Journal of Computers - Academy Publisher

Journal of Computers 

ISSN 1796-203X 

Volume 6, Number 9, September 2011 

Contents 

Special Issue: Changes in Computer Application for Economic Analysis of Law and Business 

Management 

Guest Editor: Malin Song, Dingding Pan, Jie Wu, Li Yang, Hongping Zhou, and Christopher 

Clemence 

Guest Editorial 

Malin Song, Dingding Pan, Jie Wu, Li Yang, Hongping Zhou, and Christopher Clemence 

SPECIAL ISSUE PAPERS 

Tax Evasion, Taxation Inspection and Net Tax Revenue: from an Optimal Tax Administration 

Perspective 

Bing Liu 

The Developmental Analysis of China’s Information Technology Services 

Wei Gao, Feng Wang, and Li Wang 

A Web Survey Program Based on Computer Technology and Its Application to Evaluation Model 

about Youth Self-organizations in China 

Ma-lin Song, Tong Yang, and Ya-qing Song 

The Research on the Influencing Factors of Financing Strategy of Woman Entrepreneurs in China 

Xiong Xiong, Rong Fu, Wei Zhang, Yongjie Zhang, and Lin Xiong 

A Spatial Econometric Analysis of China’s Manufacturing Agglomeration based on Geoda and 

Matlab 

Huayin Yu and Weiping Gu 

Application of Computer Technology in Efficiency Analysis of China Life Insurance Company 

Hongling Wu and XiaoFei Zeng 

A Bayesian Belief Net Model for Evaluating Organizational Safety Risks 

Li Song, Li Yang, Jing Han, and Jinkai Li 

Research and Application of J2EE and AJAX Technologies in Industry Report 

Min Hu, Ding-ding Pan, and Pei-en Zhou 

The Analysis of China New Energy Vehicle Industry Alliance Status based on UCINET Software 

Xiongfei Guo and Yingqi Liu 

Efficiency Evaluation Information System Based on Data Envelopment Analysis 

Jing Han and Malin Song 

An Optimal Inventory Control Model for a Supply Chain with Shortage Constraints 

Yinkuan Gu and Hongxia Zhang 

1797 

1799 

1805 

1812 

1819 

1825 

1832 

1842 

1847 

1852 

1857 

1862

Variable Selection for Credit Risk Model Using Data Mining Technique 

Kuangnan Fang and Hong Huang 

Corporate-, Product-, and User-Image Dimensions and Purchase Intentions —The Mediating Role of 

Cognitive and Affective Attitudes 

Xian Guo Li, Xia Wang, and Yu Juan Cai 

A Microcomputer-Based Predictive Digital Current Programmed Control System for Three-phase 

PWM Rectifier 

Zhongjiu Zheng, Guofeng Li, and Ninghui Wang 

Supply Chain Coordination under Return Policy with Asymmetric Information about Cost of Reverse 

Logistics Operations 

Ting Long Zhang 

Economic Development and Financial Support for Coal Resource Cities — A Panel Data Analysis 

Zuhuai Yuan, Li Yang, Jing Han, and Keliang Wang 

REGULAR PAPERS 

Solving the Sparsity Problem in Recommender Systems Using Association Retrieval 

YiBo Chen, ChanLe Wu, Ming Xie and Xiaojun Guo 

Integrated Structure and Control Design for Servo System Based on Genetic Algorithm and Matlab 

Dingzhen Li and Ruimin Jin 

A Model to Select System Core and Its Application 

Chongming Li and Yue Ding 

De-noise Comprehensive Research On Airplane Cockpit Signals Recorded by CVR 

Dao-Lai Cheng, Chui-JieYi, and Hong-Yu Yao 

Fuzzy Support Vector Machines Control for 6-DOF Parallel Robot 

Dequan Zhu, Tao Mei, and Lei Sun 

Parameters Optimization of Least Squares Support Vector Machines and Its Application 

Chunli Xi, Cheng Shao, and Dandan Zhao 

The Expected Value Model of Multiobjective Programming and its Solution Method Based on 

Bifuzzy Environment 

Mingfa Zheng , Bingjie Li, and Guangxing Kou 

A Method for Building Partially Connected Neural Network 

Gang Li, Xingsan Qian, Chunming Ye, and Lin Zhao 

A Cooperative Co-evolution PSO for Flow Shop Scheduling Problem with Uncertainty 

Bin Jiao, Qunxian Chen, and Shaobin Yan 

A Double Margin Based Fuzzy Support Vector Machine Algorithm 

Kai Li and Xiaoxia Lu 

A Modified Technique for Analysis of Synchronous Counters Constructed with Flip-flops 

Dangui Yan, Ruijun Tong, Chengchang Zhang, and Changyong Li 

A New Method of Detecting Multi-component LFM Signals Based on Blind Signal Processing 

Qiang Guo, Yajun Li, and Changhong Wang 

1868 

1875 

1880 

1886 

1891 

1896 

1903 

1913 

1920 

1926 

1935 

1942 

1949 

1955 

1962 

1971 

1976

Research on Self-built Digital Resource Backup Systems 

Li-zhen Shen 

Configuration Scheme for Small Scale Multi-FPGA Systems 

Chengchang Zhang, Lisheng Yang, Dangui Yan, and Changyong Li 

Order Bi-spectrum For Bearing Fault Monitoring and Diagnosis Under Run-up Condition 

Hui Li 

1983 

1988 

1994

JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 1797 

Special Issue on Changes in Computer Application 

for Economic Analysis of Law and Business Management 

Guest Editorial 

Nowadays, the computer is ubiquitous in the business world; many new computer applications have been developed 

for economic analysis of law and business management evaluation and forecasting in the past several years. With the 

rise of the Internet, computing is ever more integral to many disciplines than ever. As the average business becomes 

more and more computerized, so to does the science of studying business and economic analysis of law. These fields 

benefited greatly from the rise in quantity and power of the computer as well. 

“Tax Evasion, Taxation Inspection and Net Tax Revenue: from an Optimal Tax Administration Perspective” 

discusses the tax evasion and builds a general equilibrium model. In this paper, the interaction between tax declaration 

and taxation inspection is analyzed, and some policies and proposals about taxation inspection are proposed. 

“The Developmental Analysis of China’s Information Technology Services” makes an analysis on the 

development features, significance, present situation and existed problems of information technology services in China, 

and gives some relative suggestions on how to develop the information services of China better. 

“A Web Survey Program Based on Computer Technology and Its Application to Evaluation Model about 

Youth Self-organizations in China” studies the network and youth self-organizations based on web-platform, forecasts 

the developmental trend of adolescents by analyzing their current situation in China, and builds the evolution model for 

youth self-organizations. 

“The Research on the Influencing Factors of Financing Strategy of Woman Entrepreneurs in China” examines 

gender differences among Chinese entrepreneurs seeking financing pattern, studies the factors those affect women 

entrepreneurs’ financing strategies. 

“A Spatial Econometric Analysis of China’s Manufacturing Agglomeration based on Geoda and Matlab” uses 

spatial econometric methods to analyze the influencing factors of China’s provincial manufacturing Agglomeration by 

Geoda software and Matlab network tools. 

“Application of Computer Technology in Efficiency Analysis of China Life Insurance Company” aims at 

studying the application of computer technology in efficiency analysis of China life insurance company. 

“A Bayesian Belief Net Model to evaluating Organizational Safety Risks” presents a methodology for 

organizational risk analysis for safety management. 

“Research and Application of J2EE and AJAX Technologies in Industry Report” analyzes the weakness of the 

traditional industry report, and proposes an industry report system based on J2EE and AJAX technologies. 

“The Analysis of China New Energy Vehicle Industry Alliance Status based on UCINET Software” uses the 

software UCINET to draw up the picture of China’s new energy vehicle industry alliance network, and studies the 

cooperation relationships within the alliances through analyzing their elements. 

“Efficiency Evaluation Information System Based on Data Envelopment Analysis” studies the data envelopment 

analysis, and demonstrates the bridge between DEA and MIS for building efficiency evaluation information system. 

“An Optimal Inventory Control Model for a Supply Chain with Shortage Constraints” researches on the 

inventory decision model of the minimum total annual cost of the supply chain. 

“Variable Selection for Credit Risk Model Using Data Mining Technique” estimates long term default probability 

for developing appropriate credit risk model with the estimated default probability using Transition Matrix and mapping 

methods. 

“Corporate-, Product-, and User-Image Dimensions and Purchase Intentions” investigates the effects of 

corporate-, product-, and user image dimensions on purchase intensions with cognitive and affective attitudes as 

mediator, and conducts a questionnaire survey. 

“A Microcomputer-Based Predictive Digital Current Programmed Control System for Three-phase PWM 

Rectifier” sets up a microcomputer control system for three-phase PWM rectifier using the floating-point digital signal 

processor TMS320LF2407 from Texas Instruments. 

“Supply Chain Coordination under Return Policy with Asymmetric Information about Cost of Reverse 

Logistics Operations” predicts the return policy and supply chain coordination in a channel of one supplier and one 

retailer. 

“Economic Development and Financial Support for Coal Resource Cities — A Panel Data Analysis” considers 

the high industry concentration of financial resources, which leads to a decline in financial resource allocation 

efficiency. 

We hope that the readers of this Special Issue could find and would enjoy something, such as the academic ideas, 

methods and enlightening form the papers in this Special Issue. 

© 2011 ACADEMY PUBLISHER 

doi:10.4304/jcp.6.9.1797-1798

1798 JOURNAL OF COMPUTERS, VOL. 6, NO. 9, SEPTEMBER 2011 

Guest Editors: 

Dr. Malin Song is an associate professor in School of Statistics and Applied Mathematics, Anhui University of Finance and 

Economics. He is a Research Fellow in Economic Development Research Center, Anhui University of Finance and Economics. His 

major field of study includes Logistics, Environmental Economics and System Modeling and Analysis. Email: 

malinsong@gmail.com. 

Mr. Ding-ding Pan received his bachelor's degree in Computer Science and Technology from Anhui University of Architecture, 

Hefei, China (2008) and master's degree in Computer Application Technology from Hefei University of Technology, Hefei, China. 

His research interests include computer application, software engineering. Email: panding1986@sina.com. 

Dr. Jie Wu, School of Management, University of Science and Technology of China, Hefei, Anhui Province, P. R. China, 230026. 

Phone: +8613966717485, Email: wujie012@ustc.edu 

Dr. Li Yang is a professor in School of Economics and Management, Anhui University of Science & Technology, Huainan, Anhui, 

China. He is currently a doctor candidate in the School of Management at University of Science & Technology of China, Hefei, 

Anhui, China. His major field of study includes credit risk, strategic alliance and coal-mining eco-industrial park. E-mail: 

yangli081003@163.com. 

Dr. Hongping Zhou is Master Instructor in School of Engineering Science, University of Science and Technology of China. Her 

major fields of study are Computer Applications and Digital Circuit Design. She earned her Bachelors, Masters and Doctoral 

degrees in the School of Information Science and Technology, University of Science and Technology of China. Email: 

zhouhp@sina.cn. 

Dr. Christopher Clemence earned his Juris Doctorate at the University of Missouri-Kansas City School of Law. He practices 

real estate, energy, environmental and transaction law, and currently works for a Fortune 500 company in the United States. Email: 

conagher78@gmail.com. 

© 2011 ACADEMY PUBLISHER


Tax Evasion, Taxation Inspection and Net Tax 

Revenue: from an Optimal Tax Administration 

Perspective 

Bing Liu 

School of Economics and Management/Anhui Normal University, Wuhu, China 

Abstract— Tax evasion has always been an important topic 

to tax theory researchers and the department of government. 

However, existing research results are confined to the 

unilateral action of taxpayers, neglect the interaction 

between the tax declaration and taxation inspection. This 

paper, from an optimal tax administration perspective, 

builds a general equilibrium model, in which, taxation 

inspection cost, net tax revenue and taxpayers personal 

expected utility maximization, are included, to analyze the 

interaction between the tax declaration and taxation 

inspection. Then it proposes some policies and proposals 

about taxation inspection. 

Index Terms—tax evasion, taxation inspection, net tax 

revenue 

I. INTRODUCTION 

Tax evasion is the economic activities that taxpayers 

through illegal channels to reduce their tax payable. 

Large-scale tax evasion will not only affect a 

government's fiscal revenue, lead to the failure of a 

government's macroeconomic indicators, distortion of 

resource allocation and income distribution out of 

control. Tax evasion is widespread in the world, 

according to the statistics, in developed countries, 22 

high-income countries which per capita GNP more than 

8626 U.S. dollars (such as Germany, Japan, Switzerland, 

United Kingdom, the United States and other countries), 

and 10 upper-middle-income countries which per capita 

GNP between 2786-8625 U.S. dollars (such as Argentina, 

Brazil, Chile and other countries), the tax loss is about 

35%. 9 lower-middle-income countries which per capita 

GNP between 696-2785 dollar U.S. dollars (such as 

Colombia, the Czech Republic, Indonesia and other 

countries) and 5 low-income countries which per capita 

GNP below 695 U.S. dollars (including Egypt, India, 

Zambia and other countries) loss 30% and 60%. Such as 

the U.S. the tax loss between 30%- 45%, the Netherlands, 

the tax loss between 22% -35%. Japan's tax loss vary 

according to the size of the taxpayer, Large, medium and 

small taxpayers, the loss around 20%, 40%, 60%. 

In China, tax evasion is very serious and has become a 

well-known fact, according to the scholars calculation: in 

1999, China's tax revenue loss was about 77.6 billion 

RMB(China yuan), if calculated with the amount that 

should be collected, the loss exceeded 100 billion RMB, 


doi:10.4304/jcp.6.9.1799-1804 

if measurement by the combination of factors method, the 

loss will reach 320 billion to 430 billion, was the entire 

tax revenue’s 30% -40%. Experts conservatively 

estimated that, in 2004, tax loss was at least 450 billion, 

was the entire tax revenue’s 15%. in 2004, the National 

Audit Office audited 788 enterprises from 19 provinces 

and cities, found that in 2002 the tax should be collected 

was 117.35 billion, but actually the amount was 103.96 

billion, pay less tax 13.38 billion, accounting for the tax 

that should be collected 11.41%; January to September, 

2003, the tax should be collected 103.78 billion, 918.84 

billion actually collected, less 118.94 billion, accounting 

for 11.46% of tax should be collected, in this way, tax 

wastage nationwide in 2007 up to 500 billion. 

Tax evasion has been the important issues which 

theory researchers and government departments 

concerned about all the time. as China's economy 

development and the tax system improvement, more and 

more scholars research the issue. In this paper, based on 

Chinese and foreign scholars previous studies, from the 

point of tax audit cost and net revenue, study a general 

equilibrium model about tax evasion, tax inspection and 

tax net income, give some suggestions about the optimal 

behavior of tax collection for tax administration 

reference. 

II. THEORY REVIEW 

A. A-S model 

The first use of modern methods of economics to study 

the problem of tax evasion is U.S. economist Kagan.P, he 

used the cash ratio method to estimate the scale of U.S. 

tax loss in "The total money supply and the 

corresponding currency demand" (1958). M G Allingham 

and A Sandmo constructed a theoretical model, it’s 

theory based on Gary Becker’s study on economics of 

crime and A Sandmo’s research on risk and uncertainty. 

Allingham and Sandem’s tax evasion model established 

the theoretical framework of tax evasion, is a classic 

model of tax evasion, often referred to as A-S model. 

A-S model is a model of expected utility 

maximization, the basic assumptions are: (1) taxpayer’s 

cardinal utility maximization as objective function, and 

cardinal utility is a single function of income; (2) the 

taxpayer's marginal utility is positive and decreasing; (3) 

the taxpayer's actions are consistent with Von Neumann


and Morgenstern action rules under uncertainty; (4) 

proportional tax system; (5) tax authorities’ net income 

maximization with the budget constrained; (6) tax 

inspectors discovered a constant probability; (7) penalty 

based on the difference between taxable income and 

reported income, rather than the tax evasion, and the 

penalty ratio is higher than tax rate. In addition, tax 

audition will not add cost to taxpayers, and the taxpayer's 

real income after checking can be drawn. With these 

strict assumptions, the taxpayer's objective function can 

be expressed as: 

E( U ) = ( 1− 

p) 

U ( w − tX ) + pU[ 

W −Tx 

− r( 

W − X )] 

(1) 

Here, U is taxpayer’s disposable income utility, E (U) 

is expected utility; W is taxpayer’s real income, an 

exogenous variable, in the cases of incomplete 

information, the tax authority can not accurately grasp it; 

X for the taxpayer’s taxable income declared to the tax 

authority, a endogenous variable, 0 ≤ X ≤ W; p for the 

probability of taxpayer seized by the tax authority; t for 

the tax rate set by the tax authority, r is the penalty ratio 

when be investigated, 0


in the income areas, ∆ ≥ 0 

+ 

Y , Otherwise, in the loss 

areas and ∆ ≤ 0 

− 

+ 

Y , ∆Y 、 − 

∆Y can be understood as 

the wealth value changes relative to the reference point. 

By (5), (6) we have: 

+ + 0 + 

∆Y 

= Y −Y 

= Y − ( 1− 

t) 

W = t( 

W − D) 

≥ 0 

(7) 

− + 0 + 

∆Y 

= Y −Y 

= Y − ( 1− 

t) 

W = −( 

s + λt)( 

W − D) 

≤ 0 

(8) 

To (7), (8) into (4) can be obtained 

+ β 

β 

⎪⎧ 

( ∆ Y ) = [ t ( W − D ) ] , x ≥ 0 

V ( x ) = ⎨ 

β ∈ [ 0, 

1 , ]θ 

− β 

β 

⎪⎩ − θ ( − ∆ Y ) = −θ 

[ ( λ t + s)( 

W − D ) ] , x ≤ 0 >1 (9) 

Here, ∆Y is the amount of gain or loss relative to the 

reference point, θ is the aversion coefficient, seized 

probability p(D)(gain), not being seized probability 1p(D)(loss). 

