>of</strong> a mini-computer server with average<br />
configuration. Besides, backup jobs will also occupy<br />
network bandwidth as well as other resources. Therefore,<br />
backups should be performed when the load <strong>of</strong> the<br />
servers is minimal, should avoid casting extra load on the<br />
servers in the peak hours. Obviously, it is essential to use<br />
unattended, automated backup systems to avoid the<br />
human interactions during the backup time which, most<br />
likely, happens at late nights or in public holidays.<br />
Support multiple backup strategies. The famous<br />
Pareto principle holds in the backup area, namely, 20% <strong>of</strong><br />
the data is updated more frequently with a back up<br />
probability <strong>of</strong> 80%. If every time a full backup is<br />
performed, it will inevitably waste resources and time in<br />
some cases, thus full backup is sometimes not viable.<br />
What we need the daily backup to do is to backup the<br />
delta <strong>of</strong> two consequential full backups. Therefore, we<br />
should adopt the so-called incremental backup strategy
1984 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
and a combination <strong>of</strong> several other backup strategies.<br />
Meanwhile, it is necessary to consider the requirement <strong>of</strong><br />
long-term preservation when backing up data.<br />
Efficient and safe recovery. The fundamental purpose<br />
<strong>of</strong> backup is for recovery, a backup which cannot be<br />
restored is meaningless. An important factor which the<br />
end users will use to determine the quality <strong>of</strong> a backup<br />
system is whether the system can restore the backed up<br />
data in a safe, convenient and efficient way.<br />
Easy upgrading. It is necessary to consider possible<br />
future extension <strong>of</strong> functions when designing the system<br />
initially. For example, the designed system should be able<br />
to support database online backup, and should be easy to<br />
add functions on the client-side.<br />
Long-term preservation. Long-term preservation <strong>of</strong><br />
electronic resources is an important task for library<br />
resources construction in the new information<br />
environment. Long-term preservation is not only a new<br />
mission for the libraries, but also a major challenge as<br />
many technical, economic, legal and other problems<br />
emerge. Regardless <strong>of</strong> changes in the external<br />
environment, it is an essential characteristic for modern<br />
backup systems to effectively preserve data over long<br />
term, and to guarantee readability <strong>of</strong> the preserved data at<br />
any time.<br />
Based on the analysis <strong>of</strong> requirement for self-built<br />
backup system resources, and considering multiple<br />
factors such as unified backup management and support<br />
for future storage infrastructure, we propose a backup<br />
system which will be described in the sequel.<br />
A. Backup system structure<br />
© 2011 ACADEMY PUBLISHER<br />
III. DESIGN AND ANALYSIS<br />
Fig. 1 A self-built backup system for digital resources<br />
The proposed backup system includes self-built<br />
featured database reservoir, primary storage, application<br />
server farms, backup / media servers, file management<br />
application servers, and remote archive storage for<br />
disaster recovery.<br />
The backup system works as follows: A server is<br />
configured as a backup server which is responsible for<br />
system backup operation; a large capacity backup storage<br />
device, which consists <strong>of</strong> low-end storage, tape drive or<br />
tape array, is connected to the backup server. Other<br />
servers within the network which may need to have<br />
managed data backup will run backup client s<strong>of</strong>tware<br />
which enable centralized data backup via LAN to the<br />
primary backup storage device connected to the backup<br />
server. Prior to backup operation, digital resources are<br />
classified as backup type and archive type. Furthermore,<br />
data which is classified as backup type will go through<br />
duplication elimination equipment or s<strong>of</strong>tware to further<br />
reduce size before performing actual backup operation.<br />
Meanwhile, a comprehensive backup plan and associated<br />
backup strategy will be established using the planning<br />
functions <strong>of</strong> the backup s<strong>of</strong>tware, and all data will be<br />
backed up through centralized management. It should be<br />
noted that the local backup storage is in fact the primary<br />
storage in our case, and the primary storage will map a<br />
copy in the remote archive storage, then use the backup<br />
s<strong>of</strong>tware to provide safe disaster recovery measures. The<br />
proposed backup system can greatly shorten time<br />
required to perform backup and disaster recovery, and is<br />
capable to achieve high security and usability for<br />
network-based data backup. Structure <strong>of</strong> the proposed<br />
backup system for self-built resources is shown in Fig. 1.
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1985<br />
B. System Backup<br />
Server backup. Self-built featured library consists <strong>of</strong><br />
several processing computers and a processing<br />
management server. In order to ensure the smooth<br />
operation <strong>of</strong> the processing work, the operating system <strong>of</strong><br />
all processing computers is backed up to a compact disc<br />
media. In general, all the processing computers are the<br />
same model and were purchased in the same batch, hence<br />
mo individual backup <strong>of</strong> the operating system <strong>of</strong> each<br />
processing computer is required. It is only necessary to<br />
do an individual backup for the process management<br />
server. Using server virtualization techniques such as<br />
virtual machines, the processing management server and<br />
the application server cluster which is responsible for the<br />
web database publishing and reader service can not<br />
only achieve highly efficient server performance, but also<br />
perform file backup on the server operating system. In<br />
addition, restore is simple and fast in this case.<br />
Data backup. In the construction <strong>of</strong> self-built<br />
databases, the originally sampled or collected data need<br />
to be backed up, and the final data production also has to<br />
be backed up. However, the original and the final data<br />
types bear different usage frequency and lifecycle, hence,<br />
the archive backup <strong>of</strong> the original data <strong>of</strong>ten happens<br />
after the completion <strong>of</strong> the processing operation, while<br />
the product data is backed up afterwards.<br />
© 2011 ACADEMY PUBLISHER<br />
Fig.2 the Backup data flow diagram<br />
Implementation <strong>of</strong> long-term preservation. Long-<br />
term data preservation is expected in the proposed backup<br />
system, therefore, migration and simulation techniques<br />
are employed to explore the characteristics <strong>of</strong> the<br />
self-built data resources in the archiving process.<br />
Specifically, archiving data and files used for reading<br />
environment are archived to the tape storage using the<br />
media server, and the archived data is regularly restored<br />
to verify its readability.<br />
Duplication elimination. Many backup strategies<br />
can reduce the backed up data size, however, the results<br />
are still not satisfactory. On the other hand, duplication<br />
elimination techniques can achieve a data compression<br />
ratio <strong>of</strong> 1:20, therefore, in the backup system, it is viable<br />
to use data duplication elimination combined with<br />
incremental backup to greatly reduce the storage space<br />
required for a whole system backup. Namely, prior to<br />
backup, data files are divided into several blocks <strong>of</strong> data<br />
to store. In principle, the same data block in different data<br />
files is backed up only once, thus significantly eliminate<br />
duplication in data, and reduce data redundancy[4-5] .<br />
IV. BACKUP DATA FLOW DIAGRAM<br />
A backup data flow diagram is shown in Fig. 2 which<br />
describes in detail how to back up the digital library data.
1986 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
V. REMOTE BACKUP DATA STORAGE FOR DISASTER<br />
RECOVERY<br />
Unlike commercial data such as bank transactions, data<br />
in digital library is not necessary to be error-free, hence<br />
the main task for remote backup is to guarantee the<br />
consistency between the remote recovery data and the<br />
existing data in the local environment. Therefore, the<br />
proposed backup system uses the idle time period such as<br />
the period from 24:00 to 08:00, and employs<br />
asynchronous PPRC (Peer to Peer Remote Copy)<br />
combined with FLASHCOPY to ensure a secured copy <strong>of</strong><br />
the whole backup. In the case when data in the main<br />
corrupts, remote disaster recovery backup can be used to<br />
quickly restore data service..<br />
Currently, two basic types <strong>of</strong> disk-based remote copy<br />
are commonly used in industry, namely, synchronous<br />
PPRC and asynchronous PPRC. The major problem with<br />
synchronization PPRC is that it will occupy more<br />
bandwidth when transferring through network which<br />
influences the normal system performance. As a result,<br />
the performance <strong>of</strong> the whole system will be degraded<br />
when disaster recovery is carried out.<br />
Though asynchronous PPRC data may cause data lost<br />
problem, and asynchronous PPRC may cause<br />
inconsistency in the data if it fails to complete<br />
synchronization successfully, there is no doubt that<br />
asynchronous PPRC is far more efficient than its<br />
synchronous counterpart. As digital library has massive<br />
data, we have to select asynchronous PPRC to complete<br />
the daily remote backup.<br />
Therefore, we propose to do remote backup through<br />
asynchronous PPRC combined with Flashcopy, as<br />
asynchronous PPRC can resolve the performance<br />
problem while Flashcopy can resolve the data lost<br />
problem. Afterwards, data is synchronized. This is a<br />
quick way to ensure that backup data can be rolled back<br />
after the data loss. In fact, two techniques complement<br />
with each other, the combined application <strong>of</strong> two<br />
techniques result in more efficient, faster and safer<br />
disaster recovery than sole application <strong>of</strong> synchronous<br />
PPRC [6].<br />
VI. BACKUP DATA ARCHIVE FLOW DIAGRAM.<br />
If remote disaster recovery backup data already exists,<br />
the major task for backup and archiving is to ensure data<br />
security and to decrease data storage requirement.<br />
Firstly, for those digital resources which are under<br />
construction or which are used in online services, part <strong>of</strong><br />
the data do not require to be stored in the storage array in<br />
the form <strong>of</strong> long-term storage, these data can be archived<br />
directly; Furthermore, if some data do not have<br />
downloading or visiting for long time, these should also<br />
be archived according to the information life-cycle<br />
management theory. Therefore, the first step in Fig. 2<br />
is to evaluate data to determine whether it should be<br />
backed up or to be archived, and to decide whether to use<br />
disk media or tape in backup. In the case <strong>of</strong> archiving,<br />
using tape is relatively affordable.<br />
Secondly, duplication elimination techniques can be<br />
used to shrink size <strong>of</strong> the data that will be backed or<br />
© 2011 ACADEMY PUBLISHER<br />
archived; as a result, this also improves the efficiency <strong>of</strong><br />
backup or archive operation. For the same data type, it is<br />
obvious that coping small amounts <strong>of</strong> data is much faster.<br />
Here we use duplication elimination techniques to<br />
process the source data. However, there are some<br />
exceptions, namely, duplication elimination techniques<br />
which are based on either hashing or content<br />
identification are effective only when content <strong>of</strong> data<br />
blocks are duplicated. For example, it is quite effective<br />
to apply duplication elimination techniques to the virtual<br />
machine files <strong>of</strong> service systems to greatly reduce the<br />
amount <strong>of</strong> backup or archiving data. Nevertheless, if the<br />
document is already in a compressed format such as<br />
DJVU or compressed video files, the situation becomes<br />
less optimistic. As duplication elimination takes a lot <strong>of</strong><br />
time in comparison and calculation, it is best to do<br />
backup or archiving directly if the benefit obtained from<br />
data compression is not obvious.<br />
Thirdly, it is necessary to determine whether the data<br />
has been backed up or archived. If data has been backed<br />
up, only incremental or differential backup should be<br />
performed; otherwise, full backup has to be done. The<br />
decision to choose differential backup or incremental<br />
backup replies on the properties for data recovery which<br />
also use the same way to restore data. Differential backup<br />
is a backup all files has changed since the last full backup.<br />
Advantages <strong>of</strong> this method are that it performs well when<br />
a full recovery is demanded, as it only involves restoring<br />
a full backup and the latest differential backup.<br />
Disadvantage is that the size <strong>of</strong> differential backups<br />
grows quickly within a week. Hence the backup data can<br />
grow to a considerable scale before the next full backup.<br />
Incremental backup only backs up the data changed in<br />
files since the last backup, regardless whether the last<br />
backup is full or incremental. The main advantages <strong>of</strong><br />
this approach are that files backed up each day between<br />
two full backups significantly reduce the backup window<br />
and are more concise. The disadvantage is that in order to<br />
perform a full recovery, the latest full backup has to be<br />
restored, together with all subsequent incremental<br />
backups, thus incremental backup is more time<br />
consuming. Technical explanation for the backups can be<br />
summarized as follows: full backup and incremental<br />
backup can be used to reset the archive bit in a file to<br />
indicate the file has been backed up, while differential<br />
backup cannot perform the archive bit resetting [7]. In<br />
addition, it is also necessary at this stage to choose a<br />
backup or archive data storage media in the hope to be<br />
more cost-effective.<br />
Finally, it is important to ensure the readability <strong>of</strong><br />
backup or archive data, therefore, it is very important to<br />
perform data recovery verification. Based on the amount<br />
<strong>of</strong> backup or archiving data, as well as the differences in<br />
data backup storage hardware and archive material, it is<br />
proposed that the backup cycle should be a week or a<br />
month, while the archiving cycle should be three or six<br />
months. Of course, the six month archive cycle also takes<br />
into consideration that lower visiting amount for libraries<br />
happen at the winter and summer semester breaks for<br />
Chinese universities.