According to the Prospect theory, People tend 

to give objective probability a lower or higher subjective 

probability. Therefore, with the prospect theory, 

weighting function for the loss state (being seized) is 

W p(D) 

− 

, weighting function for the gain state(not being 

+ 

seized) is 

W [ 1− 

p( 

D)] 

. According to (9) and weight 

function, the value function of taxpayer as follow: 

[ ][ ] [ ] β 

β 

+ 

− 

V( D, 

t, 

s, 

λ , θ) 

= W 1−p( 

D) 

t( 

W−D) 

−W 

p( 

D) 

θ( 

λt 

+ s)( 

W−D) 

(10) 

In prospect theory, the taxpayer's goal is to maximize 

the value function V of (10). With boundary conditions, 

the relationship between variables can be gotten. The 

relationship between tax rates and less declaring income, 

for example, tax rate increase, ∂V/∂D will be strictly 

negative, that is, underreporting of income will allow 

taxpayers to increase the value of the function. Therefore, 

the escape of will increase with the tax rate increase, 

there is a positive correlation between them, which is the 

important difference with the traditional A-S model. It 

explains, in reality, why tax evasion is widespread in the 

high-income groups. 

C. Tax evasion cost-benefit model 

Let the real income of the taxpayer to be W, taxpayers 

are likely to avoid tax by declaration of low taxable 

income, set X for its declared income, then the taxpayers’ 

hided income R for W-X, seized probability of tax 

evasion for P, 0 C, the results of 

inequality is: 

t > mP + a + s 

(14) 

That is, when taxpayers expect that tax evasion penalty 

(mp) suffered when being seized plus the operating costs 

(a) of tax evasion and psychological costs (s) are less than 

the taxes paid in accordance with the statutory tax rate, 

then the taxpayer will choose to evade. Formula (12), 

(13) are derived with R respectively, then, relative to the 

concealed income, the marginal benefit (MB) and the 

marginal cost (MC) of tax evasion are get: 

MB = t 

(15) 

MC = a + mP + s + RPm'( 

R) 

+ Rs'( 

R) 

(16) 

To maximize taxpayers’ expectation, in accordance 

with the principles of economics, the marginal benefit 

must equals to the marginal cost, it is MB = MC. If R is 

the horizontal axis, the marginal benefit curve MB can be 

expressed as a t height horizontal line, the marginal cost 

curve MC is a tilting curve up the right. By formula (16) 

and the known conditions, when R = 0, MC has the 

minimum a + mP + s, so the starting point of MC curve 

(R = 0, MC = a + mP + s) is lower than the MB curve, the 

two curves intersect at E, the R* corresponding to the 

intersection E is the best concealing amount where the tax 

escaper’s expected income maximization, and the best 

amount of tax evasion is R*t. 

However, look at the existing research results, the 

starting point confined to the consideration of unilateral 

acts of the taxpayer, ignored the interactive relationship 

between the behavior of declaration and tax audition. In 

fact, the conditions of risk selection taxpayers facing, 

such as penalty amount and the seized probability, are 

closely related to the behavior of the tax authorities’ 

audition. If only consider the risks to taxpayer regardless 

the conditions of risk, the findings will be unconvincing. 

Based on this, to maximize the government’s net tax 

revenue, this article, including the variable of tax audit 

expenses, construct a general equilibrium model which 

cover the utility’s net tax revenue maximization and the 

taxpayer’s expectation maximization, to analyze the 

interaction between the tax inspection and the tax 

evasion.


III. MODEL CONSTRUCT 

Based on the A-S model’s parameters setting, assume, 

in a tax year, the taxpayer’s reported less taxable income 

audited by the tax authorities for b(W-X), 0 ≤ b ≤ 1. 

From the perspective of tax authorities, taxpayer’s tax 

evasion is tb (W-X), accordingly, the taxpayer to pay the 

fine by rtb (W-X). Expenditures for the tax department’s 

inspection for C, the relationship between b and C can be 

expressed as b = b (C), it certainly has: 

b c 

db 

= > 0 

dC , 

That is, with the spending of the tax department 

inspection increased, the higher the taxpayer’s taxable 

income which being inspected out. 

2 

d b 

bCC = 2 

dC , 

With the increase in audit expenses, bC into decline by 

the increase, bCC turned negative from positive, 

indicating that the maximum value b exists. 

To the interaction between tax audit and the taxpayer’s 

behavior, here should be noted that, on the choice of the 

tax department’s policy instruments, in order to curb tax 

evasion, seeking to maximize the tax revenue, tax 

inspection (inspection expenses), the tax rate adjustment 

and penalty rates and other policy tools are available, but 

in the specific application, these tools are different. Tax 

and penalty rates are legal areas, generally not free to 

change, relatively speaking, tax inspection efforts and 

configuration of inspection project, can have a moderate 

change based on the subjective views of the tax 

department, so check expenditures are choice variables, 

the tax rates and punishment rate systematic exogenous 

variables. 

To be consistent with the A-S model, here set the 

actual income of the taxpayer for W, an exogenous 

variable, under the conditions of incomplete information, 

the tax authorities can not accurately grasp; X for taxable 

income the taxpayer declare to the tax authorities, an 

endogenous variable, 0 ≤ X ≤ W; p for the seized 

probability taxpayers seized by the tax authorities; t for 

the tax rate set by the tax authorities, r is the ratio of fine 

when being investigated, 0


'' 

U ( W2) 

R2 

= − > 0 

' 

U ( W2) 

, 2 1 R R > 

, only the follow equation 

being established, the tax rate increases will lead the 

taxpayer to reduce the reported income. 

rCR2bc 

k < 

rR2b( 

W − X ) − X ( R1 

− R2) 

(25) 

Discussion above show that, when the flexibility of tax 

audition expenses to marginal tax base is not high, the tax 

department can not further improve the performance of 

tax inspection, raising tax rates will induce taxpayers to 

increase tax evasion, thus reducing tax compliance. 

B. The income of taxpayer W changes 

The impaction of taxpayer’s income changes on the 

changes in taxpayer’s reporting income depends on 

taxpayer’s utility maximization behavior, the X of 

formula (21) is derived of W: 

dX ( 1 − p ) U ′ ( W 1 ) 

= 

[ ( 1 − rtb ) R 2 − R 1 ] 

dW 

tZ 

(26) 

R1 

> ( 1− 

rtb) 

When R2 

set up, the sign of formula (26) is 

positive, that is, the higher taxpayers’ income, the higher 

the income of its report. 

Formula (26) means that, the higher taxpayers’ 

income, the higher its probability of being audited. When 

there is tax evasion, the probability of its being seized and 

the cost of being punished are higher, too. Therefore, the 

higher the taxpayers’ income, the more likely an honest 

declaration. 

As for the proportion of declared income accounted for 

reported real income, according to the definition, there is: 

⎛ X 

∂ ⎜ 

⎝ W 

∂ W 

⎞ 

⎟ 

⎠ 

= 

1 

W 

2 

⎛ 

⎜ W 

⎝ 

∂ X 

∂ W 

− 

X 

⎞ 

⎟ 

⎠ 

(27) 

Formula (27) means that, if the conclusion that the 

taxpayer’s real income W and reported income X change 

in the same direction is right, whether the proportion of 

the declared income accounted for reported real income 

increases with the rising of the real income depends on 

the flexibility of declared income, it is, if the flexibility of 

the reported income greater than 1, then the proportion of 

declared income accounted for reported real income will 

increase with the rising of the real income, or if less than 

1, it will fall. 

C. Penalty rate and seized probability change 

The punitive rate and the seized probability change, its 

impaction to the taxpayers’ compliance can get by 

deriving equation (21) with r and p respectively: 

∂X 

1 ' 

'' 

b 

= { − pU ( W2 

) − ptW( 

r −1) 

U ( W2)[ 

∂r 

tZ 

b 

2 

C 

cc 

− b( 

W − X )]} > 0 

(28) 

2 

∂X 

1 ' 

' 

'' 

rtbC 

= − [ U ( W1) 

+ ( r −1) 

U ( W2) 

+ ( r −1) 

U ( W2) 

] > 0 

∂p 

tZ 

bcc 

(29) 

Formula (28) and (29) shows that, the penalty rate and 

the rate of seizures increased, both mean adverse to tax 

evasion, thereby, the taxpayer will increase the reported 

income people, and increase tax compliance at last. 


V. CONCLUSIONS 

To the taxpayer, the tax evasion should pay the relative 

cost, that is, will face the possible punitive price. The tax 

department is no exception, to increase taxes, prevent tax 

evasion, also must spend a lot of manpower and 

resources, tax inspection and tax evasion are against each 

other and influence each other. The net revenue of the 

government and the utility of the taxpayer maximize 

respectively, in the cases there exist an optimal 

equilibrium solution, whereby the paper establish a 

general equilibrium model in which the net revenue of the 

government and the utility of the taxpayer maximize 

respectively. By analyzing the model, can draw the 

following conclusions: 

(1) The equilibrium solution to the government’s 

optimal inspection expenditures, is defined as spending 

one dollar of audit costs, must be equal to the overdue tax 

and penalty when the tax evasion to be checked out. The 

equilibrium solution to the maximization of the 

taxpayer’s utility is, the paying less tax’s expected 

benefits the taxpayer obtained by reducing a unit of 

reported taxable income, must be equal to the expected 

marginal cost of paying an overdue tax and being 

punished when the tax evasion to be checked out. 

The behavior of the government and the taxpayer is the 

opposite: the taxpayer increase (decrease) the tax evasion, 

would enable the government to increase (decrease) tax 

audit expenditures; on other hand, with the government’s 

tax audit expenditures increase (decrease), would enable 

the taxpayer to increase (decrease) their tax compliance 

correspondingly. 

(2) In general, the tax rate increase, the tax evasion 

will be expanded. However, the tax rate increase will 

increase the marginal income of the tax audition 

expenditures, incentive the tax audition spending go up, 

and promote to enhance inspection efforts. Only when the 

marginal benefit of tax inspection expenditures is 0, then 

the degree of tax audition is optimal. Therefore, Whether 

or not the taxpayer evade tax depends on the flexibility of 

tax audition expenses to marginal tax base and the degree 

of risk of their being seized. When the flexibility of tax 

audition expenses to marginal tax base is not high, the tax 

department can not further improve the performance of 

tax inspection, raising tax rates will induce taxpayers to 

increase tax evasion, thus reducing tax compliance. 

(3) The higher the taxpayers’ income, the higher its 

probability of being audited. When there is tax evasion, 

the probability of its being seized and the cost of being 

punished are higher, too. Therefore, the higher the 

taxpayers’ income, the more likely to declare honestly. 

As for the change direction of the proportion of 

declared income accounted for reported real income, if 

the conclusion that the taxpayer’s real income and 

reported income change in the same direction is right, 

whether the proportion of the declared income accounted 

for reported real income increases with the rising of the 

real income depends on the flexibility of declared 

income, it is, if the flexibility of the reported income 

greater than 1, then the proportion of declared income


accounted for reported real income will increase with the 

rising of the real income, or if less than 1, it will fall. 

(4) Penalty rate and the detected rate increase means 

that the tax evasion will assume greater costs once being 

detected, which raise more awareness of the risk 

management of the taxpayer, the taxpayer will increase 

the reported income, thereby increasing tax compliance. 

Practical significance of this study is, view China’s tax 

reality, the tax loss has the characteristics of wide range, 

large number, diversity. With the diversification of 

economic entities and the diversification of mode of 

operation, the means of illegal and crime tax-related has 

become increasingly complex. In short, with the everchanging 

tactics of corporate tax evasion, tax means more 

and more hidden, tax audit work has become increasingly 

difficult. In addition, as for the power of the tax audit and 

the inspection level, the seized probability of current tax 

authorities on tax evasion cases is very low, usually not 

more than 50% [7]. For the above, if to improve the 

seizure rate is bound to increase a large number of tax 

officials and the huge audit costs, the result will not 

necessarily bring about the increase in net revenue. As for 

the tax and punishment rates increase, not only to revise 

the relevant laws, but also will increase corporation’s tax 

burden, causing social dissatisfaction, in fact, it is 

feasible, too. Therefore, how to adjust the structure of the 


tax audit and expenditure, under the conditions of the 

existing human and material resources of the tax 

authorities, to achieve the maximum of the net tax 

revenue would be the optimal orientation of the tax 

administrative act. 

REFERENCES 

[1] Allingham, M.G. & A. Sandmo. 1972, Income tax evasion: 

a theoretical analysis. Journal of Public Economics, 1, 

pp.323-338. 

[2] Yitzhaki, S.. 1974, A note on income tax evasion: a 

theoretical analysis. Journal of Public Economics, 3, 

pp.201-202. 

[3] Lin Wei, 2001,Tax collection column. Tax Research,1, 

pp.18-23. 

[4] Ali, M. M., H. W. Cecil., & J. A. Knoblett., 2001, The 

effect of tax rates and enforcement policies on tax payers 

compliance: a study of self-employed taxpayers. 

Economics Journal, 29(2), pp.186-202. 

[5] Ruan Jiafu. 2005, On the tax status of non-compliance, 

causes and countermeasures, Modern Finance, 1,pp.78-89. 

[6] Lo Kuang, Xiao Yan Fen, 2007, Consider the tax 

compliance cost of tax evasion model, Tax Research, 1, 

pp.78-96. 

[7] Zhu Feng, 2007, Research tax evasion in China. Market 

Weekly, 4, pp.32-38.


The Developmental Analysis of China’s 

Information Technology Services 

Wei Gao 

College of Statistics and Applied Mathematics /Anhui University of Finance and Economics, Bengbu, China 

Email: gaowei.@163.com 

Feng Wang and Li Wang 

College of Statistics and Applied Mathematics/Anhui University of Finance and Economics, Bengbu, China 

Abstract—Information technology has become an 

indispensable part to modern society. Information 

technology services as an independent industry has a 

profound effect on the progress of whole society and play on 

an importance role the development of the overall socioeconomic. 

This article makes an analysis on the 

developmental features, significance, present situation and 

existed problems of China’s information technology services 

and gives some relative suggestions as to how to develop 

China’s information services better. 

Index Terms—Information Technology, Information 

Technology Services, Present Situation Strategy 


Your goal is to simulate the usual appearance of papers 

in a Journal of the Academy Publisher. We are requesting 

that you follow these guidelines as closely as possible. 

Information technology services industry uses 

networks, computers and other modern scientific 

technology to produce, collect, process, store, transmit 

and use information and it is a specialized industry 

collection for providing service to society by information 

products. Various economic fields are linked close to 

information technology. The application of global 

information technology is developing deeply and widely. 

The development of information technology makes a very 

large impact on the technology-related industries and it 

has a promotion effect on improving productivity of the 

related industries. But, information technology is an 

independent industry separated from others, whose 

contribution has a bright future. As a sunrise industry 

along with its rapid development, there is a certain degree 

of shortcomings and deficiencies. Therefore, this article 

attempts to analyze the existed problems in the 

development of information technology services. So as to 

make it play better in the related-industries and to further 

promote economic development. 

II. ANALYSIS OF CHINA'S INFORMATION TECHNOLOGY 

SERVICES FEATURES 

(a) Information technology services industry is a 

means of modern services industry. Users use modern 


doi:10.4304/jcp.6.9.1805-1811 

means to obtain information, not only using such tools as 

book-style index and abstracts but also using CD-ROM 

or on-line search to obtain needed information. 

(b) Information technology services industry is 

engaged in industry of information technology. 

Information technology services, not to industrial 

enterprises, which is a department of the third industry. 

(c) Information technology services industry is the 

services aimed at specific objects, i.e. finding specific 

information for specific users, and finding specific user 

for specific information. 

(d) Information technology services industry is a high 

value-added industry, compared with the general services 

department, information services with high value-added. 

(e) Information technology services industry is a high 

penetration industry. Information technology and 

information services can penetrate and spread to all areas 

of society and industry sectors and has an active leading 

role to the department of other industry. 

(f) Information technology services industry is one 

which provides special commodities. Information 

technology services industry is characteristic of wide 

varieties, wide range, high value of transfer, easy to copy, 

timeliness and sharing. Furthermore, information 

technology commodity achieving its own value must rely 

on specific technology and service. 

(g) Information technology services industry is 

intelligence-intensive industry. IT services has feature of 

intelligence-intensive than other services department. 

Therefore, information technology services industry is the 

product of the knowledge economy era. 

(h) Information technology services, information 

circulation of commodities market has a special industry. 

The special nature of information goods, information 

goods exchange form and scope of the diversified 

characteristics. 

III. SIGNIFICANCE OF CHINA DEVELOPING OF IT 

SERVICES 

The development of IT services as an independent 

industry has far-reaching impact on development of 

national economy and social progress. It belongs to a new 

style of industry with high value-added, high technology, 

low consumption of energy resources, low environmental


pollution, high industry association leading role and good 

use of human resources. Its development has a promoting 

role to the other industry. 

A. The Promoting Role of the Development of IT Services 

to the Whole Society and the Development of National 

Economy 

(a) The Role of IT Services Development to National 

Education 

In the information age, information is the decisive 

resource. However, in our country, people’s 

understanding that information is a kind of resources is 

very vague. The waste of information resources is great 

and the efficiency of intensive utilization of information 

resources is very low. The development of modern IT 

services plays an actively educational role in improving 

and expanding the utilization efficiency of information 

resources, educating national people to cherish and use 

information resources and developing awareness of 

people’s consumption. Trying hard to develop databases 

and consulting services as modern IT services contributes 

to progressively faster the concept “compensation for the 

use of information” and reverses vague understanding of 

information resources. 