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1987<br />
VII. SUMMARY<br />
With the rapid development <strong>of</strong> digital library, data<br />
volume in digital resources grow rapidly, and hence data<br />
storage and backup become more and more difficult. For<br />
the self-built digital resources which focus on the<br />
unstructured data, the proposed backup system can<br />
protect data well, and reduce backup storage consumption.<br />
However, due to the imperfection <strong>of</strong> long-term<br />
preservation techniques, it is still insufficient to achieve<br />
long-term preservation <strong>of</strong> data in the system.<br />
Storage media. Currently, a variety <strong>of</strong> digital<br />
resources are based on a binary 0 or 1 stored in some<br />
physical carrier, the digital information life depends on<br />
the physical carrier <strong>of</strong> life. The life <strong>of</strong> the disk is<br />
generally believed to be an average <strong>of</strong> 10 to 15 years,<br />
CD-ROM about 30 years, durable CD-ROM up to 100<br />
years, but the durable CD-ROM is expensive, hence<br />
cannot be widely used. Even the most durable discs<br />
cannot be used to save printed literature for thousands<br />
years, it is also that much difference, the long-term<br />
preservation <strong>of</strong> digital information need strengthening. In<br />
addition, because the digital document carrier prone to<br />
physical or chemical change, so demanding on the<br />
storage environment. The results <strong>of</strong> the disappearance <strong>of</strong><br />
the event information will be disastrous about digital<br />
document. Therefore, solving the long-term preservation<br />
<strong>of</strong> digital resources stored vector problem is an important<br />
challenge <strong>of</strong> digital resource conservation; it needs to be<br />
studied carefully.<br />
Diverse formats <strong>of</strong> information resources. Library<br />
digital resources, including digital resources, whether self<br />
or mirroring database which be provided by database<br />
vendors, their format is diverse: TXT, PDF, CAJ, PDG,<br />
JPG, TIF, DJVU, MP3, MPG, RMVB and so numerous,<br />
as the storage technology <strong>of</strong> digital resources continuous<br />
development, many <strong>of</strong> the digital storage format makes a<br />
variety difficulty <strong>of</strong> network data exchange between<br />
information resources, It affects the long-term use <strong>of</strong><br />
digital resources.<br />
Technology obsolete. With the network information<br />
technology and its products constantly upgrading, the one<br />
hand, enhanced information processing capability makes<br />
the network cost reduction, on the other hand, the using<br />
<strong>of</strong> stored digital resources has new difficulties. As digital<br />
resources are digital electronic information resources, it<br />
needs computer equipment with certain s<strong>of</strong>tware. With<br />
these s<strong>of</strong>tware and hardware technology continues to<br />
© 2011 ACADEMY PUBLISHER<br />
upgrade, making the old and new versions <strong>of</strong> the s<strong>of</strong>tware<br />
is not compatible, use older versions <strong>of</strong> technology to<br />
store digital resources can not read, it makes the loss <strong>of</strong><br />
resources, loss <strong>of</strong> use value. It affects the long-term<br />
preservation <strong>of</strong> digital resources.<br />
Network information security. In recent years, with<br />
the rapid popularization and development <strong>of</strong> the Internet,<br />
a variety <strong>of</strong> network information in the net, people can<br />
easily and freely on the Internet to read, browse, search,<br />
download a variety <strong>of</strong> network information, access to<br />
information to people has brought great convenience .<br />
However, a large number <strong>of</strong> computer viruses on the<br />
Internet, seriously affecting the information resources<br />
security on the network transmission and storage.<br />
Meanwhile, hackers also took the opportunity to<br />
infiltration, they use the computer system itself, there are<br />
a lot <strong>of</strong> flaws and weaknesses to attack, the light can not<br />
use the computer, the serious is causing severe paralysis<br />
<strong>of</strong> the network, so that preservation <strong>of</strong> digital resources is<br />
increasingly serious security problem, it cause permanent<br />
loss <strong>of</strong> digital information, it became one <strong>of</strong> the biggest<br />
threats on preservation <strong>of</strong> digital resources.<br />
Further research will be carried out on long-term<br />
preservation technology to establish a more<br />
comprehensive, more reliable and efficient backup<br />
system.<br />
REFERENCES<br />
[1] CHEN You hua, ZHENG Qiao ying, YANG Zong ying,<br />
WANG Shao ping, SUN Hua. Self-Developed Digital<br />
Resources <strong>of</strong> Chinese Academic Libraries [J]. <strong>Journal</strong> <strong>of</strong><br />
Shanghai Jiaotong University, 2003, (S1).<br />
[2] http://baike.baidu.com/view/2119114.htm<br />
[3] Li kezheng. Long-term preservation <strong>of</strong> digital information<br />
technology analysis [J]. Library Work and Study, 2006, (2).<br />
[4] You L L,Pollack K T,Long D D E.Deep Store:An Archival<br />
Storage System Architecture[C]//Proc.<strong>of</strong> the 21st<br />
International Conference on Data<br />
Engineering.California,USA:[s.n.],2005:804-815.<br />
[5] Bhagwat D,Pollack K T,Long D D E.Providing High<br />
Reliability in a Minimum Redundancy Archival Storage<br />
System[C]//Proc.<strong>of</strong> the 14th IEEE International Symposium<br />
on Volume.New York,USA:[s.n.],2006:413-421.<br />
[6] Zhou Jian Feng.RESEARCH OF FAST DISASTER<br />
RECOVERY BETWEEN DIFFERENT SITES[D].<br />
Shanghai Jiao Tong University,2009.<br />
[7] http://www.searchstorage.com.cn/ShowContent_14945_26.<br />
htm
1988 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
Configuration Scheme for Small Scale<br />
Multi-FPGA Systems<br />
Chengchang Zhang 1<br />
1 College <strong>of</strong> Communication Engineering<br />
Chongqing University, Chongqing, P.R. China<br />
Email:zcc_918@163.com<br />
Lisheng Yang 2 , Dangui Yan 3 , Changyong Li 4<br />
2 College <strong>of</strong> Communication Engineering, Chongqing University, Chongqing, P.R. China<br />
Email: yls@ccee.cqu.edu.cn<br />
3 College <strong>of</strong> Mathematics and Physics, Chongqing University <strong>of</strong> Post and Telecom, Chongqing, P.R. China<br />
Email: yandg@cqupt.edu.cn<br />
4 Chongqing Communication Acadimic <strong>of</strong> P.L.A. ,Chongqing, P.R. China<br />
Email: lll_ccc_yyy@163.com<br />
Abstract—Multi-FPGA systems have tremendous potential,<br />
providing a high-performance computing substrate for<br />
many different applications. These systems harness multiple<br />
FPGAs, connected in a fixed pattern, to implement complex<br />
logic structures. In order to use such a system effectively,<br />
it is a key for constructing a good performance hardware<br />
platform. The configuration scheme is an important part in<br />
hardware design. This paper aims at small scale<br />
Multi-FPGA systems composed <strong>of</strong> SRAM-based FPGAs<br />
developed by Xilinx Corporation, proposes a novel<br />
configuration technique by using Platform Flash PROM<br />
XCF32P. Using this scheme, only adopting one XCF32P and<br />
one Complex Programmable Logic Device (CPLD) we can<br />
configure four FPGAs with monolithic configuration data<br />
smaller than 8Mbit. When the number <strong>of</strong> FPGA is more<br />
than four, Design revisioning allows the user to cascade<br />
more XCF32P PROMs to realize. Since Xilinx Platform<br />
PROM and Xilinx FPGA/CPLD are used to get a<br />
single-vender solution, the design for hardware and<br />
s<strong>of</strong>tware is simplified.<br />
Index Terms—Multi-FPGA systems, XCF32P, design<br />
revision, configuration,<br />
I. INTRODUCTION<br />
There is currently tremendous interest in the<br />
development <strong>of</strong> computing platforms from multiple<br />
standard FPGAs [1,2,3,4]. One reason is that the digital<br />
system is too large to be achieved with only one FPGA,<br />
another, the growth rate <strong>of</strong> the FPGA capacity is far<br />
behind that <strong>of</strong> the ASIC(Application Specific Integrated<br />
Circuit) chip scale [5,6]. These systems harness multiple<br />
FPGAs [7], connected in a fixed pattern, to implement<br />
complex logic structures. In order to use such a system<br />
effectively, it is a key for constructing a good<br />
performance hardware platform. The configuration<br />
method plays important role for hardware platform<br />
because <strong>of</strong> two major factors. First, the configuration<br />
chips affect layout and wiring for printed circuit<br />
board(PCB). Second, the initialization and<br />
© 2011 ACADEMY PUBLISHER<br />
doi:10.4304/jcp.6.9.1988-1993<br />
reconfiguration for a multi-FPGA system is usually<br />
needed after the PCB developed, especially in system<br />
debug. A good design <strong>of</strong> configuration can optimize<br />
construction <strong>of</strong> PCB, and also make the configuration and<br />
debug processes more convenient and effective.<br />
In this paper, we focus on SRAM-based FPGAs<br />
developed by Xilinx Corporation. In SRAM-based<br />
FPGAs, the contents <strong>of</strong> the internal configuration<br />
memory are reset after power-up. As a result, the internal<br />
configuration memory cannot be used for storing<br />
configuration data permanently. SRAM-based FPGAs<br />
require external devices to initiate and control the<br />
configuration process.<br />
For Multi-FPGA systems configuration, if the number<br />
<strong>of</strong> FPGA chip and monolithic FPGA configuration files<br />
are both very large in a system, such as the DN9000K10<br />