(b) The Development of IT Services Helping to 

Promote the Fusion of Industrialization and Information 

Now, one of the major tasks in our country is to 

achieve the integration of industrialization and 

information. IT services in upgrading and transformation 

of traditional industries forms a new road to 

industrialization by pouring new content into traditional 

industries and deepening the use information. In turn, 

new industrialization focusing on the information 

application also provides a broad market space and 

material and technical basis for the development of 

information services. Thus, IT services can drive the 

integration of industrialization and industry through the 

promotion information technology. In the cycle of 

interaction of Information technology and 

industrialization, the social and economic leaps and 

bounds will be achieved by promoting the formation of 

modern agriculture, services, and new industries. 

(c) The Development of IT Services Industry Itself 

Being Able to Bring Scale Expansion and the Change of 

the Mode of Economic Growth 

IT services belongs to an important part of information 

industry. IT services has become a highlight of the new 

century. The rapid development of IT services promotes 

the rapid development of information industry. Its 

contribution to the improvement of the information 

industry improves the proportion of the information 

industry in GDP. Information production, circulation and 

consumption scale keeps expanding and further 

stimulates people to create more demand for information. 

With the development of information technology, new 

growth points emerge in economic field such as IT 

industry, information services and Internet industries, and 

meanwhile, information can promote the growth of the 

total national economy through the optimization of 

production systems. Information technology also helps 

break the time limit and geographic restrictions of the 


market and speeds up the market of information 

processing and circulation. 

(d) The Development of IT Services Industry Helping 

the Increase of Economic Efficiency and Promoting the 

Whole Social Economic Development 

IT services industry is viewed as the industry of 

processing information whose information search and the 

improvement of transmission and switching efficiency 

has an important promoting role to the reduction of 

information asymmetry and lowering transaction costs. In 

addition, IT services industry has a significant multiplier 

and diffusion effect. On the one hand, IT services is 

beneficial to raising the utilization efficiency and 

information content of productive factors in other 

industries through the penetration of all economic fields, 

directly or indirectly producing a significant impact. On 

the other hand, with its own development and innovation, 

some new type will create of such as virtual tourism and 

distance education and so on, directly creating 

employment and economic output. 

B. The Development of IT Services Industry Having a 

Promotion Effect on the Other Industries Related 

(a) The Development of IT Services Industry 

Providing a Huge Market Demand for the Development 

of and Third and Other Industries 

In recent years, the rapidly growing demand of IT 

services not only provides the market demand of rapid 

growth for the and third and other Industries but also 

putting forward an urgent request for the development of 

IT services industry itself. IT services belongs to 

knowledge and technology-intensive industries whose 

rapid development is linked close to the manufacture of a 

large number of senior technical professional talents. It 

provides broad market prospects for the development of 

the related industries and educational career. Therefore 

the rapid development of IT services has a leading role to 

the gross economic output and scale of the third and 

related industries. 

(b) The Role of IT Services Industry Development 

Optimizing the Industrial Structure and Softening 

National Economy 

IT services industry can act as “catalyst role” in the 

development of national economy. The function of IT 

services optimizing the industrial structure is not only 

reflected in the development of IT services itself directly 

promoting optimization of industrial structure but also 

reflected in IT services as a typical general purpose 

technology indirectly promoting optimization of 

industrial structure. It leads to a series of production and 

change of the related industries, opens some new services 

industries and spawns a number of new “edge industries”. 

At the same time, information service industry has shown 

a softening effect on solving the employment. The 

professional tendency of modern IT services industry 

reduces some pressure, creates new jobs and makes many 

new careers stand out. 

(c) The Effect of IT Services Development on the 

Service of the Related Industries 

The function of information services is embodied in 

the reform of traditional industrial information and the


information technology support of the related industries. 

As for the traditional industrial departments, the 

development of IT services industry can raise developing 

efficiency of the new products, the technological content 

and value-added products. As for the construction 

industry, IT services like modern engineering consulting 

services should be developed vigorously, to soften the 

industrial industries and to achieve business value-added. 

In addition, to develop vigorously IT services industry 

can not only raise the added value level of the third 

industry but also update the traditional means of service 

industry. 

IV. TENDENCY ANALYSIS OF PRESENT IT SERVICES 

STATUS 

Although late beginning, with the development of IT 

and its wide use in various fields, China’s IT services 

industry now goes into the rapid growth stage, having a 

certain scale and achieving substantial results. The scale 

of software industry continues to grow. IT services, 

represented by communications industry, has been 

developing rapidly, the development and application of 

information resources having made positive progress, 

outsourcing service becoming a new highlight, some 

excellent digital content products having been created, 

information services promoting information technology 

having preliminary breakthrough and leading to upgrade 

of optimization. But, compared with the international 

advanced level, China’s information technology industry 

is still in lower level. Specific development status and 

future trends show as follows. 

A. Analysis of Information Technology Services 

(a) Unbalance of the Development Shown from Market 

Structure, Low Level of Market Development 

Unbalanced development of information technology 

services is mainly reflected in the unbalance of balance 

industrial structure and the regional development. In 

recent years, the information service level of information 

service industry has been obviously improved, but, the 

entire IT services market presents low-level trend. At the 

same time, the development of information services level 

is vigorously unbalanced. The information service 

industry level of cities like Shanghai, Shenzhen, 

Guangzhou is far above the China’s average level. As for 

the present, the communication industry in China 

accounts for over half of IT services industry, but, the 

software, technical services and information content 

services accounts for miner proportion. The development 

is unbalanced. In addition, due to the influence of many 

aspects such as its own conditions and external 

environment, the overall strength of IT services industry 

in China is low and small. The large groups with 

internationally competitiveness are too limited. 

(b) Developing Fast Scale, a Big Gap to the 

International Advanced Level, Belonging to Growth 

Stage 

Viewing from the revenue of the whole industry, the 

growth tendency of China’s IT services is great. Although 

China’s IT services industry, especially, modern IT 


services industry has made great progress, the proportion 

in the national economy and services industry is still low. 

Compared to the developed countries the overall level 

falls behind twenty to thirty years. In the global share of 

modern IT services market, IT services market scale of 

the United States, Western Europe and Japan altogether 

accounts for 90 percent of the global market, while 

China’s IT services market share accounts for only less 

than 10%. The gap is obvious. 

(c) Falling Behind Relatively in Information Resources 

Construction, the Information Infrastructure Not Being 

Sound Enough 

Not enough is the input of China’s database 

information systems infrastructure. Nor is the 

development strength of information resources. For 

example, there are many shortcomings in database 

construction: small quantity of database construction, 

small capacity, lack of database product with high 

quality, present database information resources focusing 

on hardware neglecting software, slow updating speed, 

contents focusing on science and technology and difficult 

to meet the market requirements. 

B. Analysis on Future Development Trend of IT Services 

Industry 

(a) The Traditional IT Hardware Manufacturers 

Transformation to Software and Services 

Many well-known enterprises of today’s IT 

manufacturing fields in the whole world see IT services 

as a key development. Some manufacturers like 

internationally renowned electronic data processing 

companies successfully realize transformation from the 

traditional hardware manufacturers to software and 

hardware manufacturers and service. In the Unite States, 

the largest feature of IT services industry is meanwhile 

providing the integrated IT services of hardware and 

software. The largest project in Japan’s IT services 

industry is software development services, which 

accounts for 60% of total sales. While, now in China, IT 

services industry should do better to prepare for strategic 

transformation, leading it to transformation of software 

and services in developing hardware manufacturing 

industry and realizing IT service industry’s convergence 

with international standards. 

(b) Adaping International Trends, Realizing 

Outsourcing of Software and Services 

The local industrial cluster formed and developed 

under the global background is the carrier of the regional 

and global economy. The IT development makes the 

industrial cluster theory take on a revolutionary change. 

The border region and national boundary and therefore, 

conduct development, production and sales activities in 

different areas. Software outsourcing, compared with the 

previous one, in terms of scale or in the related content, 

has presented unprecedented differentiation and 

development. Shown as follows, first, scale enlarging and 

energy degree increasing. To ensure the quality and 

speed, the contractor establishes more and more large 

development centers where the party contracting locates. 

Second, diversification of outsourcing situation, there are 

two types: direct and indirect sub-contracting. The United


States usually adopts direct contracting. Japanese 

software enterprises adopt indirect subcontracting. In the 

initial stage of IT services development, China has 

become contractors of developed countries. 

(c) Highlighting IT Services Development of key 

Areas, Emphasizing Its Services to Producers 

All countries emphasize highlighting IT services 

development of key areas and put forward specific plans. 

The United States puts the emphasis of IT services 

industry development on software industry. Japan puts 

the emphasis of IT services industry development on 

software industry on e-government and e-commerce. 

While China should combine its own characteristics and 

simultaneously develop IT services industry by drawing 

on foreign development mode. When the manufacturing 

industry developing to a certain stage, it is bound to 

transfer to the productive services industry of high profit 

and high value-added such as research and development; 

consulting and design. It is bound to generate the demand 

for the productive IT services. At the same time, IT 

services industry gradually takes on the developing trends 

of the “outside” and “professional”. Now it becomes 

necessary requirements of China’s economic 

development to a certain period how to adapt to the new 

situation of economic development and to accelerate IT 

services industry. 

(d) Realizing the Integration of Consulting and IT 

Services Industry, Providing Professional Value-Added 

IT Services 

The integration of consulting and IT services industry 

not only promotes revolutionary breakthrough of 

consulting industry but also the development of 

information services industry. Information value-added 

services is a service mode that uses market means to 

increase the asset value of the information service by 

classifying, processing, arranging and analyzing a large 

number of original information, aiming at different 

customer’s demands and features. Information valueadded 

service is one of the developing trends of 

information services industry. Today, in such a colorful 

information contents, only to make good use of 

information resources to improve the technology of using 

information, to satisfy customer’s various demands and 

keep competitive forces. Moreover, the added value 

services of production largely reflects the professional 

level of service, so, the needs for the professional 

information service is IT services industry developing to 

a certain stage. 

V. PROBLEMS EXISTED IN CHINA’S IT SERVICES 

INDUSTRY DEVELOPMENT 

Because of late start, there is a big gap compared to the 

international advanced level. IT services industry is 

developing very rapidly in recent years, but, whether 

from the external macro environment of IT services 

industry or from the internal factors of micro aspects, 

there are some problems and shortcoming, affecting IT 

services industry to develop better. 


A. Macro Aspects 

(a) Imperfect Legal System of IT Services Industry 

At present, in the rapid development of IT services 

industry, there is no corresponding laws and regulations 

of IT services industry. The serious delay of policy and 

regulation leads to the irregularities of information 

market operation. With the development of the times and 

IT, China’s laws and regulations of promoting consulting 

industry development made in the eighties and ninties of 

last century were not suited to the demands of the Internet 

times and must accelerate replacement pace. The 

legislation of IT services industry in China has long been 

lacking a commanding basic law with a higher status. 

China’s legal system of existing information services 

technology is still in its infancy, lacking comprehensive 

legislation system and the clear legislative goals. The 

existing legislation is largely for the sector and local one 

with imcomplete system, low grade 

(b) Imperfect Management and Organization 

Coordination Mechanisms of IT Service Industry 

China’s information market, whether business income 

or employment maintains a fairly large growth every 

year, but lacks effective and integrated management. The 

structural integrity of information market system and the 

integrated development policy and planning depends on 

sound management system and organizational 

coordination mechanisms. As for our country, there aren’t 

any comprehensive, integrated and centralized leading 

departments. Such a new industrial service institution as 

IT services industry scatters various administrative 

departments of various fields and brings great difficulties 

to the organizational management. Meanwhile, the 

chaotic situation of higher-level managing departments 

and higher authorities fragmented many leaders and 

management disorder reduces the government’s control 

and operational efficiency. It is not conducive to the rapid 

development of the entire IT services industry. 

(c) Imperfect Market Operating Mechanism of IT 

Services Industry 

Information market operation mechanism based on IT 

services industry mainly consists of the price 

mechanism, competition mechanism and supply and 

demand mechanism. Now, information market of IT 

services industry in China is still at an early stage of 

development. There exists a great defect in the operation 

mechanism of information market. First, it is difficult to 

determine reasonable price of information products and 

there exists a large subjective arbitrariness in the price 

determination, causing some disorder to the information 

product price of IT services industry. So, it is difficult to 

form reasonable price mechanism. Second, duo to the 

asymmetry of social information, resulting in unbalanced 

contradiction of information product supply and demand 

in IT services industry, there often appears the 

phenomena of some information products in short supply 

or oversupply. The contrast between supply and demand 

information is fairly larger. The supply and demand 

mechanism of IT services industry is still not mature. 

Third, the Information market of IT services industry 

lacks effective competitive mechanism. At present, IT


competitive conditions between the services sectors are 

not mature. There exist many unequal competitive 

phenomena, lacking standardized competitive behavior in 

information services departments. Furthermore, IT 

services agencies themselves do not establish a sound 

internal operation mechanism. 

B. Micro Aspects 

(a) Lacking the Advanced Technology Management 

Personnel of Modern IT Services 

The rapid development of IT services industry depends 

on high-quality technical personnel and management 

personnel. This is a knowledge-intensive industry. 

Although there are a large number of qualified personnel, 

some of which is very excellent, generally speaking, the 

quality is universally low. The reason for this 

phenomenon is duo to the rapid expansion of China’s IT 

services market and the market is still in its infancy with 

a serious shortage of supply of professionals, (especially 

lacking the senior personnel of market management and 

international management.) Moreover, the staff currently 

has no unified system of accreditation and assessment 

with the uneven quality. There universally exists the 

phenomenon that expertise is obvious, but they lack the 

awareness and experience of the modern of consultation. 

At the same time, many excellent personnel of 

information services flows to developed countries and 

domestic foreign-funded enterprises, especially high-level 

brain drain is serious. 

(b) Information Products and Service Standards Are 

Not Unified, Producer Services and Information Value- 

Added Services Expecting Further Development 

It is difficult to form a united definition to the standard 

that information services software products provided by 

IT services industry to intangible products, producing the 

result that there are no uniform rules and standards in the 

products and services provided by IT service in our 

country. But, with the development of IT, the productive 

services raise the requirements of standardization and 

specialization, the degree of outside gradually increasing. 

At the same time, with the “services” of manufacturing 

industry and the adjustment of profit nodes in industrial 

chain, the world’s manufacturing industry is conducting a 

big strategic shift to China, causing the growing demand 

for the information services of production and value 

added services. 

(c) Information Language Not Compatible with 

Technology 

That information technology is not compatible with the 

language makes all the system difficult to communicate 

and link, resulting in language barriers between systems. 

Duplication of existing databases and separation from the 

market demand is mainly due to the reason that many 

information industry sectors haven’t implemented the 

national standards and international standards, causing 

irregular organization of information resources and 

seriously hindering the progress of information industry 

and network. 

(d) Poor Information Resources and Poor Service 

Chinese information in China’s online is too small and 

the information resources are poor so that it is difficult to 


meet the needs of many customers. Many information 

technology services agencies have financial problem, 

most of which only conduct a simple collection and 

accumulation and don’t deeply analyze or evaluate the 

information. So, the short-term behavior of research leads 

to the poor information services. 

(e) Insufficient Capacity of Independent Innovation 

At present, the innovation capacity China’s 

information service enterprise has been considerably 

improved, but there is still a big gap compared to 

developed countries, which is mainly reflected in the 

inadequate innovation capacity such as core technology, 

business development, products development technology 

and service model. Because of the basic institutions and 

core technology lying in the hands of international 

companies, controlled by others in core technology; due 

to inadequate protection of business ideas in the 

innovation of the service model, the copying 

phenomenon of information services being universally 

serious; difficult to grasp customer needs in business 

innovation, making few original innovation of 

information service products, the most being imitation 

and lead innovation; the homogenization phenomenon of 

business and application being serious in products 

development. 

VI. RELATED COUNTERMEASURES ON CHINA’S 

DEVELOPMENT OF MODERN IT SERVICES 

INDUSTRY 

A. Macro Aspects 

(a) Optimize the Policy Environment of Information 

Services Development, Further Improving the Legal 

System 

Some key areas of China’s IT services are still in the 

early stages of development. The information industry 

must rely on the state’s information policy and the 

corresponding legal protection and need the strong 

support of government and the society. Government 

departments should improve the legislative and legal 

system of IT services. The first is to make and improve 

the legislative and legal system closely related to IT 

services industry such as “Information Services 

Management Ordinance”, “Information Law”, “database 

revitalization Law”, “Government Information Resources 

Management Regulations”, and software standards, 

information standards, network technology standards and 

those related to IT standards and regulations of IT 

services market. The second is to make the public law of 

government information. Because of no reliable law, 

there are the contradictory relationships between 

information resource sharing and safety security, the 

public law of government development and use of no 

confidential information resources. The last is to build 

and improve the related market-oriented industrial policy, 

optimizing the external environment of IT services 

development such as financial, fiscal and tax policy. 

(b) To Further Establish and Improve a Strong 

Administration of Information Technology Services


IT services development goes deep into all fields of 

society, to change the present situation of no unified 

industrial management developments against managing 

disorder and many leaders, it is necessary to establish and 

perfect a strong national administration of information 

services so as to implement effective unified management 

to the IT services industry and coordinate the business 

between the competent departments, to formulate 

development strategy of IT services industry. At the same 

time, it is necessary for the industrial institutions to give 

play to the function of standard and self-discipline on IT 

services industry, learning form the U.S. management 

model of IT services industry and gradually forming a 

line with the management system of socialist market 

economic system. 

(c) Perfecting Operational Mechanism of Information 

Market 

Information, as a social resource, depends on the 

information market for allocate. The market of 

information transaction and the commercialization of 

information is very important way to promote the 

development of IT services industry. Only by establishing 

a unified, open, competitive and orderly information 

market system, constantly improving the operational 

mechanism of the information services market, can the 

sound development of information technology services 

industry be promoted. The main areas are as follows: 

first, using the market-oriented mode of operation to 

perfect supply and demand mechanism, putting the 

production and sales process of IT services industry into 

the orbit of market operation; second, by the 

improvement of the price mechanism, playing its full role 

in the adjustment of the market, balancing supply and 

demand relationship of information products in the 

operation of information services markets; Finally, by 

improving the competitive mechanism, having each 

sector of information services form survival of the fittest 

and a mutual competition in the information services 

market, forming a prosperous market of IT services. 