System [8] developed by Dini Company, the Xilinx<br />
Company launched a special configuration solution, that<br />
is: System ACE (System Advanced Configuration<br />
Environment), in this solution, CF(Compact Flash) Card<br />
and ACE Control Chip are used to configure the multiple<br />
FPGAs automatically [9,10], but the system is costly. For<br />
general application system (such as the number <strong>of</strong> FPGA<br />
isn’t larger than four, and the configuration files is less<br />
than 8Mbit), self-made configuration scheme is usually<br />
adopted, for example, literatures [11,12,13,14] use the<br />
configuration scheme based on CPLD and general<br />
FLASH, a special FLASH drive device is needed to<br />
program configuration file to FLASH, and a group <strong>of</strong><br />
output pins corresponding with FLASH capacity are<br />
needed to be distributed as address bus. And, designers<br />
must be clear with the first and the end address in the<br />
FLASH corresponding with configuration files <strong>of</strong> each<br />
FPGA, so that they can make sure that the counter in<br />
CPLD can start the control signal <strong>of</strong> next FPGA<br />
configuration after completing the last configuration,<br />
which is in fact very troublesome. Besides, the access<br />
speed <strong>of</strong> general FLASH is relatively slow to the FPGA<br />
and affects the system configuration speed. Literature
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1989<br />
[15] adopted the DSP + CPLD + general FLASH<br />
configuration scheme, which is based on processor, the<br />
design and debug <strong>of</strong> the circuit and program cost<br />
considerable time, and processor usually bears arduous<br />
task in addition to completing the FPGA configuration,<br />
so bus contention is appear easily.<br />
In this paper, we propose a novel configuration scheme<br />
based on Xilinx Platform Flash PROM XCF32P<br />
The chip supports FPGA serial or parallel interface<br />
configuration, basically have the following typical<br />
characteristics [16,17,18]:<br />
The embedded data decompressor compatible with<br />
Xilinx senior compression technique can decompress<br />
PROM compressed files with a highest 50% data<br />
compression ratio, and the compressed file is generated<br />
from target FPGA bit stream file. When decompression is<br />
enabled, FPGA must be in slave configuration mode and<br />
PROM first decompress the stored data then drive the<br />
clock and data to FPGA interface.<br />
There is an optional oscillator in interior and can<br />
provide a 20MHz or 40MHz clock which is output by<br />
CLKOUT pin. Among them, the 40MHz clock is used to<br />
start the internal decompressor.<br />
Design revisioning allows the user to create up to four<br />
unique design revisions on a single PROM or stored<br />
across multiple cascaded PROMs. Design revisioning can<br />
be used with compressed PROM files, and also when the<br />
CLKOUT feature is enabled. The 32Mbit storage<br />
capacity <strong>of</strong> monolithic XCF32P can be divided into<br />
several independent spaces, with 8Mbit as a unit, and<br />
each independent space can store an independent<br />
configuration file, which is called a storage version.<br />
There are many methods to manage storage versions.<br />
Shown as Fig. 2, one XCF32P can be divided into only<br />
one 32Mbit storage version, two independent 16Mbit<br />
storage versions, one independent 8Mbit storage version<br />
and one independent 24Mbit storage version, two<br />
independent 8Mbit storage versions and one independent<br />
16Mbit storage version or four independent 8Mbit<br />
storage versions, and so on. During the PROM file<br />
creation, each design revision is assigned a revision<br />
number: Revision 0 = '00', Revision 1 = '01', Revision 2 =<br />
'10', Revision 3 = '11'.<br />
© 2011 ACADEMY PUBLISHER<br />
Fig.1 Structure diagram <strong>of</strong> XCF32P<br />
to simplify the design <strong>of</strong> hardware and s<strong>of</strong>tware.<br />
II. XCF32P STRUCTURE CHARACTERISTICS<br />
XCF32P is the programmable high capacity Platform<br />
Flash PROM developed by Xilinx Company, its storage<br />
capacity is 32Mbit. The structure diagram is shown as<br />
Fig.1.<br />
Fig.2 Design Revision storage examples for a single XCF32P PROM<br />
After programming the Platform Flash PROM with a<br />
set <strong>of</strong> design revisions, a particular design revision can be<br />
selected using the external REV_SEL[1:0] pins or using<br />
the internal programmable design revision control bits.<br />
The EN_EXT_SEL pin determines if the external pins or<br />
internal bits are used to select the design revision. When<br />
EN_EXT_SEL is Low, design revision selection is<br />
controlled by the external revision select pins,<br />
REV_SEL[1:0]. When EN_EXT_SEL is High, design<br />
revision selection is controlled by the internal<br />
programmable revision telect control bits. During power<br />
up, the design revision selection inputs(pins or control<br />
bits) are sampled internally. After power up, when CE is<br />
asserted (Low) enabling the PROM inputs, the design<br />
revision selection inputs are sampled again after the<br />
rising edge <strong>of</strong> the CF pulse. The data from the selected<br />
design revision is then presented on the FPGA<br />
configuration interface.<br />
Xilinx company develops the Multiple versions design<br />
function <strong>of</strong> Platform Flash PROM is to realize the<br />
dynamic reconfigure <strong>of</strong> system or for some special<br />
application <strong>of</strong> changeable configuration when start the
1990 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
FPGA each time. The work in this paper uses the<br />
multiple independent design versions to achieve multiple<br />
FPGAs configuration.<br />
III. CONFIGURING FOUR VIRTEX XCV200 FPGAS<br />
A. System components.<br />
The system includes one Platform Flash PROM<br />
XCF32P, one CPLD XC9572 and four XCV200 FPGAs<br />
to be configured, the system structure diagram is shown<br />
in Fig. 3a, and the circuit board is shown in Fig. 3b.<br />
Fig.3a. Structure diagram <strong>of</strong> configuration system<br />
B. Configuration principle.<br />
Fig.3b. The circuit board<br />
The configuration interface circuit is shown in Fig. 4.<br />
The circuit is designed with the help <strong>of</strong> OrCAD<br />
s<strong>of</strong>tware. Because the s<strong>of</strong>tware can't identify the sign <strong>of</strong><br />
NOT operation, low-level effective is expressed as "/"<br />
(same in the following text).<br />
Virtex XCV200 FPGA supports the following four<br />
configuration modes[19]: master serial mode, slave serial<br />
mode, slave parallel (Slave SelectMAP)mode and<br />
boundary scan mode. In this work, high-speed<br />
slave-parallel mode is used and configuration clock<br />
CCLK is supplied by exterior. The frequency is<br />
determined by the formula followed:<br />
In equation (1), t is the access time <strong>of</strong> XCF32P<br />
ACC<br />
with a minimum <strong>of</strong> 25ns, t is the setup time <strong>of</strong><br />
SMDCC<br />
input data <strong>of</strong> the SelectMAP interface with a minimum <strong>of</strong><br />
2ns.Thus, the maximum frequency <strong>of</strong> CCLK is about<br />
37MHz.<br />
© 2011 ACADEMY PUBLISHER<br />
f<br />
CCLK<br />
1<br />
=<br />
t + t<br />
ACC SMDCC<br />
Fig.4 Configuration circuit interface<br />
Make a Parallel connection for the control signal<br />
CCLK, /PROGRAM, /INIT and data D[7..0] <strong>of</strong> the four<br />
FPGAs. And configure all devices orderly by setting chip<br />
selected signal /CS [4..1] respectively. When one <strong>of</strong> the<br />
three FPGAs(FPGA1, FPGA2 and FPGA3) configuration<br />
completed, it will enter its start-up stage, and send out its<br />
instructions signal DONE, set the version selection signal<br />
corresponding to the next configuration program and start<br />
configuration for next FPGA. It means that configuration<br />
is completed when the forth FPGA(FPGA4) release its<br />
signal DONE. This signal is connected to /CE, XCF32P<br />
is no longer effective and configuration process ends. The<br />
configuration flow is shown in Fig. 5.<br />
The data configuration timing diagram is shown in Fig.<br />
6. When /PROGRAM is in low state, four FPGAs begin<br />
to initialize synchronously. After initialization completed,<br />
the signal DONE turns to be low. Because the signal /CE<br />
<strong>of</strong> XCF32P is connected with the signal DONE <strong>of</strong> the<br />
forth FPGA (DONE4), the chip enable signal <strong>of</strong> XCF32P<br />
is effective. Meanwhile, the signal /INIT turn to be low<br />
automatically and begin to clear configuration memory.<br />
When the low level <strong>of</strong> the signal /INIT is input to the<br />
OE/(/RESET) interface <strong>of</strong> XCF32P, the chip XCF32P<br />
begins to reset and address pointer points to the first<br />
address <strong>of</strong> memory space. After configuration memory is<br />
emptied, the signal /INIT is set to high again, and device<br />
samples mode pins to make sure that configuration data is<br />
loaded in parallel mode.<br />
When multi-version design function is started, the<br />
internal logic <strong>of</strong> configuration PROM samples the design<br />
version selected input(pin /SEL) when power up. When<br />
/CE is set to low, the design version selected input signal<br />
(1)
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1991<br />
is sampled again at the rising edge <strong>of</strong> /CF pulse to<br />
determine which design version to provide configuration<br />
monolithic<br />
configuration<br />
process<br />
Therefore, the signal /INIT is regarded as the initial<br />
trigger signal <strong>of</strong> /CF, and /CF is triggered at the rising<br />
edge <strong>of</strong> /INIT. /CF is set to be low and the low level<br />
should delay more than 300ns duration. The version<br />