B. Micro Aspects 

(a) Raising Professional Service Level, Training and 

Introducing Professional Personnel on Information 

Services 

IT services industry belongs to knowledge-intensive 

industries and the professional service level it provides 

depends on the practitioners’ knowledge reserve and the 

professional level. At present, our country lacks the 

professional personnel of IT services, particularly in IT 

services industry and other industries of mutual 

penetration. The market needs a large number of 

professional personnel and puts forward higher 

requirements to the personnel. Therefore, it is extremely 

important to introduce and train the qualified professional 

of IT services by variety of measures improve the 

gathering space of high-end qualified personnel of 

information services, to raise the professional level and to 

promote the development of information technology 

services. First, to pay attention to “talent cultivation” 

strategically, to implement national strategy planning of 

the qualified personnel of IT services industry and strive 


to create a large number of professional personnel of high 

skill and high-level, familiar with modern IT services 

industry of international regulation and management. 

Second, full implementation vocational qualification 

certificate system, strengthening vocational job training, 

improving the basic quality of information services 

practitioners. Third, introducing domestic and 

international qualified personnel by many forms and 

channels and retaining them by establishing effective 

incentives and compensation mechanisms. Fourth, 

training high-intermediate information services personnel 

strengthening subject construction related to information 

services in colleges and universities so as to adapt to the 

needs of the rapid development of modern information 

service. 

(b) Full Developing and Using Information Resources 

Speeding Up Construction of Various Types of Databases 

Information service providers should focus on the 

development and utilization of information resources. 

Therefore, to fully develop and use information 

resources, it is necessary to build basically complete 

market demand-oriented databases and databases with 

local features, to mainly develop all kinds of the public 

and commercial databases serving the society, to adopt 

the principle of unified management, building databases 

in different places, resource sharing and multi-service, 

gradually forming a large practical database network. 

(c) Forming New Economic Growth Points, Grasping 

Innovation Project 

Innovation is the soul of information services 

development and the theme of a new era. In today’s 

network environment such as information visualization, 

digital, commercialization and globalization of 

information, it is imperative to use modern information 

technology to establish IT service innovation projects 

with high-speed information network as its main body. 

Therefore, innovation information management of 

system must be carried out well, information 

management, information systems, information services 

and information application technology to further 

promote better development of IT services industry. 

REFERENCES 

[1] ZHAN Jing. Innovation and Development of 21st 

Information Service Industries in Our County. Modern 

Library and Information Technology, 2002. 

[2] LI Jian-Ge. Developing Strategies of Modern Information 

Service Industries. China Information Fields, 2007. 

[3] WANG You-Gang. Research on Development Mode of 

Information Service Industries. Industry Analysis, 2005. 

[4] GUO Dong-Qiang. Current Situation and Solution of 

Development Information Service Industries in Our 

County. Market Weekly, 2008. 

[5] CAO Kuan-Zeng. Research on Development strategy of 

information Service Industries in Our County. Practice 

Research, 2003. 

[6] LI Jian-Ge. Development strategy of Modern Information 

Service Industries. China Information Industry, 2007. 

[7] KUANG Pei-Yuan. Information Services: Definition and 

Statistical framework. Statistics Education, 2009.


[8] WANG Xin. Development Mechanism and Measurement 

Theory of Information Industries and Methods. Jilin 

University PhD thesis. 2008. 

[9] YANG Xiang-Ming. Some Thoughts on Development of 

Information Service Industries in China. Library Theory 

and Practice, 2007. 

[10] HOU Fu-Li. Research on Current Situation and Solution of 

Modern Information Service Industries Development. 

Group Economic Research, 2007. 


Wei Gao (1964- ). Female, Han Dynasty, Nantong Jiang Su, 

College of Statistics and Applied Mathematics, Anhui 

University of Finance and Economics Vice Professor. Research 

Field: Quantity Economics. 

E-mail: gaowei.64@163.com, Mobile Phone: 15805525532. 

Feng Wang (1962- ). Male, Han Dynasty. Nantong Jiang Su, 

Anhui University of Finance and Economics. Vice Professor. 

Research Field: National Economiy 

Li Wang (1984- ) Male, Handan, Bank of Communication 

Hebei. Research Field: Quantity Economics.


A Web Survey Program Based on Computer 

Technology and Its Application to Evaluation 

Model about Youth Self-organizations in China 

Ma-lin Song 

School of Statistics and Applied Mathematics, Anhui University of Finance and Economics, Anhui Bengbu, China 

Email: songmartin@163.com 

Tong Yang and Ya-qing Song 

Anhui University of Finance and Economics, Anhui Bengbu, China 

Abstract—The network has become the second space for 

people in China, and network and youth self-organizations 

based on web-platform have influenced young people more 

than ever before. From the viewpoint of the overall 

development of youth and building a harmonious society, 

it’s an important thing to reduce the negative influence of 

the network and strengthen the sustainable development of 

media ecology. The paper forecasts the developmental trend 

of adolescents by analyzing their current situation in China 

and builds the evolution model for youth self-organizations. 

This web survey program uses the IIS web server + 

ASP.NET service + SQL Server database. Survey.aspx 

could be generated in the server dynamically, so the web 

survey program can be achieved by computer. Finally, the 

paper suggests some advices to eliminate the negative effects 

of internet and to strengthen youth self-organizations. 

Index Terms—youth self-organizations; internet media; grey 

forecasting model; analytic hierarchy process; Web Survey 

program 


Three decades after reform and opening up, China has 

undergone enormous changes. Amateur live of Chinese 

people, young people in particular, has increased more 

rich and varied. In recent years, with progress in science 

and technology, communications, and popularization of 

the Internet, the network has become the second largest 

human space, by which the impact of it on youth is 

growing. Although network have brought great 

convenience for our times, the deterioration of its 

environment, such as network information pollution, 

network security crisis, private space crisis of network, 

the shortage and expansion of networks information, also 

seriously endangers the physical and mental health of 

youth. Therefore, from the perspective of the overall 

development of young people themselves, or for building 

a harmonious society, to consider how to reduce and 

eliminate the negative impact of network on youth and 

strengthen ecological civilization construction of 

network, has increasingly become an important issue to 

be settled urgently. 


doi:10.4304/jcp.6.9.1812-1818 

II. LITERATURE REVIEW 

An Guoqi, Deng xiquan and Cao Kai (2006) pointed 

out that the Government has to face up to 

non-governmental organizations and the role of youth, 

and official organizations ought to take positive measures 

to guide and monitor the non-governmental organizations 

and effective role of youth [1]. Ma Chunlei (2007) 

considered that self-organizing system is still beyond our 

traditional work, by which the formation of its social 

forces deserves our attention and research particularly 

[2]. Shi Guoliang (2007) thought that youth 

organizations, especially the informal youth 

organizations, are increasingly becoming a social 

organization that have rapid development, strong vitality, 

increasing cohesiveness, and influence [3]. Xu Rong and 

Zheng Chen (2007) suggested an educational 

management method that the active roles of informal 

organizations in students ought to be played and their 

negative effects should be controlled [4]. 

Zeng hong considered we should concern with the 

composition and behavior characteristics of Internet users 

primarily for how to design web survey program. 

According to the Chinese Internet Network Development 

survey data in Chinese Internet Network Information 

Center (CNN IC), he made a quantitative analysis in the 

composition and behavior characteristics of Chinese 

Internet users, and then discusses the network survey 

design effects [5]. 

The paper achieves the web survey program through 

the IIS web server + ASP.NET service+ SQL Server 

database. Study on China’s youth self-organization based 

on media ecology perspective, this article suggests that 

ecological construction of young self-organization need 

to be strengthened and ecological environment of China’s 

internet media should be optimized to promote diversity, 

rationalization and ecology distribution to strengthen the 

full development of youth and harmonious society 

building.


III. ANALYSIS OF INTERNET YOUTH USERS IN CHINA 

A. The status quo of internet Youth users in China 

With the advent of the information age, improvement 

of communication facilities and increasing of people's 

income level, the Internet is getting into millions of 

households. 

The scale of Chinese internet user has showed the 

trend of sustained and rapid development, In June 2008, 

the number of Chinese Internet users is 4.52 times than it 

in June 2002. In June 2008, the number reached 25.3 

million, ranked first in the world. In June 2008, the 

number of Internet users under 24 years old is 4.15 times 

than it in June 2002. An increase from 18 to 24 year old 

Chinese Internet users is 3.50 times than it in June 2002. 

B. Trends forecast Youth Internet users 

In this paper, GM (1, 1) model is used to predict the 

size of China's young Internet users and Internet users. 

The gray system theory is proposed by Professor Deng 

Julong, a China scholar, in the 1980's, which is used to 

control and prediction and is widely applied in 

agriculture, socio-economic and other fields [6]. In this 

paper, GM (1,1) model is used to forecast China's total 

Internet users and its change in the trend. The simulation 

model and the residual difference are shown in table I. As 

a result of p = 1.0000, c = 0.1942, the current model is in 

a very good level of prediction. 

TABLE 1 CHANGES OF THE TRENDS IN THE TOTAL NUMBER OF 

CHINESE NETIZENS 

Sequences 

Original 

value 

(0) 

x () i 

Predictive 

value 

(0) 

xˆ () i 

Residual 

errors 

(0) 

ε () i 

Relative 

errors 

(%) 

X(2) 5910.0000 5127.3129 782.6871 13.2434 

X(3) 6800.0000 5868.6508 931.3492 13.6963 

X(4) 7950.0000 6717.1758 1232.8242 15.5072 

X(5) 8700.0000 7688.3857 1011.6143 11.6278 

X(6) 9400.0000 8800.0191 599.9809 6.3828 

X(7) 10300.0000 10072.3790 227.6210 2.2099 

X(8) 11100.0000 11528.7046 -428.7046 -3.8622 

X(9) 12300.0000 13195.5944 -895.5944 -7.2813 

X(10) 13700.0000 15103.4933 -1403.4933 -10.2445 

X(11) 16200.0000 17287.2477 -1087.2477 -6.7114 

X(12) 21000.0000 19786.7426 1213.2574 5.7774 

X(13) 25300.0000 22647.6296 2652.3704 10.4837 

The predicted results of other variables (18 ~ 24 years 

of age the number of users) is available similarly, of 

which the proportion of 18 to 24 years old of Internet 

users is get through each stage netizens divides total 

number. 

As can be seen through the forecast, the next three 

years the total number of Chinese Internet users and the 

number of Internet users of 18 - 24 years old will 


continue to increase. In June 2011, the number reached 

13.893 million, the Chinese youth will account for about 

27.24 percent of China's total Internet users. 

It is generally believed that China's rapid development 

of Internet network bring about opportunities for the 

youth self-organizations’ flourish. The Internet goes into 

millions of households, in which its fashion and 

convenience attract a lot of young people involved. 

Internet provides equality, freedom, easy platform 

exchange for young people’s activities, by which it 

brought more opportunities for the formation of 

self-organizations. Low-cost of Internet resources’ 

network configuration also carries out facilitations for the 

formation of self-organizations’ establishment, 

management and activities. 

IV. INVESTIGATION DESIGNATION OF YOUTH 

SELF-ORGANIZATIONS BASED ON INTERNET MEDIA 

PROSPECTIVE 

A. Index system for youth self-organizations evaluation 

and its quantification 

In this paper, four-level index system is used to 

evaluate youth self-organizations, in which the target 

level is youth self-organizations indicators index, criteria 

level includes eight indicators used to measure the 

members’ feeling of youth self-organizations, indicator 

level includes a total of 28 indicators and the last level 

mainly includes questionnaire design for indicator level. 

B. Determination Indicators’ Weights 

In this paper, evaluation system of indicators of youth 

self-organization is composed of the multi-level index 

cluster. It constructs judgment matrix structure after 

seeking the advices from experts and determines weigh 

by mathematical treatment in some forms. Therefore, this 

article will make it more scientific to combine qualitative 

and quantitative weigh determination by Analytic 

Hierarchy Process (AHP) [7]. 

Analysis Hierarchy Structure are Constructed with 

Index System for the Calculation, which includes object 

layer A; rule hierarchy B1 ~ B8; individual indicators are 

just index hierarchy. After using “1 to 9 scales”, 

judgment matrix of the criteria to the objective is 

constructed. It is important to carry out Consistency test 

of judgment matrix and level-ranking, which can be seen 

in table II. 

TABLE 2 CONSISTENCY TEST OF JUDGMENT MATRIXES 

CI CR 

A 0.0721 0.0511 

B1 0.0018 0.0032 

B2 0 0 

B3 0.0652 0.0693 

B4 0.0853 0.0761 

B5 0.0193 0.0332 

B6 0.0198 0.0220 

B7 0.0193 0.0332 

B8 0.0046 0.0079


Under such circumstances, the judgment matrixes of 

the CR are less than 0.10, which can be considered sort of 

single-level structure with consistency. As results, 

Hierarchy general ranking results are as follows: 

W ′ B1 

=[ 0.1427 0.0269 0.0506]′ 

(1) 

W ′ B2 

= [ 0.0075 0.0075 0.0226]′ 

(2) 

W ′ B3 

= [ 0.0041 0.0095 0.0160 0.0294]′ 

(3) 

W ′ B4 

= [ 0.0480 0.0032 0.0060 0.0151 0.0205]′ 

(4) 

W ′ B5 

=[ 0.0158 0.0390 0.0962]′ 

(5) 

W ′ B6 

= [ 0.1904 0.0692 0.0316 0.1073]′ 

(6) 

W ′ B7 

= [ 0.0107 0.0018 0.0043]′ 

(7) 

W ′ B8 

=[ 0.0134 0.0041 0.0074]′ 

(8) 

These matrixes from (1) to (8) are the corresponding 

weights of single indicators. 

V. EMPIRICAL ANALYSIS OF PROGRAM DESIGN 

A series of investigations are carried out in a university 

surrounding schools via the design of the questionnaire of 

the authors, by which some youth self-organizations are 

known, 15 self-organizations being more influential and 

Internet-based, can be selected to be conducted a 

questionnaire survey on. The specific names of 

self-organization are as follows: Economic Research 

Institute (Y01), Employment and Entrepreneurial 

Associations of University Students (Y02), Mutual 

Assistance Center of college students (Y03), Students 

Association of Financial Investment (Y04), Association 

of popular science of Students (Y05), Computer 

Association (Y06), Basketball Association (Y07), Table 

Tennis Union (Y08), Management Institute (Y09), Art 

Troupe of university students (Y10), Advertising Art 

Association (Y11), English Society (Y12), Green IN 

Society (Y13), Association of Wushu Enthusiasts (Y14) 

and Press Corps of university students (Y15). According 

to tests in among small proportion of their numbers, the 

revised final version, including 28 issues, is concluded. 

Because of limited space, the programs do not list; and it 

can be obtained from the author if necessary. 

Because of the difficulty in implement stabile retest 

reliability, most questionnaires use consistency reliability 

testing generally, in which reliability coefficient α is the 

most commonly used method. Cronbach α, being 

reliability coefficient, can be used for test of consistency. 

In general, α may be accepted if it is larger than 0.5. If the 

reliability coefficient is greater than 0.7, it means a very 

high reliability; when the range between 0.7 and 0.35, it 

means so-so; if it is less than 0.35, it means low 

reliability. Web survey can be carried out through emails, 

by which emails with questionnaire send. 


VI. ACHIEVEMENT OF INTERNET SURVEY PROGRAM 

A. Introduction 

This web survey program uses the IIS web server + 

ASP.NET service+ SQL Server database. Survey.aspx 

could be generated in the server dynamically. 

The entire program is divided into three layers: client 

layer, service layer and data layer. Client base in the 

surveyed users computers. The user could request 

survey.aspx page by IE browser. Service layer in the IIS 

web server, survey.aspx is generated in the IIS server 

dynamically and passed to the customer's IE browser. 

Data layer in the SQL Server database; all the issues 

involved in the web survey, survey and the user 

participated in the investigation are stored in the database. 

The hierarchical structure as shown in Figure 1: 

IIS 网页服务器 SQL 

Server 

数据库 

Figure 1 hierarchical structure chart 

B. Design 

For this program, the design includes class design in 

server and database design in data table. 

1. class design 

Mainly survey.aspx web pages generated dynamically. 

Survey.aspx initially only contains the user table 

information and the submit button, the specific details are 

shown in survey.aspx file. For different request survey 

class handle survey web pages and the result web pages 

dynamically generated. 

Survey class view is as follows: 

Figure 2 Survey class view 

In the class, Page_Load () function is executed for 

each user apply for survey.aspx page, according to 

different application parameters sub-function generated 

different pages dynamically. In order to build the survey 

and result web page in the web survey, we design two 

different classes: Submit Page Creator and Result Page 

Creator. Submit Page Creator class is used to generate the


survey website; Result Page Creator class is used to 

generate the result page. Their class diagram as follows: 

Figure 3 Submit Page Creator and Result Page Creator class chart 

In Survey class, Submit_Click () function is executed 

when the user clicks on the submit button each time. This 

function collects the information in survey and user 

tables, to keep the information to the database. 

2. The design of the data table structure 

The network survey program involves 7 data table: 

ref_Organization, ref_QuestionType, ref_Question, 

ref_SurveyCatalog, Survey, Survey_Detail and User 

table. 

1) Ref_Organization table 

The table saves the information of survey; the table 

structure is defined as: 

Figure 4 Ref_Organization table structure chart 

ID is used to identify each organization; Name is the 

organization's name; Description gives a brief description 

of the organization. 