selected input signal /SEL is triggered and set to be “00”<br />
at the same time, namely, configuration data is output<br />
form the zero version. Trigger piece selected signal /CS1<br />
is effective at the rising edge <strong>of</strong> /CF signal. The zero<br />
version data <strong>of</strong> XCF32P is output to the first FPGA and<br />
begin to configure FPGA1 at the affection <strong>of</strong> CCLK.<br />
When the first FPGA is configured, it releases the signal<br />
DONE, by this way, DONE1 turns to be high level. /CF<br />
signal is triggered by the rising edge <strong>of</strong> DONE1 and is<br />
reset to be low level, at the same time, /SEL signal is set<br />
© 2011 ACADEMY PUBLISHER<br />
data for the FPGA. The version selected pin should be set<br />
before sampling is triggered at least 300ns.<br />
Start Clear the configuration<br />
memory and set DONE<br />
to be low<br />
/PROGRAM is low?<br />
Clear the configuration<br />
memory again<br />
/INIT is low?<br />
Begin to configure FPGA1<br />
Set the chip selected /CS to be low; Set the version code<br />
REV_SEL corresponding to this chip; Set the version<br />
initialization signal /CF to be low, and the low level stay<br />
for longer than 300ns.<br />
Write data in<br />
BUSY is low?<br />
DONE is high?<br />
Set the chip selected /CS<br />
to be high, and enter the<br />
starting process<br />
Repeat monolithic configuration<br />
process, and configure FPGA2,<br />
FPGA3 and FPGA4<br />
DONE4 is high?<br />
Set /CE to be high<br />
End<br />
N<br />
Y<br />
Y<br />
Y<br />
Y<br />
Fig.5 Configuration flow<br />
N<br />
N<br />
N<br />
N<br />
to be “01”. When the rising edge <strong>of</strong> /CF signal arrives,<br />
configuration data is sent out by the first version <strong>of</strong><br />
XCF32P. By this time, /CS2 is set to be effective and the<br />
second FPGA is selected to begin receive configuration<br />
data. Besides, /CS1 is set to be ineffective and starts to<br />
configuration the second FPGA. The configuration <strong>of</strong> the<br />
third and forth FPGA is similar to above. After the forth<br />
FPGA configured, /CE <strong>of</strong> XCF32P is set to be high level<br />
by DONE4 signal released by this FPGA. That is to say,<br />
the chip enable signal <strong>of</strong> XCF32P is ineffective and the<br />
whole configuration process ends.<br />
C.The s<strong>of</strong>tware design <strong>of</strong> CPLD<br />
The design <strong>of</strong> internal control circuit in CPLD is a key
1992 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
<strong>of</strong> the system. Providing the needed timing sequence<br />
when configuring, coordinating the configuration process,<br />
and ensuring that multi-FPGA configuration completed<br />
Design is realized by combining the hardware<br />
description language with schematic diagram. Control<br />
circuits are made up <strong>of</strong> a delay model, a counter and a<br />
shift register, as shown in Fig. 7.<br />
Delay module tests the rising edge <strong>of</strong> /INIT,<br />
DONE(1), DONE(2) and DONE(3) and trigger internal<br />
delay circuit to produce the negative pulses longer than<br />
300ns which is need by /CF signal. It is difficult to detect<br />
rising edge <strong>of</strong> four signals simultaneously, so, there are<br />
four independent delay circuits in delay model to detect<br />
four trigger signals respectively and to produce four<br />
negative pulses which can get /CF signal when they are<br />
done the AND operation. The shift register is triggered<br />
by the rising edge <strong>of</strong> /CF and produces the chip selected<br />
signal /CS(4:1). The falling edge <strong>of</strong> /CF triggers counter<br />
and produce version selected signal /SEL(1:0).<br />
The simulation results <strong>of</strong> control circuit are shown in<br />
Fig. 8.<br />
IV. CONCLUSIONS<br />
A new configuration scheme for small scale<br />
multi-FPGA systems based on XCF32P is given. In this<br />
scheme, a XCF32P and a CPLD are used to configure<br />
© 2011 ACADEMY PUBLISHER<br />
Fig.6 Configuration timing<br />
Fig.8 Control circuit simulation results<br />
as the predetermined process are the main functions <strong>of</strong><br />
this work.<br />
Fig.7 Control circuit block<br />
four Virtex XCV200 FPGAs. The design has certain<br />
universality, and can be used to configure multiple Xilinx<br />
FPGAs with monolithic configuration data smaller than<br />
8Mbit. When starting the internal decompression in<br />
XCF32P, monolithic FPGA configuration data can reach<br />
16Mbit. When the number <strong>of</strong> FPGA is more than four,
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1993<br />
Design revisioning allows the user to cascade more<br />
XCF32P PROMs to realize.<br />
Due to the XCF32P is special configuration chip<br />
developed by XILINX Company, the chip access time is<br />
short, and the configuration speed is fast. Meanwhile,<br />
Xilinx Platform and Xilinx FPGA/CPLD are used to get a<br />
single-vender solution to make the design for hardware<br />
and s<strong>of</strong>tware simplified.<br />
ACKNOWLEDGMENT<br />
The authors thank their colleagues at College <strong>of</strong><br />
Communication and Engineering for fruitful discussions.<br />
This work was supported by Natural Science Foundation<br />
Project <strong>of</strong> CQ CSTC under contract no: 2010BB2240.<br />
REFERENCES<br />
[1] Panella, A.; Santambrogio, M.D.; Redaelli, F.; Cancare,<br />
F.; Sciuto, D.. A design workflow for dynamically<br />
reconfigurable multi-FPGA systems. VLSI System on<br />
Chip Conference (VLSI-SoC), 2010 18th IEEE/IFIP, pp.<br />
414-419.<br />
[2] Jain, S.C.; Kumar, S.; Kumar, A.. Evaluation <strong>of</strong> various<br />
routing architectures for multi-FPGA boards. VLSI<br />
Design, 2000. Thirteenth International Conference on ,<br />
pp.262-267.<br />
[3] Khalid, M.A.S.; Rose, J.. A novel and efficient routing<br />
architecture for multi-FPGA systems. Very Large Scale<br />
Integration (VLSI) Systems, IEEE Transactions on, Vol. 8<br />
, Issue. 1, 2000 , pp.30-39.<br />
[4] Zhang Cheng-chang; Yan Dan-gui; Yang Li-sheng; Qi<br />
Huai-long; Li Chang-yong. DLL-based multi-FPGA<br />
systems clock synchronization. Industrial Electronics and<br />
Applications (ICIEA), 2010 the 5th IEEE Conference on,<br />
pp. 1420-1423.<br />
[5] Krupnova, H. Mapping multi-million gate SoCs on<br />
FPGAs:industrial methodology and experience. Design,<br />
Automation and Test in Europe Conference and<br />
Exhibition, Proceedings, Volume 2, 16-20 Feb.2004 Vol.2,<br />
pp.1236-1241.<br />
[6] Melnikova, O.; Hahanova, I.; Mostovaya, K.. Using<br />
multi-FPGA systems for ASIC prototyping. CAD<br />
Systems in Microelectronics, 2009. CADSM 2009. 10th<br />
International Conference-The Experience <strong>of</strong> Designing<br />
and Application <strong>of</strong> 24-28 Feb.2009, pp.237-239.<br />
[7] Scott Hauck. Multi-FPGA Systems, Doctor <strong>of</strong> Philosophy,<br />
University <strong>of</strong> Washington. 1995.<br />
[8] http://www.dinigroup.com/index.php/.<br />
[9] Yang Sen; Chen Jian-jun; Wang Jian-guo. System ACE<br />
CF Technology--A New Configuration Solution for<br />
FPGAs. Radar & Ecm, 2002(4), pp.72-77.<br />
[10] Alonso, R.; Barbara, D.; Cova, L.L.. A file storage<br />
implementation for very large distributed systems.<br />
© 2011 ACADEMY PUBLISHER<br />
Workstation Operating Systems, 1989., Proceedings <strong>of</strong><br />
the Second Workshop on , pp. 1-5.<br />
[11] Li Peng; Lan Ju-long. The configuration method for<br />
FPGA based on CPLD and Flash. Application <strong>of</strong><br />
Electronic Technique, 2006(6), pp.101-103.<br />
[12] Guo Tian-tian. Interface Circuit <strong>of</strong> Configuring Virtex<br />
FPGA Through SelectMAP. Microprocessors, 2000(4),<br />
pp.17-19.<br />
[13] Zhang Hong-gang; Xin Fan-ge; Wang De-shi. The fast<br />
configuration circuit design for FPGA based on CPLD.<br />
Application <strong>of</strong> Electronic Technique, 2006(2),<br />
pp.123-125.<br />
[14] Xiao Jin-qiu; Liu Chuan-yang; Feng Yi; Zhong Jia-lin.<br />
Design <strong>of</strong> FPGA Initialization Configure System at High<br />
Speed with LPC Bus. Computer Engineering,<br />
2005(13):176-178.<br />
[15] She You-jun; Wang Dan. Design for double FPGA<br />
configuration based on TMS320C61416 EMIF bus<br />
[J].Microcontrollers & Embedded Systems,<br />
2007(7):29-31.<br />
[16] Platform Flash In-System Programmable Configuration<br />
PROMS. http://www.xilinx.com, DS123 (v2.6) March<br />
14, 2005.<br />
[17] LI Yan-bin; LI Yan-chun. Fast Dynamic Reconfiguration<br />
<strong>of</strong> FPGA with XCF32P. Telecommunication Engineering,<br />
2006(6):199-202.<br />
[18] Platform Flash PROM User Guide. http://www.xilinx.com,<br />
UG161(v1.5) October 26, 2009.<br />
[19] Virtex 2.5V Field Programmable Gate Arrays.<br />
http://www.xilinx.com, DS003-1(v2.5) April 2, 2001.<br />
Chengchang Zhang was born in Lichuan, China, 1975. He<br />
received the BS degree in automation engineering from the<br />
Wuhan Institute <strong>of</strong> Chemical Technology in 1997. He received<br />
the MS degree in College <strong>of</strong> Communications Engineering from<br />
Chongqing University in 2005. He is currently PhD candidate<br />
<strong>of</strong> Chongqing University majoring in Communication and<br />
information systems. His research interests are s<strong>of</strong>tware radio<br />
and FPGA design.<br />
Lisheng Yang was born in Chongqing, China, 1972. He is an<br />
pr<strong>of</strong>essor <strong>of</strong> Chongqing University. His research interest<br />
includes s<strong>of</strong>tware radio, radar ,TT&C, etc.<br />
Dangui Yan was born in Luotian, China, 1975. She is an<br />
lecturer <strong>of</strong> Chongqing University <strong>of</strong> Post and Telecom. Her<br />
research interest is logic algebra.<br />
Changyong Li was born in Chongqing, China, 1971. His<br />
research interests are s<strong>of</strong>tware radio and ultra-wide band radar.
1994 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
Order Bi-spectrum For Bearing Fault Monitoring<br />
and Diagnosis Under Run-up Condition<br />
Hui Li<br />
Department <strong>of</strong> Electromechanical Engineering, Shijiazhuang Institute <strong>of</strong> Railway Technology, Shijiazhuang, China<br />
Email: Huili68@163.com<br />
Abstract—Varying speed machinery condition detection and<br />
fault diagnosis are more difficult due to non-stationary<br />
machine dynamics and vibration. Therefore, most<br />
conventional signal processing methods based on time<br />
invariant carried out in constant time interval are<br />
frequently unable to provide meaningful results. This paper<br />
deals with the detection <strong>of</strong> bearing faults in gearbox under<br />
non-stationary run-up <strong>of</strong> gear drives. In order to process the<br />
non-stationary vibration signals such as run-up or rundown<br />
vibration signals effectively, the order bi-spectrum<br />
technique is presented. This new method combines<br />
computed order tracking technique with bi-spectrum<br />
analysis. First, the vibration signal is sampled at constant<br />
time increments during run-up <strong>of</strong> gearbox and then uses<br />
numerical techniques to resample the data at constant angle<br />
increments. Therefore, the vibration signals are<br />
transformed from the time domain transient signal to angle<br />
domain stationary one. Second, the re-sample signal is<br />
processed by bi-spectrum analysis method. The procedure is<br />
illustrated with the experimental vibration data <strong>of</strong> a gearbox.<br />
The experimental results show that order bi-spectrum<br />
technique can effectively diagnosis and diagnosis the faults<br />
<strong>of</strong> bearing.<br />
Index Terms—fault diagnosis, gearbox, bearing, vibration,<br />
signal processing, order tracking, bi-spectrum<br />
I. INTRODUCTION<br />
Rotating machine fault diagnosis is typically based on<br />
vibration. The spectral contents <strong>of</strong> emitted vibration<br />
signals are analyzed to ascertain the current condition <strong>of</strong><br />
the monitored process. At present, for the fault diagnosis<br />
<strong>of</strong> rotating machinery, many research outcomes have<br />
been obtained in the stationary process. However, little<br />
research has been done for monitoring the vibrations <strong>of</strong><br />
varying speed condition such as the run-up or run-down<br />
process. The reason why we stress the run-up or rundown<br />
process is that non-stationary vibrations signals<br />
from varying speed machinery may include more<br />
abundant information about its condition. Some<br />
phenomena, which are usually not obvious at constant<br />
speed operation, may become more apparent under<br />
varying speed conditions. Therefore, the behavior<br />
characteristics <strong>of</strong> the run-up or run-down process have a<br />
distinct diagnostic value, and the fault diagnosis <strong>of</strong> run-up<br />
Manuscript received November 5, 2010; revised December 23,<br />
2010; accepted January 28, 2011.<br />
Corresponding author: Hui Li.<br />
© 2011 ACADEMY PUBLISHER<br />
doi:10.4304/jcp.6.9.1994-2000<br />
or run-down process has owed its distinct standing in the<br />
fault diagnosis <strong>of</strong> rotating machinery. In the last decade<br />
vibration analysis and condition monitoring techniques<br />
for varying speed machinery have attracted the attention<br />
<strong>of</strong> scientists and engineers. Lopatinskaia et al. [1,2]<br />
presented the application <strong>of</strong> recursive filtering and angle<br />
domain analysis to non-stationary vibration analysis. The<br />
approach is implemented and validated through computer<br />
simulation and experiments. Meltzer [3,4] dealt with the<br />
recognition <strong>of</strong> faults in gear tooth during non-stationary<br />
start-up and run-down <strong>of</strong> planetary gear drives using the<br />
time-frequency approach and the time-quefrency<br />
approach. Wu et al. [5] presented the application <strong>of</strong><br />
adaptive order tracking fault diagnosis technique based<br />
on recursive Kalman filtering algorithm to gear-set defect<br />
diagnosis and engine turbocharger wheel blades damaged<br />
under various conditions. Li et al. [6] presented the<br />
hidden Markov model-based fault diagnosis method in<br />
speed-up and speed-down process for rotating machinery.<br />
However, the vibration signal <strong>of</strong> the run-up or rundown<br />
process is more complex than that <strong>of</strong> the stationary<br />
process. Conventional signal processing methods, which<br />
were developed for constant speed machinery monitoring,<br />
are based on digital sampling carried out in equal time<br />
intervals. If the machine operates under varying speed or<br />
load, its dynamic and vibrations become non-stationary.<br />
The vibration signal sampled from the rotating machinery<br />
is a non-stationary signal, whose amplitudes and<br />
frequencies both vary with time. Fixed time sampling<br />
cannot cope with the varying rotational frequency <strong>of</strong> the<br />
machine, resulting in increasing leakage error and<br />
spectral smearing [1,2]. Therefore, most <strong>of</strong> the<br />
conventional methods for signal processing become<br />
inappropriate when monitoring the vibrations <strong>of</strong> varying<br />
speed machinery [1,2]. Some progress has been made in<br />
the theoretical analysis [7,8], the signal processing<br />
methodology [9,10], measurements and practical<br />
applications <strong>of</strong> varying speed machinery monitoring<br />
[11,12,13].<br />
At present, two techniques are mainly used to process<br />
the non-stationary signal: time frequency analysis (such<br />
as the short time Fourier transform (STFT), wavelet<br />
transform (WT) [14], Wigner-Ville distribution (WVD)<br />
[15,16,17] and Hilbert-Huang transform [18,19,20]) and<br />
order tracking technique [11,12,13]. The time frequency<br />
analysis involves three-dimensional functions that allow<br />
for visualizing the frequency and amplitude variations <strong>of</strong><br />
the spectral components [14]. However, when the
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1995<br />
analyzed vibration signal is composed <strong>of</strong> many spectral<br />
components and with large changes <strong>of</strong> the machine speed<br />
during measurement, they become very difficult to<br />
analyze. Recently, order tracking has been become one <strong>of</strong><br />
the important methods for fault diagnosis in rotating<br />
machinery [11,12,13]. Vibration signals produced from<br />
rotating machinery are speed dependent and hence orders<br />
as opposed to absolute frequencies are preferred as the<br />
frequency base. Orders represent the number <strong>of</strong> cycles<br />
per revolution and are thus ideal for representing speeddependent<br />
vibrations. Therefore, order tracking normally<br />
exploits a vibration or a noise signal supplemented with<br />
the information <strong>of</strong> shaft speed for fault diagnosis <strong>of</strong><br />
rotating machinery. The order spectrum gives the<br />
amplitude <strong>of</strong> the signal as a function <strong>of</strong> harmonic order<br />
and shaft speed in rotating machinery [11].<br />
In this work, the computed order tracking approach<br />
and bi-spectrum analysis are introduced and applied<br />
specifically to gearbox fault diagnosis during run-up.<br />
This method is based on the re-sampling technique and<br />
the bi-spectrum estimation <strong>of</strong> the re-sampling signal,<br />
which is a function <strong>of</strong> the angle <strong>of</strong> the input shaft <strong>of</strong> the<br />
gearbox. This re-sampling signal can be obtained by resampling<br />
<strong>of</strong> the vibration signal that has been previous<br />
sampled in the time domain. The order power spectrum<br />
and order bi-spectrum techniques are based on the signal<br />
processing <strong>of</strong> the angle domain signal, where the<br />
resample signal is in accordance with the shaft angle <strong>of</strong><br />
the gearbox. The order power spectrum and order bispectrum<br />
are then evaluated for the vibration signal resampled<br />
constantly in angle at equidistant phases <strong>of</strong> the<br />
input shaft <strong>of</strong> the gearbox. In this case, the results <strong>of</strong> the<br />
order power spectrum or order bi-spectrum are expressed<br />
as results <strong>of</strong> order analysis where the frequency axes are<br />
changed to the axes <strong>of</strong> orders independent <strong>of</strong> the input<br />
shaft speed. The usefulness <strong>of</strong> this approach will be<br />
shown by experimental example in Section VI.<br />
To address the issues discussed above, this paper is<br />
organized as follows. Section I gives a brief introduction<br />
<strong>of</strong> the order tracking analysis technology. Section II<br />
briefly describes the bi-spectrum. Section III presents the<br />
principles and procedure <strong>of</strong> the computed order tracking.<br />
Section IV gives the method and procedure <strong>of</strong> the fault<br />
diagnosis based on computed order tracking and order bispectrum.<br />
Section V looks at the experimental set-up.<br />
Section VI gives the applications <strong>of</strong> the method based on<br />
computed order tracking and order bi-spectrum to faults<br />
diagnosis <strong>of</strong> bearing faults. Finally, the main conclusions<br />
<strong>of</strong> this paper are provided in Section VII.<br />
II. A BRIEF INTRODUCTION OF BI-SPECTRUM<br />
x be a real, discrete, zero-mean stationary<br />
process with third-order cumulant R xx ( τ1,<br />
τ 2 ) defined<br />
as [21]<br />
Let { (n)<br />
}<br />
Rxx τ , τ ) = E[<br />
x(<br />
n)<br />
x(<br />
n + τ ) x(<br />
n + τ )] (1)<br />
( 1 2<br />
1<br />
2<br />
Then the bi-spectrum <strong>of</strong> { (n)<br />
}<br />
expression<br />
© 2011 ACADEMY PUBLISHER<br />
x is given by the<br />
+∞ +∞<br />
− j(<br />
ω1<br />
τ1+<br />
ω2τ<br />
2 )<br />
B xx ( ω1<br />
, ω2<br />
) = ∑∑Rxx<br />
( τ1,<br />
τ 2 ) e<br />
(2)<br />
τ1= −∞ τ2=<br />
−∞<br />
where ω1 ≤ π , ω2 ≤ π , ω1 + ω2<br />
≤ π .<br />
Therefore, in the same way that the power spectrum<br />
decomposes the power <strong>of</strong> a signal, the bi-spectrum<br />
decomposes the third-order cumulant. The bi-spectrum is<br />
a function <strong>of</strong> two frequency variables, ω1 and ω 2 , and<br />
whilst the power spectrum includes the contribution <strong>of</strong><br />
each individual frequency component independently, the<br />
bi-spectrum analyses the frequency interaction between<br />
the frequency components at ω 1 , ω 2 and ω 1 + ω2<br />
[22-<br />
23].<br />
III. THE PRINCIPALS OF COMPUTED ORDER TRACKING<br />
There are two popular techniques for producing<br />
synchronously sampled data: the traditional approach that<br />
uses special hardware to dynamically adapt the sample<br />
rate and a technique where the vibration signals and a<br />
tachometer signal are synchronously sampled, that is,<br />
they are sampled conventionally at equal time increments.<br />
From the synchronously sampled tachometer signal resample<br />
times required to produce synchronous sampled<br />
data are calculated. This process is referred to as<br />
computed order tracking and is particularly attractive, as<br />
it requires no special hardware. Also, this approach is<br />
more flexible than the traditional method, as for example<br />
different sample rates may be synthesized. The computed<br />
order tracking is considerably more flexible than the<br />
traditional approach. It may be organized to produce<br />
equally accurate or more accurate results than the<br />
traditional method. An added benefit is that computed<br />