Ref_Question Type table 

The table holds the type of survey questions. The table 


Figure 5 Ref_Question Type table structure chart 

Type is used to identify each problem type; Choices is 

all the alternative answers for the kind of problem 

(separated different answers by semicolon); Weights is 

scores for the corresponding optional answer (separated 


different scores by semicolon); Description is a brief 

description of such problems. 

Ref_Question table 

The table holds all the survey questions. The table 


ID is used to represent each problem; Name gives the 

content of issue; Type specifies the type of problem; 

Description gives a brief description of the problem. 

Ref_Survey Catalog table 

The table holds all the web survey by the system 

launched. The table structure is defined as: 

CatalogID is used to distinguish different surveys; 

Questions gives all the problems in the survey (separated 

different issues by semicolon); Organizations gives all the 

surveyed organizations involved in the survey (separated 

different organizations by semicolon); Valid Survey 

Count represents the total number of all valid 

questionnaire in the survey; Total Survey Count 

represents the total number of all submitted questionnaire 

in the survey; Descriptions gives a brief description for 

this questionnaire. 

2) Survey table 

Figure 6 Ref_Question table structure chart 

Figure 7 Ref_Survey Catalog table structure chart 

The table participates in the questionnaire submitted by 

the user each time through save the system. The table 


Figure 8 Survey table structure chart


ID is used to identify this survey questionnaire; 

CatalogID represents the questionnaire belongs to which 

network survey; UserID represents the user ID who 

submitted the questionnaire; Time represents the 

submitted time of questionnaire; Valid indicates the 

validity of this questionnaire; SourceIP represents the IP 

address submitted to the client in the questionnaire. 

3) Survey_Detail table 

The table holds each question and participates in the 

user's choice in the entire questionnaire. The table 


SurveID identifies each specific survey item in the 

questionnaire. QuestionID gives the question identity 

involved in this investigation; OrganizationID gives the 

organization identity involved in this investigation; 

Record shows the results of the survey items. 

4) User table 

Figure 9 Survey_Detail table structure chart 

The table holds the detail of the involved user. The 

table structure is defined as: 

ID is used to identify each user; Gender gives the user 

gender; AgeRange gives the age range of users; Domain 

gives the industry of user; Name gives the user's name; 

Email gives the e-mail of users; Address gives the contact 

of user; Comments gives additional user information. 

C. Demonstrate 

User need to provide id parameter to apply for 

survey.aspx, this parameter is used to distinguish 

different web survey. The IE browser displays as follows: 


Figure 10 User table structure chart 

Figure 11 Web survey chart 

In the table, the users make a choice for the overall 

impression of 15 self-organizations, options can be 

divided into very satisfied, satisfied, more satisfied, in 

general, less satisfied, dissatisfied, very dissatisfied. 

However, if the users have not participated in the 

self-organization, some of the problems they are not 

interested or do not know the answer, please do not 

answer. 

Finally, the user provides some necessary personal 

information. 

Figure 12 Personal information chart 

In the table, the users need to select the relevant 

information, including gender, age range, and 

professional. If the users need the results of this survey, 

please provide name, address, zip code, and Email. 

After the user clicks the submit button, according to 

different user information, it will show different findings 

slightly. In the result page of web survey, for each 

question, page provides to the total users number of 

answering the question and the current result. The result 

represents by the color section, the shorter color section, 

the closer to green, and the result is more close to the left 

of alternative answers in the list.


When the user to provide personal contact, the result 

page as shown below: 

In the table, we can see the web survey results for the 

overall impression of 15 self-organizations; the second 

column shows the number of result options. 

After the user provides personal contact, the results as 

shown below: 

The table shows that the latest survey results will be 

sent to the user’s e-mail or postal address. 

VII. CONCLUSIONS 

Different types of self-organizations have different 

impacts on the growth of young people, while young 

people also have their own objective assessment of 

different self-organizations. The diversification, 

co-existence and symbiosis of self-organizations will 

enrich the lives of young people and promote the 

comprehensive development of youth. At present, 

network environment should be optimized to provide 

good platforms for young and healthy development of 

self-organization. 


Figure 13 Web survey results chart 

Figure 14 Web survey results chart 

The development of Networks is breaking down the 

temporal and spatial boundaries of ideological and 

political education and provides new opportunities for 

further strengthening of educational influence. 

Through the building of network platform, new moral 

space could be opened up. Strengthening the building of 

communication channels can help students explore a 

variety of ideological confusion or communication issues 

freely with their teachers, parents and students and keep 

abreast of all kinds of information in society, by which 

expectations of communities, school and parents are 

conducted together through the network society. 

These will cherish the transfer for students to increase 

the original space of narrow education into whole society 

and develop ideological and political education, which 

make the original lag content of ideological and political 

education into a more forward-looking for students. The 

timeliness of content makes room for the ideological 

and political education that be extended to the entire 

network. Publicity through the network, young people 

may know that indulging in online games is dangerous. 

Understanding, proper use of the Internet and ability to 

selection of useful information will enhance their own 

ability to resist information pollution. 

It is believed that through the integration of network 

information and practical resources, combining with the 

establishment of a good cultural atmosphere of the 

network, various types of self-organization will 

strengthen exchanges and cooperation among them. 

Meanwhile, vigorous, healthy and civilized organization 

activities could enrich the lives of amateurs of youth. 

ACKNOWLEDGMENT 

The authors wish to thank Yang Jie, from School of 

Adult Education, Anhui University of Finance and 

Economics, Anhui Bengbu, China, for his help of the 

finish of this paper. This paper is supported by Supported 

by National Natural Science Funds of China for 

Innovative Research Groups (70821001), National 

Natural Science Foundation of China (70901069), Social 

Science Foundation of Ministry of Education of China 

(10YJC630208), Social Science Foundation of Anhui, 

China (AHSK07-08D25, AHSKF09-10D116, 

AHSK09-10D14), and Anhui Provincial Natural Science 

Research Project for Universities (KJ2011A001). 

REFERENCES 

[1] An Guoqi, Deng xiquan and Cao Kai. Research on the 

roles and development trends of contemporary youth in 

non-governmental organizations [J]. Youth Studies, 

2006(5): 3-5. (In Chinese) 

[2] Ma Chunlei. Status quo of youth non-governmental 

organizations and theirs guides [J]. China Youth Study, 

2007(11): 38-39. (In Chinese) 

[3] Shi Guoliang. Analysis of development trends of youth 

organizations in today's world [J]. China Youth Study, 

2007(12): 22-24. (In Chinese) 

[4] Xu Rong & Zheng Chen. College students' informal 

organizations and educational management 

countermeasures [J]. Journal of Ningbo Radio & TV 

University, 2007(2): 91-93. (In Chinese)


[5] Zeng hong. See the program design of web survey from 

Chinese internet users features [J].Economic 

Issues,2006(1): 145-147.(In Chinese) 

[6] Deng Julong. Basic methods of gray system [M].Wuhan: 

Publishing House of Huazhong University of Science and 

Technology, 2006. (In Chinese) 

[7] Xiong li, liang Liang and Wang Guo-hua. Method research 

on selection and valuation of numeric scale in analytic 

hierarchy process [J]. Systems Engineering-theory & 

Practice, 2005(3): 72-79. (In Chinese) 


Malin Song, corresponding author, is a teacher in School of 

Statistics and Applied Mathematics, Anhui University of 

Finance and Economics, Bengbu, Anhui, China. His major field 

of study includes management of computer manufacturing 

enterprise, credit risk, strategic alliance and eco-industrial park 

(E-mail: songmartin@163.com). He is the corresponding author 

of this article.


The Research on the Influencing Factors of 

Financing Strategy of Woman Entrepreneurs in 

China 

Xiong Xiong 

College of Management and Economics, Tianjin University ,Tianjin , China 

Email: xxpeter@tju.edu.cn 

Rong Fu, Wei Zhang, Yongjie Zhang 

College of Management and Economics, Tianjin University,Tianjin, China 

Email: diana1228cn@yahoo.com.cn, weiz@tju.edu.cn, yjz@tju.edu.cn 

Abstract—Based on the data from the nationwide surveys of 

SMEs in "China's Private Economic Research in 2002, this 

paper examines gender differences among Chinese 

entrepreneurs seeking financing pattern, including external 

and internal financing, and studies on the factors those 

affect women entrepreneurs’ financing strategies through 

theoretical analysis and model validation from the human 

capital and social capital perspective. We find that human 

capital and social capital have positive influence on seeking 

external financing. There is also some evidence that the 

impact in Administrative system may promote external 

financing in China. 

Index Terms—Women entrepreneurship, Human capital, 

Social capital, financing strategy 


With the development of economy, increasing women 

begin to start their own business, no matter in developed 

or developing countries. Although female 

entrepreneurship has drawn wide attention only in recent 

20 years, the female entrepreneurship developed rapidly, 

and women enterprise has become an important driving 

force for the global economic growth. According to the 

GEM report (2005), female entrepreneurship is booming 

worldwide, and more than a third of entrepreneurs are 

woman. According to GEM 2007, the Chinese women's 

entrepreneurial activity index was as high as 11.16%, 

higher than the global average. However, contrast to the 

entrepreneurial enthusiasm, woman enterprises rely more 

on self-accumulation of capital and develop relatively 

slow. 

For the research on women entrepreneurs’ financing 

strategy, many scholars believe that there are some 

Corresponding author: Yongjie Zhang. 


doi:10.4304/jcp.6.9.1819-1824 

Lin Xiong 

The Robert Gordon University ,Scotland, Aberdeen ,UK 

Email: l.xiong@rgu.ac.uk 

differences between male entrepreneurs and women 

entrepreneurs in financing patterns. Women 

entrepreneurs face more difficult in obtaining financing, 

and seem to have some specific financing [1]. Women 

face significant difficulties in external financing, 

particularly bank loans, venture financing. In the past 40 

years, the United States about 40% of the enterprises are 

owned or managed by women, but less than 5% of the 

venture capital invest in the women-led enterprises [2]. 

In order to analyze this subject, scholars give various 

explanations from different points of view. On the supply 

side (behaviors of bankers and public funders), some 

scholars believe that women entrepreneurs encountered 

with some credit discrimination when seeking to external 

financing. Women were required higher interest rates and 

more additional conditions when applying for loan [3]. 

Using the methods of experimental and qualitative 

analyze, Sara Carter et al [4] found that the loans lenders 

assess different conditions when dealing with the 

application of loan from male and female appliers. When 

the variables such as business industry, credit market 

structure are controlled, women business owners still 

have to pay more for the loan, and there is no evidence to 

prove that women enterprises are greater risk than men’s. 

On the demand side (behaviors of women 

entrepreneurs), Orser [5] believe that women 

entrepreneurs are not that eager for business growth, and 

usually consider more on risks. In the process of applying 

for bank loans, women business owners are more 

negative, although their applications will not be easier to 

reject by banks. Some scholars also attempted to explain 

this phenomenon from human capital and social capital 

perspective. Nancy [2] find that only education level of 

women business owners has a significant impact on the 

choice of external equity financing strategy, while social 

capital is not directly affected.



Women entrepreneurs’ financing pattern has been 

noticed by many scholars recently. Some of them tried to 

study the issue from human capital perspective. By 

studying the Finnish company, Cooper et al. [6] found 

entrepreneurs’ education level have a significant impact 

on the enterprise's survival and development. Loscocco et 

al. [7] believe one of the key factors leading to the 

success of small businesses is the relevant industry 

experience. They found that women has disadvantage in 

this area, because the female entrepreneurs than male 

entrepreneurs usually have less relevant experience in 

specific industry. Bosma et al. [8] studied more than 

1,000 Dutch companies, and confirmed the industry 

experience prior to starting one’s own business has 

played an important role in profitability and growth of 

small business. Bates [9] found that college-educated 

entrepreneurs has a higher rate of success of starting 

one’s own company than less educated entrepreneurs, and 

they apply for a loan from commercial banks easier. 

Fabowale et al [10] found that banks increased rejection 

rate of lending loan when women business owners had 

few management experience. Boden and Nucci [11] 

believe that women has less opportunity to accumulate 

human capital because of their lower payment and less 

management experience. 

Some of scholars also tried to carry out the research 

from social capital perspective. Uzzi’s [12] research 

shows that the strong and weak links of the network 

between companies and banks are favorable for applying 

loan, and obtaining lower interest rates. Higgins and 

Gulati [13] found that the more extensive networks 

business owners have, the more excellent investment 

banks willing to underwrite the company’s IPO. Shane 

and Stuart [14] found that venture capitalists seem to be 

more willing to invest in emerging companies those they 

are familiar with, especially the entrepreneurs who had 

previously sponsored or had close contacted before. 

Priscilla chu[15] and other scholars studied the Chinese 

entrepreneurs in Hong Kong and Canada, and found that 

social capital of entrepreneurs could provide an access to 

critical resources for enterprises’ development, such as 

market, technology, capital, knowledge. In addition, 

Tjosvold [16] also believe social capital could be used in 

getting support from government. Pearce and Robinson 

[17] also pointed out that in China, enterprises leaders 

usually set up long term relationship among political 

parties, administrative leaders and other business 

executives, and the relationship with government officials 

is often essential for industrial and commercial 

enterprises on business success. David [18] found that at 

present stage, both private entrepreneurs and managers of 

state-owned enterprises in China all maintain close and 

good social relations with the government and the Party, 

which is one of important ways to obtain economic 

resources. On the other hand, Renzulli, Aldrich, and 

Moody [19] found that compared with men’s, most 

women’s networks are lack of diversity, which will 

hinder women entrepreneurs to identify entrepreneurial 

opportunities as well as search for scare resources. Just as 


Nancy [20] mentioned, the lack of diversity of female’s 

networks results from the lack of contact with those who 

control key resources and critical introducers in their past 

experience. 

III. RESEARCH QUESTIONS AND METHOD 

A.Research questions 

Several specific financing patterns of women-owned 

businesses and relative explanations have been identified 

in the previous section. In this review, we try to verify 

several hypothesis and interpret the reasons of women 

entrepreneurs’ financing strategies from human capital 

and social capital perspective. 

To identify human capital, we use 2 variables. ① the 

result of formal education – education level; ② 

knowledge gained from work experience and practice- 

years of work experience. 

To quantify social capital, we design 3 categories of 

variables, including the type of social network, the scale 

of social network, and the intensity of social network. 

The network we surveyed are following 6 types of 

network, the member of Deputy to People’s Congress, 

CPPCC National Committee member, the Federation of 

Industry member, member of the Individual and Private 

Entrepreneur Association, Chinese Communist Party, and 

Democratic Party. We also investigated the highest duties 

and the levels of network in the first 4 networks. The 

scale of network is measured by the number of 

membership. The greater the number of membership is, 

the larger network is. The intensity of network is 

measured by the following indicators: ① the cost spends 

on social activities in that year; ② average time spend on 

social activities weekly. The greater the value of these 

two indicators shows, the greater the intensity of the 

network is. 

Meanwhile, we have three hypotheses to verify in 

Chinese market. 

H1: There are gender differences when entrepreneurs 

choose financing strategy. 

H2: More human capital female entrepreneurs have, 

external financing are more likely to be used. 

H3: More social capital female entrepreneurs have, 

external financing are more likely to be used. 

The conceptual model could be established in 

following way: 

Figure.1 conceptual model of analysis


B.The data and research method 

This study is based on the data from the 2002 China 

Private Economy Research survey, a national wide 

survey conducted by the CPC Central Committee United 

Front Work Department, China Federation of Industry 

and State Administration for Industry. To ensure data’s 

integrity, reliability, we use Microsoft Excel to eliminate 

invalid original survey for the preliminary screening. 

After deleting the data those are incomplete or 

obviously invalid data, we carry out mean value analysis 

According to above analysis, we obviously notice that 

men and women business owners have significant 

difference when choosing whether using external 

financing (Sig = 0.001), and averagely men prefers to 

external financing than women (Mean Male > Mean 

Female). However, the proportion of start-up capital in 

individual investors does not appear significant difference 

in genders. 

To identify the number of factors, and the correlation 

between the observed variables, we carried correlation 

between the observed variables, we carried out 

TABLE 4-1 SAMPLES INFORMATION 

Then we use SPSS17.0 statistics software to process the 

data, and the main analysis methods include reliability 

and validity analysis, descriptive statistics analysis, 

ANOVA, exploratory factor analysis, Logistic regression. 

IV. EMPIRICAL ANALYSIS 

Before starting our analysis, we summarize basic 

information of our samples as follows, shown in table 4-1: 

Gender Education Age 

Type: Male 

No:2328 

89.9% 

Type: Female 

No:261 

10.1% 

Type No. % Type No. % 

Primary 49 2.1 Under 25 11 0.5 

Junior 434 18.6 26~35 301 12.9 

Senior 983 42.2 36-45 1024 44.0 

College 764 32.8 46-55 798 34.3 

Master 98 4.2 Above 55 194 8.3 

Primary 7 2.7 Under 25 0 0 

Junior 31 11.9 26~35 44 16.9 

Senior 117 44.8 36-45 122 46.7 

College 90 34.5 46-55 76 29.1 

Master 16 6.1 Above 55 19 7.3 

Table 4-2 ANOVA Result 

Gender Work experience Industry 

and ANOVA based on gender differences. The results are 

shown in Table 4-2: 

Type No. Percent Type No. Percent 

Male Less than 1 year 37 1.6 Manufacturing category ( Agriculture, 

Mining, Manufacturing, Geology, 

Construction, Electricity & Gas) 

1306 56.1 

2-5 years 142 6.1 Service category(Food service, Finance, 

Insurance, Real State, Social Services, 

Education, Scientific Research, Health) 

806 34.6 

6-10 years 339 14.6 Others 216 9.3 

More than 10 

years 

1810 77.7 

Female Less than 1 year 7 2.7 Manufacturing category ( Agriculture, 

Mining, Manufacturing, Geology, 

Construction, Electricity & Gas) 

102 39.1 

2-5 years 14 5.4 Service category(Food service, Finance, 

Insurance, Real State, Social Services, 

Education, Scientific Research, Health) 

136 52.1 

6-10 years 34 13.0 Others 23 8.8 

More than 10 

years 

206 78.9 


exploratory factor analysis by using formal questionnaire 

date. The data’s KMO = 0.660, and Bartlett's test of the 

value of spherical is 6678.330, whose Sig =.000


TABLE 4-3 FACTORS MATRIX AFTER ROTATION 

Component 

1 2 3 4 5 

F10 Post in Private 

Entrepreneur Association 

.841 

F11 Level in Private 


.815 

F3 No. of social networks .638 

F8 Industry and Commerce 

Level 

.511 

F9 Industry and Commerce Post .444 

F6 CPPCC Post .897 

F7 CPPCC Level .873 

F4 NPC Post .887 

F5 NPC Level .882 

F2 Work experience -.763 

F1 Education .761 

F12 Cost for social activities .681 

F13 Time spends on social 

activities 

.530 

To further study the factors those impact women 

entrepreneurs’ financing strategy, we tried to establish a 

Logistic Regression Model. The results are as follows, 

shown in table 4-4, 4-5, and 4-6: 

Table 4-4 Hosmer and Lemeshow Test 

Step Chi-square df Sig. 