order tracking requires no specialized hardware, which is<br />
an important factor in many conditions monitoring<br />
applications. Therefore, computed order tracking<br />
techniques are introduced and applied in this paper.<br />
The objective <strong>of</strong> computed order tracking (COT) [9] is<br />
a calculation <strong>of</strong> the vibration signal sampled constant in<br />
angle from sampled constant in time. From the<br />
mathematical point <strong>of</strong> view, this task could be solved by<br />
interpolation theory.<br />
To determine the resample times, it will be assumed<br />
that the shaft is undergoing constant angular acceleration.<br />
With this basis, the shaft angle θ (t)<br />
can be described by<br />
a quadratic equation <strong>of</strong> the following form [9]:<br />
2<br />
θ ( t ) = b0<br />
+ b1t<br />
+ b2t<br />
(3)<br />
where b 0,b 1 and b3 are unknown coefficients, which are<br />
found by fitting three successive key-phasor arrival times<br />
( 1 t , 1 t and t 3 ) which occur at known shaft angle<br />
increments ∆ φ . This can be obtained by the following<br />
conditions:
1996 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
⎧ θ ( t1)<br />
= 0<br />
⎪<br />
⎨ θ ( t2<br />
) = ∆φ<br />
⎪<br />
⎩θ<br />
( t3<br />
) = 2∆φ<br />
Substituting these conditions into Eq. (3) and arranging<br />
in a matrix format gives,<br />
(4)<br />
2<br />
⎛ 0 ⎞ ⎡1<br />
t ⎤ 1 t1<br />
⎧b0<br />
⎫<br />
⎜ ⎟ ⎢<br />
2 ⎥⎪<br />
⎪<br />
⎜ ∆φ<br />
⎟ = ⎢1<br />
t2<br />
t2<br />
⎥⎨b1<br />
⎬ (5)<br />
⎜ ⎟ ⎢<br />
2 ⎥⎪<br />
⎪<br />
⎝2∆φ<br />
⎠ ⎣1<br />
t3<br />
t3<br />
⎦⎩b2<br />
⎭<br />
This set <strong>of</strong> equations is then solved for the unknown<br />
b components. Once these values are known, Eq.(3)<br />
{ }<br />
i<br />
may be solved for t , yielding<br />
2<br />
2<br />
[ 4b2<br />
( k∆<br />
− b0<br />
) + b1<br />
b1<br />
]<br />
1<br />
t = θ − (6)<br />
2b<br />
where k is the interpolation coefficient which can be<br />
obtained as follow<br />
θ = k ∆θ<br />
(7)<br />
where θ is the shaft angle and ∆ θ is the desired<br />
angular spacing between re-samples.<br />
Once the resample times are calculated, the<br />
corresponding amplitudes <strong>of</strong> the signal are calculated by<br />
interpolating between the sampled data. After the<br />
amplitudes are determined, the re-sample data are<br />
transformed from the angle domain to the order domain<br />
by means <strong>of</strong> an FFT.<br />
The order spectrum and order bi-spectrum techniques<br />
are based on the signal processing <strong>of</strong> the angle domain<br />
signal, where the resample signal is in accordance with<br />
the shaft angle <strong>of</strong> the gearbox. The order spectrum and<br />
order bi-spectrum are then evaluated for the resample<br />
signal. The usefulness <strong>of</strong> this approach will be shown<br />
with an experimental example in Section VI.<br />
IV. PROPOSED ORDER BI-SPECTRUM METHOD FOR<br />
FAULTS DETECTION OF BEARING<br />
The procedure <strong>of</strong> proposed order bi-spectrum method<br />
is given as follows:<br />
1) Non-stationary vibration signal under run-up<br />
condition is sampled using a constant time increment;<br />
2) Non-stationary vibration signal is re-sampled at a<br />
constant angle increment. Then the non-stationary<br />
vibration signal in time domain is transformed into<br />
stationary one in angle domain;<br />
3) To demodulate the constant angle increment signal<br />
using Hilbert transform;<br />
4) The order bi-spectrum is calculated according to Eq.<br />
(2);<br />
5) The diagnostic conclusions are drawn according to<br />
the order bi-spectrum.<br />
© 2011 ACADEMY PUBLISHER<br />
V. EXPERIMENTAL SET-UP<br />
The test apparatus used in this study is shown in Fig.1<br />
[24,15]. The experimental set-up consists <strong>of</strong> a singlestage<br />
gearbox, driven by a 4.5 kW AC governor motor.<br />
The driving gear has 30 teeth and the driven gear has 50<br />
teeth. Therefore, the transmission ratio is 50/30, which<br />
means that an decrease in rotation speed is achieved. The<br />
module <strong>of</strong> the gear is 2.5 mm. The monitoring and<br />
diagnostic system is composed <strong>of</strong> three accelerometers,<br />
amplifiers, a speed and torque transducer, B&K 3560<br />
spectrum analyzer and a computer. The sampling span is<br />
3.2 kHz, the sampling frequency is 8192 Hz and the<br />
sampling time is 2 seconds. This time included one speed<br />
up <strong>of</strong> the gearbox from idle speed up to steady. After<br />
sampling, the measured vibration signals were loaded<br />
into MATLAB from data-files. Then, the vibration<br />
signals were re-samples. For their re-sampling, the<br />
algorithm described in the previous section was used. As<br />
a result <strong>of</strong> experiment, the vibration signals generated by<br />
the tested gearbox were obtained sampled constant in<br />
time as well as sampled constant in angle.<br />
Figure 1. Experimental set-up<br />
VI. BEARING FAULTS DIAGNOSIS BASED ON ORDER BI-<br />
SPECTRUM<br />
In this section, the order power spectrum and order bispectrum<br />
will be applied to vibration signal analysis<br />
measured from a gearbox during speed-up process.<br />
Ball bearings are installed in many kinds <strong>of</strong> machinery.<br />
Many problem <strong>of</strong> those machines may be caused by<br />
defects <strong>of</strong> the ball bearing. Generally, localized defects<br />
may occur on inner race, outer race or rollers <strong>of</strong> bearing.<br />
A local fault may produce periodic impacts, the size and<br />
the repetition period which are determined by the shaft<br />
rotation speed, the type <strong>of</strong> fault and the geometry <strong>of</strong> the<br />
bearing. The successive impacts produce a series <strong>of</strong><br />
impulse response, which maybe amplitude modulated<br />
because <strong>of</strong> the passage <strong>of</strong> fault through the load zone. The<br />
spectrum <strong>of</strong> such a signal would consists <strong>of</strong> a harmonics<br />
series <strong>of</strong> frequency components spaced at the component<br />
fault frequency with the highest amplitude around the<br />
resonance frequency. These frequency components are<br />
flanked by sidebands if there is an amplitude modulation<br />
due to the load zone. According to the period <strong>of</strong> the<br />
impulse, we can judge the location <strong>of</strong> the defect using<br />
characteristic frequency formulae.
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1997<br />
The tested bearing was used to study only one kind <strong>of</strong><br />
surface failure: the bearing was damaged on the inner<br />
race or outer race. The ball bearing tested has a groove on<br />
the inner race or outer race. Localized defect was seed on<br />
the inner race or outer race by an electric-discharge<br />
machine to keep their size and depth under control. The<br />
size <strong>of</strong> the artificial defect was 1mm in depth and the<br />
width <strong>of</strong> the groove was 1.5 mm. The type <strong>of</strong> the ball<br />
bearing is 206. There are 9 balls (z=9) in a bearing and<br />
the contact angle α = 0°<br />
, ball diameter d=9.5mm,<br />
bearing pitch diameter D=41.75mm. Then the<br />
characteristic frequency <strong>of</strong> the inner race or outer race<br />
defect can be calculated according to the Eq.(8), Eq.(9),<br />
respectively.<br />
z ⎛ d ⎞<br />
= + α f<br />
finner ⎜1<br />
cos ⎟<br />
2 ⎝ D ⎠<br />
z ⎛ d ⎞<br />
= − α f<br />
fouter ⎜1<br />
cos ⎟<br />
2 ⎝ D ⎠<br />
where f r is the rotating frequency <strong>of</strong> the input shaft.<br />
Therefore, according to Eq.(8) and Eq.(9), the<br />
characteristic frequency <strong>of</strong> the inner race and outer race<br />
defect are given as follows:<br />
inner<br />
r<br />
r<br />
r<br />
(8)<br />
(9)<br />
f = 5.<br />
42 f<br />
(10)<br />
f = 3.<br />
58 f<br />
outer<br />
r<br />
(11)<br />
Then the characteristic order <strong>of</strong> the inner race and<br />
outer race are obtained as follows:<br />
O = 5.<br />
42<br />
(12)<br />
inner<br />
O = 3.<br />
58<br />
(13)<br />
outer<br />
A. Application <strong>of</strong> Order Bi-spectrum to Fault Diagnosis<br />
<strong>of</strong> Inner Race<br />
The rotating speed signal <strong>of</strong> the input shaft for the<br />
tested gearbox is displayed in Fig.2. Fig.2 (a) represents<br />
the sampling pluses <strong>of</strong> the input shaft from the optical<br />
encoder (60 pulses per rotational period). The encoder<br />
signals consist <strong>of</strong> 16384 points and have a total duration<br />
<strong>of</strong> 2 seconds. To obtain approximate values <strong>of</strong> rotational<br />
speed for every data point, polynomial curve fitting was<br />
used. It was found that linear approximate was sufficient<br />
for this research. polynomial coefficients were<br />
determined for each data and analytical descriptions <strong>of</strong><br />
the rotational speed were obtained. Fig.2 (b) is the<br />
calculated instantaneous rotating speed using<br />
interpolating method. Fig.2 (b) clearly shows that the<br />
rotating speed <strong>of</strong> the input shaft runs up from idle to<br />
steady speed about 700 rpm.<br />
The original vibration signal with inner race fault is<br />
displayed in Fig.3 (a). Fig.3 (a) shows that the vibration<br />
signals are non-stationary which the amplitude <strong>of</strong> the<br />
© 2011 ACADEMY PUBLISHER<br />
vibration is increasing during the input shaft speed up.<br />
The result <strong>of</strong> applying conventional spectral analysis<br />
(FFT) to the specified non-stationary signal is given in<br />
Fig.3 (b). Fig.3 (b) displays the FFT <strong>of</strong> the vibration<br />
signals with inner race fault. It is very clear that the<br />
resulting spectrum is significantly obscured by spectral<br />
smearing. Besides, traditional spectral averaging cannot<br />
be applied to the non-stationary signal during the input<br />
shaft run-up process. Fig.3 (b) clearly shows that spectral<br />
smearing substantially affects the result <strong>of</strong> conventional<br />
analysis based on time sampling. Therefore, classical<br />
Fourier analysis has some limitation such as being unable<br />
to process non-stationary signals.<br />
Figure 2. Rotating speed <strong>of</strong> the input shaft<br />
Figure 3. Time-domain vibration signal with inner race fault and FFT<br />
Figure 4. Angular resample signal <strong>of</strong> Fig.3 (a)
1998 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
The angular re-sampling technique is applied to the<br />
vibration signal <strong>of</strong> Fig.3 (a). Fig.4 displays the re-sample<br />
vibration signal with uniform angular increment <strong>of</strong><br />
0.008722 rad. It is clear that there are periodic impacts in<br />
the angle domain vibration signal. There are significant<br />
fluctuations in the peak amplitude <strong>of</strong> the signal. However,<br />
it is hardly possible to evaluate the bearing fault condition<br />
only through such angle domain vibration signal. Fig.5<br />
shows the order power spectrum <strong>of</strong> the re-sample<br />
vibration signal. The order power spectrum, as shown in<br />
Fig.5, is dominated by the repetition order <strong>of</strong> the gear<br />
mesh order and its harmonics. It can be seen from Fig.5,<br />
that the order power spectrum represents the complicated<br />
quantities. This complexity <strong>of</strong> the order power spectrum<br />
follows from the frequency smearing and modulation<br />
effects. Therefore, the conventional order power<br />
spectrum was not capable <strong>of</strong> revealing the characteristic<br />
order <strong>of</strong> inner race fault that was corrupted by the<br />
modulation and noise.<br />
Figure 5. Order spectrum <strong>of</strong> inner race fault<br />
Figure 6. Order bi-spectrum <strong>of</strong> inner race fault (contour)<br />
The order bi-spectrum was evaluated according to the<br />
conventional direct method [2] after the re-sample signal<br />
has been demodulated by Hilbert transform. The order bispectrum<br />
is depicted in Fig.6 (contour plot) and Fig.7<br />
(mesh plot). From Fig.6 we can see that the graphs <strong>of</strong> the<br />
order quantities are much simple than that <strong>of</strong> the order<br />
power spectrum <strong>of</strong> Fig.5. In case <strong>of</strong> the order bi-spectrum,<br />
it can be identified that the characteristic order ( O ) <strong>of</strong><br />
inner<br />
inner race fault and its harmonics are represented clearly<br />
in the order bi-spectrum. The simplicity <strong>of</strong> the order<br />
© 2011 ACADEMY PUBLISHER<br />
quantity representation can be put down to the ability <strong>of</strong><br />
the order signal processing method to eliminate<br />
undesirable spectral smearing and modulation effects.<br />
Fig.6 and Fig.7 demonstrate the advantage <strong>of</strong> the order<br />
quantity application for the analysis vibration signals<br />
generated by gearbox under running up condition.<br />
Especially, the order bi-spectrum better identifies the<br />
order components and consequently leads to a better<br />
understanding <strong>of</strong> the transient vibration characteristics<br />
than that <strong>of</strong> the order power spectrum.<br />
Figure 7. Order bi-spectrum <strong>of</strong> inner race fault (mesh)<br />
B. Application <strong>of</strong> Order Bi-spectrum to Fault Diagnosis<br />
<strong>of</strong> Outer Race<br />
Figure 8. Time-domain vibration signal with outer race fault and FFT<br />
Figure 9. Angular resample signal <strong>of</strong> Fig.8 (a)
JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1999<br />
Figure 8(a) shows the original vibration signal with<br />
outer race fault during the input shaft speed-up. Fig.8 (b)<br />
displays the FFT <strong>of</strong> the vibration signal with outer race<br />
fault. It is clear that the resulting spectrum is the same as<br />
the inner race fault that is significantly obscured by<br />
spectral smearing.<br />
Figure 10. Order spectrum <strong>of</strong> outer race fault<br />
Figure 11. Order bi-spectrum <strong>of</strong> outer race fault (contour)<br />
Figure 12. Order bi-spectrum <strong>of</strong> outer race fault (mesh)<br />
Figure 9 displays the re-sample vibration signal with<br />
uniform angular increment. Fig.10 is the order power<br />
spectrum <strong>of</strong> the re-sample vibration signal. The<br />
conventional order power spectrum was not capable <strong>of</strong><br />
© 2011 ACADEMY PUBLISHER<br />
revealing the characteristic order <strong>of</strong> outer race fault in the<br />
same way. The order bi-spectrum is depicted in Fig.11<br />
(contour plot) and Fig.12 (mesh plot), respectively. It can<br />
be seen clearly from Fig.11 and Fig.12 that there are the<br />
characteristic order ( O outer ) <strong>of</strong> outer race fault and its<br />
harmonics. Therefore, the outer race fault can be easily<br />
detected by using order bi-spectrum. Fig.11 and Fig.12<br />
demonstrate the advantage <strong>of</strong> the order bi-spectrum for<br />
the analysis vibration signals generated by gearbox<br />
during run-up process.<br />
VII. CONCLUSIONS<br />
A method for fault diagnosis <strong>of</strong> bearing under run-up<br />
condition was presented based on a newly developed<br />
signal processing technique termed as computed order<br />
tracking and order bi-spectrum. Using computed ordertracking<br />
technique, the non-stationary vibration signals <strong>of</strong><br />
bearing faults in time domain can be transformed into<br />
stationary ones in the angle domain. The definition <strong>of</strong> the<br />
order bi-spectrum for analysis <strong>of</strong> vibration signals<br />
generated by rotating machinery was introduced. This<br />
method is based on the bi-spectrum estimation from the<br />
vibration signal sampled constant in angle with respect to<br />
the shaft speed <strong>of</strong> the gearbox. The order bi-spectrum<br />
method assists in the elimination <strong>of</strong> spectral smearing and<br />
modulation effects caused by the variation in shaft speed.<br />
The experimental results show that order bi-spectrum can<br />
be effectively used as a diagnostic feature for bearing<br />
faults.<br />
ACKNOWLEDGMENT<br />
The authors are grateful to the National Natural<br />
Science Foundation <strong>of</strong> China (No.50975185), Zhejiang<br />
Provincial Natural Science Foundation (No.Y1080040).<br />
The authors are also grateful to the editors and reviewers<br />
for their constructive comments.<br />
REFERENCES<br />
[1] E.Lopatinskaia, J.Zhu, J.Mathew, “Monitoring varying<br />
speed machinery vibration- I. The use <strong>of</strong> non-stationary<br />
recursive filters”, Mechanical Systems and Signal<br />
Processing, vol.9, no.6, pp.635-645, 1995.<br />
[2] E.Lopatinskaia,J.Zhu, J.Mathew, “Monitoring varying<br />
speed machinery vibration -II. Recursive filters and angle<br />
domain,” Mechanical Systems and Signal Processing,<br />
vol.9, no.6, pp.647-655, 1995.<br />
[3] G.Meltzer,Y.Y.Ivanov, “Fault detection in gear drives with<br />
non-stationary rotational speed- part I: the time-frequency<br />
approach,” Mechanical Systems and Signal Processing,<br />
vol.17, no.5, pp.1033-1047, 2003.<br />
[4] G.Meltzer,Y.Y.Ivanov, “Fault detection in gear drives with<br />
non-stationary rotational speed- part II: the time-quefrency<br />
approach,” Mechanical Systems and Signal Processing,<br />
vol.1, no.2, pp.273-283, 2003.<br />
[5] JianDa Wu, ChinWei Huang, Rongwen Huang, “An<br />
application <strong>of</strong> a recursive kalman filtering algorithm in<br />
rotating machinery fault diagnosis,” NDT&E International,<br />
vol.37, no.3, pp.411-419, 2004.<br />
[6] Zhinong Li, Zhaotong Wu, Yongyong He, Chu Fulei,<br />
“Hidden Markov model-based fault diagnostics method in
2000 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011<br />
speed-up and speed-down process for rotating machinery,”<br />
Mechanical Systems and Signal Processing, vol.19 no.2,<br />
pp.329-339, 2005.<br />
[7] R.Potter, M.Gribler, “Computed Order Tracking Obsoletes<br />
Older Methods,” Proceedings <strong>of</strong> the SAE Noise and<br />
Vibration conference, pp. 63-67, 1989.<br />
[8] R.Potter, “A New Order Tracking Method for Rotating<br />
Machinery,” Sound and Vibration, vol.7, pp.30-34, 1990.<br />
[9] K.R.Fyfe, E.D.S.Munck, “Analysis <strong>of</strong> computed order<br />
tracking,” Mechanical Systems and Signal Processing,<br />
vol.11, no.2, pp.187-205, 1997.<br />
[10] K. M. Bossley and R. J. Mckendrick, “Hybrid computed<br />
order tracking,” Mechanical Systems and Signal<br />
Processing, vol.13, no.4, pp.627-641, 1999.<br />
[11] J.R.Blough, “Development and analysis <strong>of</strong> time variant<br />
discrete Fourier transform order tracking,” Mechanical<br />
Systems and Signal Processing, vol.17, no.6, pp.1185-1199,<br />
2003.<br />
[12] Hui Li, Yuping Zhang, Haiqi Zheng, “Angle Order<br />
Analysis Technique for Processing Non-stationary<br />
Vibrations,” Proceedings <strong>of</strong> 7th International Symposium<br />
on Test and Measurement, vol.5, pp.4000-4003, 2007.<br />
[13] Hui Li, Yuping Zhang, “Order Tracking and AR Spectrum<br />
Based Bearing Fault Detection Under Run-up Condition,”<br />
Proceedings <strong>of</strong> the First International Congress on Image<br />
and Signal Processing, vol.5, pp.286-290, 2008.<br />
[14] J.Lin and L. Qu, “Feature extraction based on Morlet<br />
wavelet and its application for mechanical fault diagnosis,”<br />
<strong>Journal</strong> <strong>of</strong> Sound and Vibration, vol.234, no.1, pp135-148,<br />
2001.<br />
[15] Y.S.Shin and J.J.Jeon, “Pseudo Wigner-Ville timefrequency<br />
distribution and its application to machinery<br />
condition monitoring,” <strong>Journal</strong> <strong>of</strong> Shock and Vibration,<br />
vol.1, no.4, pp. 65-76, 1993.<br />
[16] Hui Li, Haiqi Zheng, Liwei Tang, “Wigner-Ville<br />
Distribution Based on EMD for Faults Diagnosis <strong>of</strong><br />
Bearing,” Lecture Notes in Computer Science, vol.4223<br />
pp.803-812, 2006.<br />
[17] W.J.Staszewski, K. Worden and G.R.Tomlinson, “Thefrequency<br />
analysis in gearbox fault detection using the<br />
Wigner-Ville distribution and pattern recognition,”<br />
Mechanical Systems and Signal Processing, vol. 11, no.5,<br />
pp. 673-692, 1997.<br />
[18] Hui Li, Yuping Zhang, Haiqi Zheng, “Hilbert-Huang<br />
transform and marginal spectrum for detection and<br />
diagnosis <strong>of</strong> localized defects in roller bearings,” <strong>Journal</strong><br />
<strong>of</strong> Mechanical Science and Technology, vol.23, no.2,<br />
pp.291-301, 2009.<br />
[19] H.Li, H.Q.Zheng, L.W.Tang, “Faults Monitoring and<br />
Diagnosis <strong>of</strong> Ball Bearing Based on Hilbert-Huang<br />
Transformation,” Key Engineering Material, vol.291,<br />
pp.649-654, 2005.<br />
[20] Hui Li, Yuping Zhang, Haiqi Zheng, “Wear Detection in<br />
Gear System Using Hilbert-Huang Transform,” <strong>Journal</strong> <strong>of</strong><br />
© 2011 ACADEMY PUBLISHER<br />
Mechanical Science and Technology, vol.20, no.11,<br />
pp.1781-1789, 2006.<br />
[21] W.B.Collis, P.R.White, J.K.Hammond, “Higher-order<br />
spectra: the bispectrum and trispectrum,” Mechanical<br />
Systems and Signal Processing, vol.12, no.3, pp.375-394,<br />
1998.<br />
[22] J.W.A.Fackrell, P.R.White, J.K.Hammond, R.J.Pinnington,<br />
“The interpretation <strong>of</strong> the bispectra <strong>of</strong> vibration signals- I.<br />
Theory,” Mechanical Systems and Signal Processing, vol.9,<br />
no.3, pp. 257-266, 1995.<br />
[23] J.W.A.Fackrell, P.R.White, J.K.Hammond, R.J.Pinnington,<br />
“The interpretation <strong>of</strong> the bispectra <strong>of</strong> vibration signals- II.<br />
Experimental results and application,” Mechanical Systems<br />
and Signal Processing, vol.9, no.3, pp.267-274, 1995.<br />
[24] Hui Li, Haiqi Zheng, Liwei Tang, “Bearing Fault<br />
Detection and Diagnosis Based on Order Tracking and<br />
Teager-Huang Transform,” <strong>Journal</strong> <strong>of</strong> Mechanical Science<br />
and Technology, vol.24, no.3, pp.811-822, 2010.<br />
[25] Hui Li, Haiqi Zheng, Liwei Tang, “Bearing Fault<br />
Detection and Diagnosis Based on Teager-Huang<br />
Transform,” International <strong>Journal</strong> <strong>of</strong> Wavelets,<br />
Multiresolution and Information Processing, vol.7, no.5,<br />
pp.643-663, 2009.<br />
Hui Li was born in Hebei province,<br />
China, on August 23,1968. He received<br />
his B.S. degree in Mechanical<br />
Engineering from the Hebei Polytechnic<br />
University, Hebei, China, in 1991. He<br />
received his M.S. degree in Mechanical<br />
Engineering from the Harbin University<br />
<strong>of</strong> Science and Technology, Heilongjiang,<br />
China, in 1994. He received his PhD<br />
from the School <strong>of</strong> Mechanical<br />
Engineering <strong>of</strong> Tianjin University, Tianjin, China, in 2003. He<br />
was a postdoctoral researcher in Shijiazhuang Mechanical<br />
Engineering College from August 2003 to September 2005, and<br />
in Beijing Jiaotong University from March 2006 to December<br />
2008.<br />
He is currently a pr<strong>of</strong>essor in Mechanical Engineering at<br />
Shijiazhuang Institute <strong>of</strong> Railway Technology, China. His<br />
research and teaching interests include hybrid driven<br />
mechanism, kinematics and dynamics <strong>of</strong> machinery,<br />
mechatronics, CAD/CAPP, signal processing for machine<br />
health monitoring, diagnosis and prognosis. He has written<br />
more than 170 papers and conference proceedings.<br />
Dr. Li is currently a senior member <strong>of</strong> the Chinese Society <strong>of</strong><br />
Mechanical Engineering.