1 3.798 5 .579 

2 2.955 6 .814 

3 1.076 6 .983 

4 6.893 7 .440 

According to above tables, we can conclude our 

equation: 

P=1/(1+e-z) 

And z=-3.383 +0.504*Work experience+0.057*Cost 

for social activities+0. 774*Industry and Commerce+0. 

.375*Level in Private Entrepreneur 

Association+21.669*NPC Post. 

We may notice that the coefficients of work 

experience, cost for social activities, whether to join the 

Industry and Commerce, Level in Private Entrepreneur 

Association and the NPC Post, which are as measures of 

human capital and social capital, are positively correlated 

with external financing. Thus, this result confirmed 

hypothesis H2 and H3, that is the more human capital and 

social capital women entrepreneurs have, the more they 

tend to use external financing. 

We can explain these results from the previous 

literature and the social status. 

(a) human capital: The work experience, as an 

indicator of human capital, may promote using external 

financing. From the demand side, it may enhance the 

relevant skills and accumulate of the managerial 

experience. From the investors’ points of view, whether 

having relative work experience is considered an 


important measure, when applying bank’s loan and 

attracting venture investment, which has also been 

confirmed in previous research. However, another 

measure of human capital, education background, was not 

evolved in our final model equation. Therefore, the 

higher education of women entrepreneurs does not means 

they would tend to use external financing. But this result 

is in accordance with Hu Huaimin's finding, that is the 

number of women entrepreneurs and their education level 

distributed in inverted “U ", that is to say, both less 

educated and higher educated women are not that 

interested in starting their own business. 

(b) the intensity of network: Cost for social activities, 

as an indicator of the intensity of network, may reflect the 

maintenance of social networks by women entrepreneurs 

to some extent. Scholars have found that there are 

triangular interactions among emotions, resources and 

interactive reciprocal relationship, which is the more the 

interaction between individuals; the more likely they are 

to participate in group activities to share feelings, the 

more likely to exchange resources. Therefore, spending 

more on social activities and participating in more 

interactive activities, women entrepreneurs may obtain 

external financing resources easier. 

(c) networks between enterprises and industries: 

Although Chinese people values family relationship 

deeply, family network still cannot meet the needs of 

enterprises’ development. Therefore, joining in the 

Industry and Commerce, Association of Private 

Entrepreneurs may expand women entrepreneurs’ social 

circle and benefit their career. On one hand, among these 

networks between enterprises and industries, network 

members are engaged in similar activities, and exchange 

information of different market, related technical advice 

and financing with women entrepreneurs. On the other 

hand, female entrepreneurs have more opportunity to the 

key figures mastering scarce financial resources by these 

networks. 

(d) Administrative impact:The coefficient of NPC 

Post, a measure of administrative impact, is much 

greater than the other three, which explains this factor has 

larger impact of using external financing. In China, the 

people's congress is China's highest authority. Deputies 

generally have high social reputation in social life, and 

also have some administrative influence in the 

administrative system. Therefore, when women 

entrepreneurs take some deputies duties, such as director 

or deputy director of the Standing Committee, it will 

increase their personal reputation to a large extent. 

Because of financial market imperfections and lack of 

policy stability, informal constraints in the economy 

during the transition period has played an important role 

in society, and network and people's trust has become 

extremely important. Thus, taking some duties in NPC is 

beneficial to attract external financing. Besides, based on 

previous studies, women, lacking of access to critical 

resources to grasp chances, are often disadvantaged in the 

network status. In China, about 75% of the deputies to the 

NPC are officials. Therefore, by serving as a certain NPC 

Post, women entrepreneurs will have an access to the


some scarce financial resources and optimize their 

network infrastructure. 

According to above analysis, we establish our model as 

follows. 

Figure.2 Women entrepreneur’s financing strategy 

V. CONCLUSION 

The main finding of this paper including: 

(a) From the result of ANOVA, we may notice that 

during the development of enterprise, men and women 

Table 4-5 Variable in equation 

entrepreneurs have a significant difference in using 

internal or external financing. 

(b) Female entrepreneurs who have more human 

capital are more likely to use external financing. Work 

experience of women entrepreneurs has positive 

correlation to external financing, while education 

background is not significant in China. 

(c) Social capital on entrepreneurial financing 

strategies was significant. From the logistic regression 

equation, we find that Cost for social activities, Industry 

and Commerce, Association of Private Entrepreneurs and 

NPC Post are the key factors of promoting using external 

capital for Chinese women entrepreneurs. 


This research is supported by NSFC (Project 

70603021) and Royal Society of Edinburgh and National 

Natural Science Foundation of China for financial 

support (Project 70911130020). 

B S.E. Wald df Sig. Exp(B) 

Step 1 a Cost for social 

activities 

.062 .023 7.664 1 .006 1.064 

Constant -.707 .152 21.555 1 .000 .493 

Step 2 b Cost for social 

activities 

.062 .023 7.052 1 .008 1.064 

Level in Private 

Entrepreneur 

Association 

.428 .159 7.216 1 .007 1.534 

Constant -.844 .164 26.611 1 .000 .430 

Step 3 c Cost for social 

activities 

.061 .023 6.966 1 .008 1.062 

NPC Post 21.636 19923.521 .000 1 .999 2.491E9 


Entrepreneur 

Association 

.422 .160 6.948 1 .008 1.525 

Constant -.881 .165 28.519 1 .000 .415 

Step 4 d Work experience .495 .242 4.185 1 .041 1.641 

Cost for social 

activities 

.060 .023 6.758 1 .009 1.061 

NPC Post 21.502 19939.319 .000 1 .999 2.178E9 


Entrepreneur 

Association 

.409 .162 6.423 1 .011 1.506 

Constant -2.719 .928 8.591 1 .003 .066 

Step 5 e Work experience .504 .245 4.241 1 .039 1.655 

Cost for social 

activities 

.057 .022 6.625 1 .010 1.059 

Industry and 

Commerce 

.774 .401 3.730 1 .053 2.169 

NPC Post 21.669 19544.921 .000 1 .999 2.575E9 


Entrepreneur 

Association 

.375 .164 5.259 1 .022 1.455 

Constant -3.383 1.004 11.359 1 .001 .034 

a. Variable(s) entered on step 1: Cost for social activities. b. Variable(s) entered on step 2: Level in Private Entrepreneur Association.c. 

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 

Commerce. 



REFERENCES 

[1] Nancy Carter, Candida Brush, Patricia Greene, Elizabeth 

Gatewood and Myra Hart, “Financing High-Growth 

Enterprise: Is Gender an Issue?”, J. Women in Business: 

Access and Success, 2003: 45-52. 

[2] Candida G. Brush, Nancy M. Carter, Patricia G. Greene, 

Myra M. Hart, “The role of social capital and gender in 

linking financial suppliers and entrepreneurial firms: a 

framework for future research”, J. Venture Capital, 2002, 

Vol. 4, Issue 4:305 – 323. 

[3] Coleman S, “Access to capital and terms of credit: A 

comparison of men-and women owned small businesses”, 

J. Journal of Small Business Management, 2000, 38 

(3):37-5. 

[4] Sara Carter, Eleanor Shaw, Wing Lam and Fiona Wilson, 

“Gender, Entrepreneurship and Bank Lending: The 

Criteria and Processes Used by Bank Loan Officers in 

Assessing Applications”, J. Entrepreneurship Theory and 

Practice, 2007, 31(3):427 – 444. 

[5] Barbara Orser, Sandra Hogarth-Scott, “Opting for Growth: 

Gender Dimensions of Choosing Enterprise Development”, 

Canadian Journal of Administrative Sciences, 2002, 

19(3): 284 – 300. 

[6] Cooper, A. C, F. J. Gimeno-Gascon and C. Y. Woo, 

“Initial Human and Financial Capital as Predictors of New 

Venture Performance”, Journal of Business Venturing, 

1994, 9: 371–395. 

[7] Loscocco, K. A., J. Robinson, R. H. Hall and J. K. Allen, 

“Gender and Small Business Success: An Inquiry into 

Women’s Relative Disadvantage”, J. Social Forces 1991, 

70(1): 65–85. 

[8] Bosma, N., M. van Praag, R. Thurik and G. de Wit, “The 

Value of Human and Social Capital Investments for the 

TABLE 4-6 THE MODEL AFTER REMOVING SOME OF THE VARIABLES 

Variable Model Log Likelihood 

Change in -2 Log 

Likelihood df Sig. of the Change 

Step 1 Cost for social activities -171.306 11.290 1 .001 




-165.660 7.569 1 .006 


NPC Post -161.876 7.332 1 .007 



-161.830 7.239 1 .007 

Step 4 Work experience -158.210 4.916 1 .027 

Cost for social activities -160.773 10.042 1 .002 

NPC Post -159.101 6.697 1 .010 



-159.101 6.698 1 .010 

Step 5 Work experience -156.215 4.995 1 .025 

Cost for social activities -158.475 9.515 1 .002 

Industry and Commerce -155.752 4.069 1 .044 

NPC Post -157.328 7.221 1 .007 



-156.461 5.487 1 .019 


Business Performance of Startups”, J. Small Business 

Economics, 2004, 23(3): 227–236. 

[9] Bates, T, “Entrepreneur Human Capital Inputs and Small 

Business Longevity”, J. The Review of Economics and 

Statistics, 1990, 72(4): 551–559. 

[10] Fabowale, L., Orser, B. and Riding, A. ,”Gender, 

structural factors and credit terms between Canadian small 

businesses and financial institutions”, J. Entrepreneurship 

Theory and Practice, 1995,19(4): 41 – 66. 

[11] Boden, R. J. and Nucci, A. R., “On the survival prospects 

of men’s and women’s new business ventures”, J. Journal 

of Business Venturing, 2000, 15(4): 347 – 362. 

[12] Uzzi B, “Embeddedness in the Making of Financial 

Capital: How Social Relations and Networks Benefit 

Firms Seeking Financing”, J. American Sociological 

Review, 1999, 64(4): 481-505. 

[13] Higgins M, Gulati R, “Getting off to a good start: the 

effects of upper echelon affiliations on underwriter 

prestige”, J. Organization Science, 2003, 14(3): 244–263. 

[14] Shane S, Stuart TE, “Organizational endowments and the 

performance of university startups”, J. Management 

Science, 2002, 48(1): 154–170. 

[15] Priscilla Chu, “Social Network Models of Overseas 

Chinese Entrepreneurship: The Experience in Hong Kong 

and Canada”, J. Canadian Journal of Administrative 

Sciences, Dec.1996; 13(4): 358-365. 

[16] Tjosvold, Dean, Weieker, David, “Cooperative and 

competitive networking by entrepreneurs: A critical 

incident study”, J. Journal of Small Business 

Management. Milwaukee: 1993, 31(1):11 一 22. 

[17] Zhilong Tian, Yongqiang Gao et al, “Chinese Enterprises’ 

Political Strategy and Behavior Study”, J. Management 

World, 2003, 12.


A Spatial Econometric Analysis of China’s 

Manufacturing Agglomeration based on Geoda 

and Matlab 

Huayin Yu 

Department of Statistics, Anhui University of Financial and Economics, Bengbu, China 

Email: y_hyin@163.com 

Weiping Gu 

Department of Statistics, Anhui University of Financial and Economics, Bengbu, China 

Abstract—Industrial agglomeration has gradually become 

an economic focus in recent years. Scholars has done a lot of 

research about the formation mechanism of industry 

agglomeration and its influencing factors, but the spatial 

correlation of variables has still been neglected. Firstly this 

paper gives a brief introduction about Geoda software and 

Matlab neural network toolbox, then use spatial statistical 

methods to describe the 1999-2008 China's manufacturing 

industry agglomeration. Secondly this paper uses spatial 

econometric methods to analyze the influencing factors of 

China’s provincial manufacturing Agglomeration. The 

results show that the spatial econometric model is superior 

to the traditional econometric models and the analysis based 

on spatial econometric model are more accurate. Finally, the 

paper also gives a brief forecast of the manufacturing 

Agglomeration. 

Index Terms—manufacturing Agglomeration, spatial 

correlation, spatial lag model, spatial error model, BP 

neural network 


Geoda is a collection of software developed by Luc 

Anselin. It has a friendly and graphical interface that 

users can easily implement exploratory spatial data 

analysis (ESDA) with it, such as spatial autocorrelation 

analysis and spatial econometric analysis. The Geoda 

software includes an interactive environment that 

combines maps with statistical graphics, using 

dynamic-linked-window technology. Its original version 

date back to the first contribution made to develop a 

bridge between ESRI’s ArcInfo GIS and the SpaceStat 

software. The second version of Geoda made an 

improvement to ESRI’s ArcView 3.x GIS that it can 

implement linked windows and brushing. In contrast to 

the previous versions, the current Geoda is independent 

software that runs under any of the Microsoft Windows 

operating systems without a specific GIS system. 

Matlab is an advanced language and interactive 

environment that users can implement numerical 

computation with it. And its operational efficiency is 

much higher than traditional programming languages 

such as C, C++, and FORTRAN due to the excellent 


doi:10.4304/jcp.6.9.1825-1831 

design. Matlab can perform many complex tasks such as 

signal and image processing, computation, control system 

design, test and measurement, financial modeling and 

analysis. There are more than 30 Toolboxes in Matlab 

and they can be divided into two categories: functional 

toolbo x and field-based toolbox. The functional toolbox 

is mainly used to expand symbolic computing, modeling 

and simulation capabilities, word processing and 

hardware real-time interactivity. Functional toolbox can 

be used in a variety of disciplines. In the Opposite, the 

field-based toolbox is highly professional, such as the 

control system toolbox, signal processing toolbox and 

finance toolbox. And neural network toolbox is one of 

them. It extends Matlab with tools for designing, 

implementing, visualizing, and simulating neural 

networks. 

II. DESCRIPTIVE STATISTICS 

As a branch of econometrics, spatial econometrics 

focuses on dealing with spatial interaction and spatial 

structure in cross-sectional data and panel data regression 

model. This area has developed rapidly in recent years. 

Spatial econometrics is widely used in applied economics 

and policy analysis, particularly in regional economics, 

residential economics, environmental and resource 

economics and development economics and other fields. 

Firstly, this paper made an exploratory spatial data 

analysis of China’s manufacturing agglomeration with 

Geoda. Secondly, we performed a spatial econometric 

analysis on influencing factors of china’s manufacturing 

agglomeration using spatial lag model and spatial error 

model. Finally, we used Matlab neural network toolbox 

to predict China’s manufacturing agglomeration based on 

the existing data. From an economic point of view, this 

article can also be seen as an example of spatial 

econometric analysis. 

A. Dependent Variable and Indicators 

We have many indicators to measure the industrial 

agglomeration in the actual study. In this paper, we chose


Location Quotient (LQ) to describe China’s 

manufacturing agglomeration. It is defined as follows: 

LQ ij = ( Eij 

Ei 

) ( Ekj 

Ek 

) 

In the formula above, ij E indicate the employment 

in j 

industry of i district; Ei indicate the total 

employment of i kj district; E indicate the employment 

in j industry of the total district k ; Ek indicate the total 

employment of the total district k . It is generally believed 

that the greater the Location Quotient coefficient, the 

higher the level of the region's industry agglomeration. 

B. Spatial Statistical Analysis of China’s Manufacturing 

Agglomeration 

(a) Spatial distribution of China's manufacturing 

industry 

To give a better analysis of the spatial variation 

process of China’s manufacturing agglomeration, we 

mapping the spatial percentile chart (three periods: 

1999-2001, 2002-2005, 2006-2008) with Geoda095i, 

based on provincial Location Quotient coefficient 

(Calculated average for each period). The results is 

shown in chart 1. 

(a): 1999-2001 

(b): 2002-2005 

(c):2006-2008 

chart 1: Spatial percentile chart of China’s provincial 

manufacturing agglomeration (1999-2008, three periods) 

From chart 1 we can see that: From 1999 to 2001, 

Shanghai, China’s economic center, got the highest LQ 

coefficient and rank the first echelon; Beijing, Tianjin 


rank the second echelon; Liaoning, Hebei, Shandong, 

Jiangsu, Zhejiang, Fujian, Guangdong and some 

provinces ( 1of 

) central region rank the third echelon; 

however, the agglomeration of manufacturing industry in 

Xizang, Yunnan and Hainan are still at a low level. From 

2002 to 2005, Shanghai still rank the first echelon; 

however, instead of Beijing, Guangdong came into the 

second echelon; both Inner Mongolia and Ningxia move 

forward to the third echelon; Shanxi, Gansu drop to the 

fourth echelon. Spatial percentile chart of 2006-2008 

hasn’t changed compared to the 2002-2005’s. 

As can be seen from the above analysis, China's 

manufacturing industry mainly concentrated in the 

southeast coastal areas. Manufacturing sector of coastal 

areas showed an increasing trend, but this trend is 

gradually slowing down. 

(b) Spatial autocorrelation analysis of China's 

manufacturing agglomeration 

In actual research we often use Moran'I index to test 

the existence of spatial autocorrelation, which is defined 

as follows: 

n n 

∑∑Wij( 

Yi 

−Y 

)( Yj 

−Y 

) 

Mora 

i= 

1 j= 

1 

n′ 

s I = 

n n 

2 

S 

∑∑ 

i= 

1 j= 

1 

W 

ij 

n 

= ∑ i − 

In the formula above, 

i= 

Y Y 

n 

2 1 

2 1 

S ( ) Y = ∑ Yi 

n 1 , n i= 

1 , 

Yi is the value ofi district, n is the number of district, 

Wij is the Contiguity Based Spatial Weights: if 

region i and region 

j ij 

is adjacent, 

W =1; otherwise, 

W ij =0. 

i = 1, 2, 

⋅⋅⋅, 

n ; j = 1, 

2, 

⋅⋅⋅, 

m ; m = n 

or n ≠ m . 

Moran’s I rank from -1 to 1. 

For Moran's I index results, we can use standardized 

statistic Z to test the existence of spatial autocorrelation 

between the regions. 

I −E( 

I) 

Z = 

( 3) 

VAR( 

I) 

Under the assumption of normal distribution, the 

expectation and variance of Moran's I can be calculated 

as follows: 

2 

2 

1 n w1 

+ nw2 

+ 3w0 

2 

E( I) 

= − , VAR( 

I) 

= 

−E 

( I) 

( 4) 

2 2 

n−1 

w0 

( n −1) 

In the formula above, 

n n 

n n 

n 

1 

2 

2 

w0 

= ∑∑w 

ij, 

w1 

= ∑∑( 

wij 

+ wji) 

, w2 

= ∑( 

wi. 

+ wj.) 

, 

i= 

1 j= 

1 2 i= 

1 j= 

1 

i= 

1 

wi. wj. and are the sum of row i and column 

j 

of the 

spatial weight matrix respectively. Both mean and 

variance are theoretical. 

We can make a significant test of the spatial 

H 

autocorrelation based on the statistic Z caculated. 0 : 

spatial autocorrelation between the regions does not exist. 

In Geoda095i, we use Monte Carlo method to test the 

existence of spatial autocorrelation, and the significant 

( 2)


level is determined by the p value of the statistic Z. If p < α , 

0 H is denied; otherwise, 

Table 1: Moran’s I value of China’s manufacturing agglomeration (1999-2008) 

Year Moran’s I Mean 

Standard dev 

iation 

p value 

1999 0.3191 -0.0333 0.1073 0.0060 

2000 0.3216 -0.0330 0.1087 0.0058 

2001 0.3342 -0.0333 0.1072 0.0042 

2002 0.3063 -0.0337 0.1081 0.0062 

2003 0.3147 -0.0321 0.1094 0.0053 

2004 0.3294 -0.0337 0.1097 0.0042 

2005 0.3421 -0.0345 0.1112 0.0031 

2006 0.3572 -0.0335 0.1131 0.0031 

2007 0.3377 -0.0332 0.1100 0.0032 

2008 0.3496 -0.0324 0.1090 0.0032 

From table 1 we can make a conclusion that there was 

a significant positive spatial autocorrelation between 

China’s provincial manufacturing industries. This 

indicates that China's manufacturing industry did not 

distribute randomly, and the spatial distribution of 

manufacturing industry showed a clear concentration 

trend over the last decade: the provinces that have similar 

LQ coefficient tend to concentrate geographically. 

III. METHOD AND MODEL 

A. Research Methods 

(a) Spatial Lag Model 

Spatial Lag Model (SLM) is mainly used to discuss 

whether there is a spillover effect of variables in a region. 

The model is expressed as follows: 

y = ρ Wy + Xβ 

+ ε 

In the formula above, y is a dependent variable; 

X is a n × k matrix of exogenous explanatory 

variables; ρ is a spatial regression coefficient, reflecting 

the effect of spatial dependence in observations; W is a 

n× n matrix of spatial weight; 

Wy 

is a spatial lagged 

dependent variable; ε is a random error vector. 

Parameter β reveals the effect the explanatory 

variable X has to dependent variable y . Spatial lagged 

dependent variable Wy is a exogenous variable 

reflecting how spatial distance influence the act of 

regions. The act of regions is strongly affected by the 

cultural environment and the transfer cost related to 

spatial distance. 

(b) Spatial Error Model 

Spatial Error Model (SEM) is expressed as follows: 

y = Xβ 

+ ε, 

ε = λWε 

+ µ 

In the formula above, ε is a random error vector; λ 

is a n × 1 spatial error matrix of dependent variable 

vector; µ is a random error vector in normal 

distribution. 

Parameter β reveals the effect the explanatory 

variable X has to dependent variable y . Parameter λ 


0 H is accepted. 

reflects the effect of spatial dependence in observations. 

The spatial dependence in random error term measures 

how the error impact of dependent variable of 

neighboring areas influence the observations in this 

region. 

(c) Estimation method 

Considering the endogeneity of explanatory variables 

in spatial regression model, coefficient estimates will be 

biased or invalid if we use OLS method to estimate the 

coefficient in Spatial Lag Model and Spatial Error Model. 

Instead, we can use other methods to estimate, such as 

Instrumental Variable method, Maximum Likelihood 

method, Generalized Least Squares method and 

Generalized Method of Moments. Anselin (1998) 

recommended Maximum Likelihood method for 

estimating the coefficient in SLM and SEM. 

(d) ( A 5) 

choice between SLM, SEM and spatial 

autocorrelation test 

Anselin and Florax (1995) proposed the following 

criterion: We can determine that Spatial Lag Model 

would be more appropriate if (a) LMLAG is more 

significant than LMERR statistically; (b) R-LMLAG is 

significant but R-LMERR is not significant. In the 

Opposite, Spatial Error Model would be better if (a) 

LMERR is more significant than LMLAG; (b) 

R-LMERR is significant but R-LMLAG isn’t. 

Besides R-squared, some common criterion includes 

Log likelihood, Likelihood Ratio, Akaike Information 

Criterion (AIC) and Schwartz Criterion (SC). The higher 

the Log likelihood, the lower the AIC and SC, the 

better the model. These indicators can also be used to 

compare the regression effect between OLS, SLM and 

SEM. 

B. Econometric Model 

In this ( article, 6) 

we proposes the main factors that 

affecting industrial agglomeration from four perspectives: 

comparative advantage, new economic geography, 

knowledge spillovers and the role of government. 

Considering measurability of indicators and availability 

of data, we proposes the following twelve indicators in 

table 2.


Table 2: Explanatory variables and their settings 

Explanatory variables Symbol Meaning 

Agriculture endowment agrgift Added value of agriculture/GDP 

Endowment of natural resources natgift Output of extractive industries/GDP 

Capital endowment capgift Rate of capital formation 

Endowment of human resources humgift Number of students in higher school/Population 

Level of urbanization city Urban employment/Total employment 

Market demand demand Provincial GDP per capita/National GDP per capita 

Transportation road Classified road density 

Industrial foundation industry Added value of industry /GDP 

Openness open Foreign direct investment/GDP 

Supporting services service Added value of tertiary industry /GDP 

Patent approvals patent Provincial patent approvals/National patent approvals 

Financial income fisc-income Financial income/GDP 

In table 2, Agriculture endowment, Endowment of 

natural resources, Capital endowment and Endowment of 

human resources are the indicators of comparative 

advantage; Level of urbanization, Market demand, 

Transportation and Industrial foundation are the 

indicators of new economic geography; Openness, 

Supporting services and Patent approvals are the 

indicators of knowledge spillovers; Financial income is 

the indicator that reflect the role of government. On this 

basis, We recommend the following double logarithmic 

model: 

ln LQ = β0 

+ β1 

ln agrgift + β2 

ln natgift + β3 

ln capgift 

+ β6 

ln demand + β7 

ln road + β8 

lnindustry 

+ β9 

+ β11 

ln patent + β12 

ln fisc − income + ε 

β 

In the equation above, i are regression coefficient, 

i = 1, 2, 

⋅⋅ 

⋅30, 

ε is a random error term. In the 

following empirical analysis, adjustment may be made to 

the model based on the actual situation. 

The sample in this paper includes all the provinces, 

autonomous regions and municipalities in China except 

Hong Kong, Macao and Taiwan (Chongqing is taken into 

Sichuan for convenience). All data can be found in 

"China Statistical Yearbook", "China Industrial Economy 

Statistical Yearbook" from 2000-2009 and the website of 

The People's Bank of China. 

IV. EMPIRICAL ESTIMATION AND RESULTS 

A. Econometric Analysis 

Considering the formation of industrial agglomeration 

is a process, it will take some time for the effect 

becoming apparent. In this article, we set the Location 

Quotient of provincial Manufacturing of 2008 as 

+ dependent β4 

ln humgift variable + β5 

ln and city the twelve indicators of 2006 as 

lnexplanatory 

open + β variables. 

10 ln service 

( 7) 

First of all, we make a OLS estimation including all 

the factors. As can be seen from estimation 1 of table 3, 

the t value of most variables is not significant and severe 

multicollinearity exists in the model. So some adjustment 

need to be made until t value of most variables become 

significant and model’s multicollinearity is weaken. 

Results after adjustment is shown in estimation 2 of table 

3. Obviously, the model of estimation 2 is better than the 

model of estimation 1. 

Table 3: OLS estimation of the factors of manufacturing agglomeration (2006-2008) 

estimation 1 estimation 2 

Variables 

Coefficient 

Standard 

deviation 

t value p value Coefficient 

Standard 

deviation 

t value p value 

C 1.0524 2.1146 0.4976 0.6250 1.5389 0.5981 2.5728 0.0166 

lnagrgift -0.2014 0.2413 -0.8345 0.4155 

lnnatgift -0.0147 0.0641 -0.2294 0.8212 

lncapgift 0.3917 0.3991 0.9814 0.3401 

lnhumgift -0.0791 0.2610 -0.3031 0.7654 

lncity 0.2485 0.3863 0.6432 0.5286 

lndemand 0.2833 0.4734 0.5984 0.5574 0.8064 0.1133 7.1131 0.0000 

lnroad 0.1051 0.1081 0.9720 0.3446 0.0898 0.0522 1.7205 0.0982 

lnindustry 1.2566 0.4098 3.0661 0.0069 1.3435 0.2508 5.3550 0.0000 

lnopen 0.0744 0.0769 0.9678 0.3466 

lnservice 1.1140 0.8528 1.3062 0.2088 1.3270 0.6093 2.1779 0.0394 

lnpatent 0.0852 0.0849 1.0029 0.3299 

lnfisc-income -0.5922 0.3889 -1.5228 0.1461 -0.3536 0.2448 -1.4441 0.1616 

R 2 0.8936 0.9142 

LogL 6.6697 4.7362 

AIC 12.6606 2.5274 

SC 30.8761 10.9347 

F 21.3008 0.0000 62.8634 0.0000 



As can be seen from the above analysis, the main 

factors that influence China’s provincial manufacturing 

agglomeration include Market demand, Transportation, 

Industrial foundation, Supporting services and Financial 

income. Industrial foundation and Supporting services 

have positive impact on manufacturing agglomeration. 

Their elasticities are 1.3435 and 1.3270. While Financial 

income has a negative impact on manufacturing 

agglomeration with its elasticity -0.3536. Due to the fault 

of OLS method dealing with spatial autocorrelation, we 

intend to use spatial econometric models to analyze the 

influencing factors of manufacturing agglomeration. The 

new model is the one in estimation 2 of table 3: 

ln LQ = β0 

+ β1 

ln demand + β2 

ln road + β3 

ln 

+ β4 

ln service + β5 

ln fisc − income + ε 

We need to verify the existence of spatial 

autocorrelation before peforming spatial econometric 

analysis. The result of spatial dependence test is shown in 

table 4. 

Table 4: Spatial dependence test 

Spatial dependence MI/DF Statistics p value 

Moran’s I (Error) 0.1047 1.9679 0.0489 

LMLAG 1 0.2102 0.6465 

R-LMLAG 1 0.4540 0.5003 

LMERR 1 0.0931 0.7602 

R-LMERR 1 0.3370 0.5615 

LM-SARMA 2 0.5472 0.7606 

As is shown in table 4, LMERR, LMLAG, R-LMERR 

and R-LMLAG do not pass the test at significance level 

of 5%. And SEM is better comparing the value of LogL, 

industry AIC, SC and LR of SLM and SEM. The result of 

estimation of SLM and SEM ( 8) 

is displayed in table 5. 

Table 5: SLM and SEM estimation of the factors of manufacturing agglomeration (2006-2008) 

SLM SEM 

Variables 

β 

Standard 

deviation 

t value p value β 

Standard 

deviation 

t value p value 

C 1.6128 0.5600 2.8797 0.0039 1.5984 0.5300 3.0154 0.0025 

lndemand 0.7757 0.1192 6.5054 0.0000 0.8018 0.0975 8.2223 0.0000 

lnroad 0.0742 0.0573 1.2952 0.1952 0.0851 0.0437 1.9457 0.0516 

lnindustry 1.3642 0.2306 5.9145 0.0000 1.3741 0.2224 6.1785 0.0000 

lnservice 1.3462 0.5464 2.4635 0.0137 1.3585 0.5442 2.4960 0.0125 

lnfisc-income 0.3370 0.2200 1.5319 0.1255 0.3516 0.2199 1.5985 0.1099 

ρ/λ 0.0621 0.1208 0.5146 0.6067 0.3324 0.1275 2.6064 0.0074 

Statistical test DF Statistics p value DF Statistics p value 

R 2 0.9296 0.9495 

LogL 4.8505 4.8034 

AIC 4.2989 2.3930 

SC 14.1073 10.8002 

LR 1 0.2285 0.6325 1 0.1344 0.7138 

As is shown in table 5, the coefficients of all variables 

are positive. It indicate that this estimation is more 

consistent with theoretical analysis. Industrial foundation 

and Supporting services have positive impact on 

manufacturing agglomeration. Their elasticities are 

1.3642 and 1.3462 in SLM, 1.3741 and 1.3585 in SEM. 

B. Prediction of Location Quotient of China’s Provincial 

Manufacturing 

BP neural network is error back propagation neural 

network and feedforward network. It is widely used in 

function approximation, pattern recognition and data 

compression. In this article, we will predict the 2009’s 

Location Quotient of China’s Provincial Manufacturing 

using Neural Network Toolbox in Matlab (2006, 2007 

and 2008’s location quotient are references for 

comparison). The Matlab program of prediction is as 

follows (take Beijing for example): 

clc 

close all 

clear all 

p0=[2.4289 2.1880 2.0585 1.6110 1.3667 1.4216 

1.4169 1.3470 1.2994 1.3316]; 

day=1999:2008; 

plot(day, p0,'b+') 


hold on 

plot(day, p0, 'r-.') 

p1=(p0-min(p0))./(max(p0)-min(p0)); 

for i=1:5; 

p(:,i)=p1(i:i+2); 

t(:,i)=p1(i+3); 

end 

p; 

t; 

for i=1:5; 

testp(:,i)=p1(i+1:i+3); 

testt(:,i)=p1(i+4); 

end 

net=newff(minmax(p),[20,1],{'logsig','purelin'},'trainl 

m'); 

net.trainParam.lr=0.8; 

net.trainParam.epochs = 500; 

net.trainParam.goal = 0.001; 

net=train(net,p,t); 

y=sim(net,testp); 

E=testt-y;


p01=p1(5:7)'; 

y01=sim(net,p01); 

y08=min(p0)+y01*(max(p0)-min(p0)); 

E01=(p0(8)-y08)/p0(8); 

p02=p1(6:8)'; 



E02=(p0(9)-y09)/p0(9); 

p03=p1(7:9)'; 



E03=(p0(10)-y10)/p0(10); 

y=[y08 E01;y09 E02;y10 E03] 

p04=p1(8:10)'; 


y11=min(p0)+y04*(max(p0)-min(p0)) 

p2=[p0 y11]; 

day1=[day 2009]; 

figure 

plot(day, p0,'g-.') 

hold on 

plot(day, p0,'b+') 

hold on 

plot(2006,y08,'ro') 

plot(2007,y09,'ro') 

plot(2008,y10,'ro') 

plot(2009,y11,'ro') 

plot(day1,p2, 'k-.') 