Aims and Scope.<br />
Call for Papers and Special Issues<br />
<strong>Journal</strong> <strong>of</strong> <strong>Computers</strong> (JCP, ISSN 1796-203X) is a scholarly peer-reviewed international scientific journal published monthly for researchers,<br />
developers, technical managers, and educators in the computer field. It provide a high pr<strong>of</strong>ile, leading edge forum for academic researchers, industrial<br />
pr<strong>of</strong>essionals, engineers, consultants, managers, educators and policy makers working in the field to contribute and disseminate innovative new work<br />
on all the areas <strong>of</strong> computers.<br />
JCP invites original, previously unpublished, research, survey and tutorial papers, plus case studies and short research notes, on both applied and<br />
theoretical aspects <strong>of</strong> computers. These areas include, but are not limited to, the following:<br />
• Computer Organizations and Architectures<br />
• Operating Systems, S<strong>of</strong>tware Systems, and Communication Protocols<br />
• Real-time Systems, Embedded Systems, and Distributed Systems<br />
• Digital Devices, Computer Components, and Interconnection Networks<br />
• Specification, Design, Prototyping, and Testing Methods and Tools<br />
• Artificial Intelligence, Algorithms, Computational Science<br />
• Performance, Fault Tolerance, Reliability, Security, and Testability<br />
• Case Studies and Experimental and Theoretical Evaluations<br />
• New and Important Applications and Trends<br />
Special Issue Guidelines<br />
Special issues feature specifically aimed and targeted topics <strong>of</strong> interest contributed by authors responding to a particular Call for Papers or by<br />
invitation, edited by guest editor(s). We encourage you to submit proposals for creating special issues in areas that are <strong>of</strong> interest to the <strong>Journal</strong>.<br />
Preference will be given to proposals that cover some unique aspect <strong>of</strong> the technology and ones that include subjects that are timely and useful to the<br />
readers <strong>of</strong> the <strong>Journal</strong>. A Special Issue is typically made <strong>of</strong> 10 to 15 papers, with each paper 8 to 12 pages <strong>of</strong> length.<br />
The following information should be included as part <strong>of</strong> the proposal:<br />
• Proposed title for the Special Issue<br />
• Description <strong>of</strong> the topic area to be focused upon and justification<br />
• Review process for the selection and rejection <strong>of</strong> papers.<br />
• Name, contact, position, affiliation, and biography <strong>of</strong> the Guest Editor(s)<br />
• List <strong>of</strong> potential reviewers<br />
• Potential authors to the issue<br />
• Tentative time-table for the call for papers and reviews<br />
If a proposal is accepted, the guest editor will be responsible for:<br />
• Preparing the “Call for Papers” to be included on the <strong>Journal</strong>’s Web site.<br />
• Distribution <strong>of</strong> the Call for Papers broadly to various mailing lists and sites.<br />
• Getting submissions, arranging review process, making decisions, and carrying out all correspondence with the authors. Authors should be<br />
informed the Instructions for Authors.<br />
• Providing us the completed and approved final versions <strong>of</strong> the papers formatted in the <strong>Journal</strong>’s style, together with all authors’ contact<br />
information.<br />
• Writing a one- or two-page introductory editorial to be published in the Special Issue.<br />
Special Issue for a Conference/Workshop<br />
A special issue for a Conference/Workshop is usually released in association with the committee members <strong>of</strong> the Conference/Workshop like<br />
general chairs and/or program chairs who are appointed as the Guest Editors <strong>of</strong> the Special Issue. Special Issue for a Conference/Workshop is<br />
typically made <strong>of</strong> 10 to 15 papers, with each paper 8 to 12 pages <strong>of</strong> length.<br />
Guest Editors are involved in the following steps in guest-editing a Special Issue based on a Conference/Workshop:<br />
• Selecting a Title for the Special Issue, e.g. “Special Issue: Selected Best Papers <strong>of</strong> XYZ Conference”.<br />
• Sending us a formal “Letter <strong>of</strong> Intent” for the Special Issue.<br />
• Creating a “Call for Papers” for the Special Issue, posting it on the conference web site, and publicizing it to the conference attendees.<br />
Information about the <strong>Journal</strong> and <strong>Academy</strong> <strong>Publisher</strong> can be included in the Call for Papers.<br />
• Establishing criteria for paper selection/rejections. The papers can be nominated based on multiple criteria, e.g. rank in review process plus<br />
the evaluation from the Session Chairs and the feedback from the Conference attendees.<br />
• Selecting and inviting submissions, arranging review process, making decisions, and carrying out all correspondence with the authors.<br />
Authors should be informed the Author Instructions. Usually, the Proceedings manuscripts should be expanded and enhanced.<br />
• Providing us the completed and approved final versions <strong>of</strong> the papers formatted in the <strong>Journal</strong>’s style, together with all authors’ contact<br />
information.<br />
• Writing a one- or two-page introductory editorial to be published in the Special Issue.<br />
More information is available on the web site at http://www.academypublisher.com/jcp/.
A Modified Technique for Analysis <strong>of</strong> Synchronous Counters Constructed with Flip-flops<br />
Dangui Yan, Ruijun Tong, Chengchang Zhang, and Changyong Li<br />
A New Method <strong>of</strong> Detecting Multi-component LFM Signals Based on Blind Signal Processing<br />
Qiang Guo, Yajun Li, and Changhong Wang<br />
Research on Self-built Digital Resource Backup Systems<br />
Li-zhen Shen<br />
Configuration Scheme for Small Scale Multi-FPGA Systems<br />
Chengchang Zhang, Lisheng Yang, Dangui Yan, and Changyong Li<br />
Order Bi-spectrum For Bearing Fault Monitoring and Diagnosis Under Run-up Condition<br />
Hui Li<br />
1971<br />
1976<br />
1983<br />
1988<br />
1994
(Contents Continued from Back Cover)<br />
The Analysis <strong>of</strong> China New Energy Vehicle Industry Alliance Status based on UCINET S<strong>of</strong>tware<br />
Xiongfei Guo and Yingqi Liu<br />
Efficiency Evaluation Information System Based on Data Envelopment Analysis<br />
Jing Han and Malin Song<br />
An Optimal Inventory Control Model for a Supply Chain with Shortage Constraints<br />
Yinkuan Gu and Hongxia Zhang<br />
Variable Selection for Credit Risk Model Using Data Mining Technique<br />
Kuangnan Fang and Hong Huang<br />
Corporate-, Product-, and User-Image Dimensions and Purchase Intentions —The Mediating Role <strong>of</strong><br />
Cognitive and Affective Attitudes<br />
Xian Guo Li, Xia Wang, and Yu Juan Cai<br />
A Microcomputer-Based Predictive Digital Current Programmed Control System for Three-phase<br />
PWM Rectifier<br />
Zhongjiu Zheng, Gu<strong>of</strong>eng Li, and Ninghui Wang<br />
Supply Chain Coordination under Return Policy with Asymmetric Information about Cost <strong>of</strong> Reverse<br />
Logistics Operations<br />
Ting Long Zhang<br />
Economic Development and Financial Support for Coal Resource Cities — A Panel Data Analysis<br />
Zuhuai Yuan, Li Yang, Jing Han, and Keliang Wang<br />
REGULAR PAPERS<br />
Solving the Sparsity Problem in Recommender Systems Using Association Retrieval<br />
YiBo Chen, ChanLe Wu, Ming Xie and Xiaojun Guo<br />
Integrated Structure and Control Design for Servo System Based on Genetic Algorithm and Matlab<br />
Dingzhen Li and Ruimin Jin<br />
A Model to Select System Core and Its Application<br />
Chongming Li and Yue Ding<br />
De-noise Comprehensive Research On Airplane Cockpit Signals Recorded by CVR<br />
Dao-Lai Cheng, Chui-JieYi, and Hong-Yu Yao<br />
Fuzzy Support Vector Machines Control for 6-DOF Parallel Robot<br />
Dequan Zhu, Tao Mei, and Lei Sun<br />
Parameters Optimization <strong>of</strong> Least Squares Support Vector Machines and Its Application<br />
Chunli Xi, Cheng Shao, and Dandan Zhao<br />
The Expected Value Model <strong>of</strong> Multiobjective Programming and its Solution Method Based on<br />
Bifuzzy Environment<br />
Mingfa Zheng , Bingjie Li, and Guangxing Kou<br />
A Method for Building Partially Connected Neural Network<br />
Gang Li, Xingsan Qian, Chunming Ye, and Lin Zhao<br />
A Cooperative Co-evolution PSO for Flow Shop Scheduling Problem with Uncertainty<br />
Bin Jiao, Qunxian Chen, and Shaobin Yan<br />
A Double Margin Based Fuzzy Support Vector Machine Algorithm<br />
Kai Li and Xiaoxia Lu<br />
1852<br />
1857<br />
1862<br />
1868<br />
1875<br />
1880<br />
1886<br />
1891<br />
1896<br />
1903<br />
1913<br />
1920<br />
1926<br />
1935<br />
1942<br />
1949<br />
1955<br />
1962