Table 6 shows the details about the prediction of the 

location quotient of China's provincial manufacturing. 

table 6: Result of BP neural network prediction (2006-2009) 

2006 2007 2008 2009 

Provinces Actual Predicted Relative Actual Predicted Relative Actual Predicted Relative Predicted 

value value error value value error value value error value 

Beijing 1.3471 1.3471 0.0000 1.2995 1.3513 -0.0399 1.3317 1.3417 -0.0076 1.3234 

Tianjin 3.0678 3.0680 -0.0001 3.3412 3.2362 0.0314 3.1160 3.1768 -0.0195 3.3121 

Hebei 0.9264 0.9268 -0.0004 0.9620 0.9702 -0.0085 0.9860 0.9975 -0.0117 0.9678 

Shanxi 1.1078 1.1077 0.0001 1.1452 1.1093 0.0314 1.1828 1.1927 -0.0084 1.2705 

Inner Mongolia 0.8289 0.8293 -0.0004 0.8529 0.8756 -0.0266 0.8782 0.8837 -0.0063 0.9291 

Liaoning 1.3766 1.3741 0.0019 1.4282 1.3442 0.0589 1.4928 1.5446 -0.0347 1.7410 

Jilin 0.9328 0.9325 0.0004 0.9951 1.0093 -0.0142 1.0031 1.0477 -0.0445 1.1994 

Heilongjiang 0.5891 0.5921 -0.0052 0.6210 0.5937 0.0440 0.6533 0.6673 -0.0215 0.7849 

Shanghai 3.5631 3.5114 0.0145 3.8370 3.8636 -0.0069 4.0874 4.0309 0.0138 4.0215 

Jiangsu 2.3332 2.3264 0.0029 2.4230 2.4105 0.0052 2.5108 2.4587 0.0207 2.4824 

Zhejiang 2.1312 2.1315 -0.0002 2.1339 2.2604 -0.0593 2.2352 2.2784 -0.0193 2.2766 

Anhui 0.4583 0.4587 -0.0008 0.4807 0.4570 0.0492 0.5093 0.4927 0.0326 0.4968 

Fujian 1.2987 1.2985 0.0002 1.3475 1.3451 0.0018 1.3897 1.3361 0.0386 1.3027 

Jiangxi 0.6510 0.6511 -0.0001 0.6724 0.6739 -0.0022 0.7028 0.6804 0.0318 0.6711 

Shandong 1.4882 1.4889 -0.0005 1.5514 1.5563 -0.0031 1.6114 1.5556 0.0346 1.5575 

Henan 0.5783 0.5780 0.0005 0.6038 0.5920 0.0196 0.6261 0.6312 -0.0081 0.6476 

Hubei 0.8619 0.8687 -0.0079 0.8907 0.9224 -0.0356 0.9010 0.9064 -0.0060 0.8571 

Hunan 0.6870 0.6870 0.0000 0.7222 0.7075 0.0205 0.7530 0.7612 -0.0110 0.7743 

Guangdong 2.6261 2.6265 -0.0001 2.7000 2.7044 -0.0016 2.8909 2.9614 -0.0244 2.9368 

Guangxi 0.3574 0.3576 -0.0006 0.3706 0.3692 0.0036 0.3812 0.3970 -0.0413 0.4505 

Hainan 0.2636 0.2641 -0.0019 0.2611 0.2734 -0.0472 0.2768 0.2730 0.0139 0.2851 

Sichuan 0.6065 0.6068 -0.0005 0.6266 0.6306 -0.0064 0.6456 0.6691 -0.0364 0.6766 

Guizhou 0.3089 0.3089 0.0002 0.3182 0.3157 0.0078 0.3310 0.3386 -0.0227 0.3437 

Yunnan 0.3094 0.3111 -0.0055 0.3119 0.2997 0.0389 0.3156 0.3227 -0.0224 0.3359 

Xizang 0.1206 0.1207 -0.0004 0.1162 0.1125 0.0318 0.1152 0.1125 0.0234 0.1125 

Shaanxi 0.6657 0.6649 0.0011 0.7090 0.6814 0.0389 0.7447 0.7702 -0.0342 0.6921 

Gansu 0.5786 0.5789 -0.0005 0.6076 0.5766 0.0511 0.6346 0.6518 -0.0272 0.6594 

Qinghai 0.4761 0.4763 -0.0004 0.4896 0.4783 0.0231 0.5105 0.5374 -0.0527 0.5895 

Ningxia 0.7761 0.7792 -0.0039 0.8054 0.8101 -0.0058 0.8593 0.8469 0.0145 0.8456 

Xinjiang 0.4726 0.4714 0.0026 0.4837 0.4655 0.0376 0.4990 0.4868 0.0244 0.4834 

We have high precision in the prediction using BP 

neural network (Relative errors are within 6%). This also 

shows that BP neural network has a strong function in 

learning, association, fault-tolerant and highly nonlinear 

function mapping with a good ability of generalization. 



The authors are grateful to Professor Huanming Zhang 

and Professor Erpo Lu of Anhui University of Financial 

and Economics (AUFE) for their valuable comments. The 

authors are also grateful to Zhongsheng Xu, a graduate 

student of department of statistics of AUFE for our earlier 

joint work. However, any remaining errors are the 

author’s responsibility. 

This work was supported in part by a grant from XYZ.


REFERENCES 

[1] Anselin L. Spatial Econometrics: Methods and Models. 

Dordrecht: Kluwer Academic, 1988. 

[2] Anselin L. “Space and applied econometrics”. Special 

Issue, Regional Science and Urban Economics, Vol. 22, 

No. 3, pp.509-536, 1992. 

[3] Anselin L, Florax R. “Small sample properties of tests for 

spatial dependence in regression models: some further 

results”. In New Directions in Spatial Econometrics, 

Edited by Anselin and Florax. Berlin: Springer-Verlag, 

1995, pp.21–74. 

[4] Anselin L. GeoDa 0.9.3 User's Guide. Center for Spatially 

Integrated Social Science, 2003. 

[5] Mohamad H. Hassoun. Fundamentals of Artificial Neural 

Networks. The MIT Press, 1995. 

[6] Saeed Moshiri and Norman Cameron. “Neural Network 

Versus Econometric Models in Forecasting Inflation”. 

Journal of Forecasting, No. 19, pp.201-217, 2000. 

[7] Paul Krugman. Geography and Trade. The MIT Press, 

1992. 

[8] G. Ellison and E. Glaeser. “Geographic concentration in 

US manufacturing industries: A dartboard approach”. 

Journal of Political Economy, Vol. 105, No. 5, 

pp.889–927, 1997. 


[9] Marius Brulhart. “Economic Geography, Industry location 

and trade: the evidence”. The World Economy, Vol. 21, 

No. 6, pp.775-801, 1998. 

[10] Antje. “Determinations of Geographical concentration 

patterns in central and eastern European countries”, 

unpublished. 

[11] Stuart S. Rosenthal. “The determinants of 

agglomenration”. Journal of Urban, Vol. 50, No. 2, 

pp.191-229, 2001. 

[12] Kim, S. “Expansion of Markets and the Geographic 

Distribution of Economic Activities: The Trends in U.S. 

Regional Manufacturing Structure, 1860-1987”. Quarterly 

Journal of Economics, Vol. 110, pp.881-908, 1995. 

Huayin Yu (1962-), professor in Department of Statistics of 

Anhui University of Financial and Economics, master tutor. 

Professor Huayin Yu majors in numerical calculation and data 

analysis. He is the corresponding author of this article. 

Weiping Gu (1985-), graduate student in Department of 

Statistics of Anhui University of Financial and Economics. He 

majors in numerical calculation and data analysis.


Application of Computer Technology in 

Efficiency Analysis of China Life Insurance 

Company 

Hongling Wu 

School of Economics/Anhui University of Technology, Maanshan, China 

Email: wuhongling76@163.com 

XiaoFei Zeng 

School of Economics/Anhui University of Technology, Maanshan, China 

Abstract—During the recent 100 years, the third 

technological revolution has promoted the development of 

computer technology dramatically, which thus has brought 

a great change in the economic society of human beings such 

as economics structure, employment direction, the form of 

international economic and the form of business. Besides, 

new concepts and ideas have been brought into the mode of 

production and life style of human beings. By the 

instrumentality of LINDO soft ware and SAS system, this 

research was conducted to evaluate the super efficiency, 

technical efficiency, pure technical efficiency and scale 

efficiency of life insurance companies of China in recent 

years by using the method of DEA and to analyze and find 

out the main and secondary factors that influenced the 

operational efficiency of insurance companies by using the 

measurement method. On this basis, it was concluded that 

efficiency of life insurance companies in our country could 

be enhanced by increasing underwriting quality, 

strengthening service awareness and optimizing business 

structure, etc. 

Index Terms—DEA model; efficiency; Life Insurance 

Company; insurance market; software LINDO; SAS system 


Application of computer technology and computer 

programs pervades every field of human life and 

production and also alters the development mode of 

human economic society. For example, in recent years, 

Lindo and Lingo are widely used in the fields of 

economic management and empirical analysis. Software 

of Lindo and Lingo which were developed by American 

Lindo System Company are computer programs to solve 

the problem of optimization. The basic function of Lindo 

is to solve problems of linear programming and quadratic 

programming. Furthermore, Lingo not only has all the 

functions of Lindo but also can solve the problem of 

nonlinear programming and Lingo can be used in the 

solution of linear and nonlinear equations. In the practical 

process of application, we find that the most significant 

Foundation item: Project of Non Fiction of Department of Education 

Anhui Province (Grant No.2010sk181). 


doi:10.4304/jcp.6.9.1832-1841 

feature of Lindo and Lingo is that an integer acting as a 

decision variable is available (integer programming) and 

the execution speed of these two kinds of software is 

much faster. 

In fact, Lingo is a modeling language of the problem of 

optimization, which includes many common functions of 

mathematics, economics and management and it is 

available for users’ fitting the optimization model and it 

can supply interfaces of other data files such as text files, 

excel, database files and so on and it is very convenient, 

fast and simple for inputting, solving and analyzing 

spacious problems of optimization. 

Maybe thanks to these characteristics, Lindo and 

Lingo’s solving programs of linear, nonlinear and integer 

programming are used to analyze maximizing profits and 

minimizing costs by the broad masses of theory 

researchers and practical managers and the programs can 

be used in various fields and have been proved to be 

playing a significant role in commercial, industry, 

research and government including affairs of production 

distribution, ingredient mixing, arrangement between 

production and personal affairs, inventory management 

etc and especially the field of finance and insurance. 

SAS is a large-scale integrated computer software 

system in which a set of computer programs worked 

together. The SAS users can make reasonable choices 

according to their demands. Since SAS is a kind of 

integrated system, it has complete functions of data 

access, data management, data analysis, data report and 

so on. This computer system was promoted by American 

SAS Software Research Institution in 1976, and now has 

been adopted by 120 countries and 30,000 departments in 

the world. SAS when running under WINDOWS 

environment can fully utilize the eminent graphical 

interface of WINDOWS operating system and good 

connectivity with other system and data, which brings a 

lot of convenience on program editing and data 

manipulation and management. The operation of SAS 

system is flexible and functional; furthermore, its 

language is a powerful programming designing language 

and it integrates a variety of high-level language features 

and flexible format. It is an integration of data


progressing and statistical analysis and also has a strong 

scalability. Therefore, the system is widely used by lots 

of general theory researchers and practical managers. 

Since 1980 when the domestic business was restored, 

the insurance industry of China has developed rapidly. 

The total volume of premium income was 1,600,000,000 

yuan in 1980, and an increase to 139,322,000,000 yuan 

(life insurance premium income was 87, 210,000,000 

yuan), and an increase to 978,400,000,000 yuan (life 

insurance premium income was 744, 738,000,000 yuan). 

The total volume of premium income in 2008 was 

increased by 6 times that in 1998 and life insurance 

premium income was increased 7.5 times. The capital of 

insurance increased from 260,409,000,000 yuan in 1999 

to 3,341,844,000,000 yuan in 2008, an increase of 11.8 

times. The total investment of insurance increased from 

89,142,000,000 yuan in 1999 to 2,246,522,000,000 yuan 

in 2008, an increase of 24.2 times. The insurance density 

increased from 110 yuan per person in 1999 to 736.74 

yuan per person in 2008, and the insurance density was 

increased. As a result, insurance is playing a more and 

more important role in the development of economics of 

the society. On aspect of attracting foreign investment: 

there were 18 overseas-funded enterprises of all insurance 

companies in China in 2000, and this number increased to 

89 in 2009. The increase of insurance companies 

especially on overseas-funded enterprises will lead to an 

increased competition in the insurance market. The 

financial strength, product development technology, 

development of industry approach and the business 

management level of foreign insurance companies are 

obviously better than those of domestic insurance 

companies. And because of the better salary, higher 

strategy on investment and management, a large number 

of excellent talents of management will be attracted by 

foreign insurance companies and this is a huge pressure 

for domestic insurance companies. Insurance companies 

of China always pay attention to underwriting income 

and scale of growth, and ignore claims service, efficiency 

and investing management and emphasize the premium 

income, thus under the macroeconomic environment that 

large numbers of foreign insurance companies flush into 

the market of insurance of China, the efficiency of 

insurance companies is becoming a focus in this field. 

With the linear, nonlinear and quadratic solution 

programs of Lindo and through the method of Date 

Envelopment Analysis (DAE), this research is conducted 

to evaluate values of super efficiency, comprehensive 

efficiency, pure efficiency and scale efficiency, and 

analyze changes of efficiency of different insurance 

companies, and establish relevant econometric model to 

analyze the key factors affecting the efficiency of 

insurance companies, and make appropriate comments 

and suggestions on enhancing the efficiency of insurance 

companies. Related researches in China only measured 

the efficiency of a certain value; however, this paper 

especially estimates the super efficiency of insurance 

companies to compare the pros and cons between 

insurance companies of which technology are effective 


and at last it analyzes the influencing factors on 

corresponding values of efficiency. 

II. INTRODUCTION EVALUTAION OF THE EFFICIENCY OF 

INSURANCE COMPANIES OF CHINA 

A. Sample Selection 

According to the principle of availability and 

comparability on data, 22 life insurance companies in 

2003-2008 were selected as the research samples. While 

newly established insurance companies that have been 

operated for 10 months and the premium income of 

which was in the forefront of all newly established 

insurance companies were also selected, and at last, 22 

companies in 2003, 25 companies in 2004, 29 companies 

in 2005, 33companies in 2006, 35 companies in 2007, 

and 39companies in 2008 were chosen. In the data of 

sampling companies, because PICC (People’s Life 

Insurance Company of China) was establish on 6.30. 

2003, and inherited relevant insurance business of CICL 

(China Life Insurance Company Limited), the increase of 

reserve fund of PICC in 2003 is presented as the product 

of total increase reserve fund of CICL*( premium 

income of PICC / premium income of CICL), and 

compensation duty and profit margin are instead 

according CICL, others are instead according to data of 

PICC. Relevant data come from the “China Insurance 

Yearbook” and relevant documents. 

B. Variable Selection and Comparison 

Efficiency is a reflection of result on microscopic 

behavior of enterprises, and it specially presents 

relationship between input and output or costs and 

benefits of insurance companies. According to the 

definitions given by Charnes and Coopers, the most 

important characteristics of input and output are that the 

increase of output and the decrease of input are the 

fundamental approach of pursuing aim and improving 

efficiency level of a production decision-making unit. 

There are three main methods to define input and output 

of a financial institution, namely intermediate approach, 

cost approach and added value. Financial institutions are 

generally calculated as a pure financial intermediary 

financial institution in the intermediate approach, that is, 

financial institution only earns the differences of interests 

through borrowing funds and transforming funds into 

assets. Obviously, this method is not proper for insurance 

companies. It is determined by the contribution to the 

income of financial institutions that whether a financial 

product can be acted as an input or output. If the proceed 

of the asset is greater than the opportunity cost of assets , 

or liabilities of the financing cost is less than the 

opportunity cost, then the product can be considered to be 

financial outputs; otherwise that is input. This approach is 

theoretically feasible; however, it is not practically 

available because it needs accurate data of product 

benefits and opportunity costs which are difficult to 

estimate. Berger& Humphery(1997) considered that 

added value was an appropriate method to measure the 

output in researches on the efficiency of insurance


companies, and this method can bring significant added 

value factors as output variables and the value reduction 

factors as input variables. There is basic agreement on the 

selection of input variables in China and the variables are 

mainly labor investment, capital investment (including 

forms of physical assets, paid-in capital, total capital and 

etc.) and operating costs (including forms of claim 

amounts and operating expenses etc.); however, there is a 

big difference on output variables. Zhao Xu (2003) 

adopted profits as the variable, while Hui min and Li Xin 

dan (2003) the asset profit margin and business income, 

Hou jing and Zhu lei (2004) the actual expected loss and 

investment income, Yao Shu jie et. (2005) premium 

income and investment income, Sun lin and Li Guang jin 

(2005) per capita profit and asset margin and He jing and 

Li Cunpu (2005) premium income. Insurance companies 

are different from general companies. It is clearly 

inappropriate to measure the operating efficiency by 

using a particular index of profits, profit margin or the 

premium income, such as premium income, and per 

capita premium is a quantitative measurement of 

operating results, and it is difficult to evaluate the income 

and risk status objectively only by considering the 

premium income. The reserve fund is a indicator of 

measuring business risk of insurance companies. The 

more adequate reserve fund the stronger ability of 

insurance companies resisting risks, and profits and profit 

margins are profitability indicators of insurance 

companies, and the higher profit margin, the greater 

development potential of companies. The amount of 

investment income presents the management and 

investment competence of companies. Modern insurance 

companies should not only pursue profits and investment 

income but also carry out their social duties, thus it is a 

key to measure the performance of insurance companies 

that considers premium income, profits, changes in the 

insurance reserve and investment income 

comprehensively. 

In summary, we adopt added value approach on input 

and output. Total fixed assets (equal to half of total fixed 

assets in early and the fixed assets at the end), total cost 

(including fees, commission costs and operating 

expenses), net amount of compensation payout (including 

direct insurance and reinsurance claims net of 

compensation ), and total number of employees are 

selected as the input indicators. Premium income (equal 

to the direct insurance and reinsurance premium income), 

total profits, the amount of reserve growth (the amount of 

preparation for the end of the year - the early mount of 

preparation) and the amount of investment income are 

selected as output indicators. 

C. Selection of Model 

Estimating efficiency of insurance companies using 

DEA linear model includes: 1.Measure the technical 

efficiency value by using C2R model, thus in order to 

compare the comprehensive efficiency of insurance 

companies; 2. Measure the pure efficiency value by using 

BC2 model and compare efficiency of insurance 

companies after removing the scale factor; 3. Measure the 

super efficiency value by using super efficiency model 


and thus in order to compare and distinguish the 

achievements and failures among insurance companies; 4. 

Measure the returns to scale changes by NIRS model, 

when the technical efficiency value in NIRS model is not 

equal to that in BC2 model that TENIRS ≠ TE BC2 ,it 

means the unit being evaluated is in the increase region of 

returns to scale, and the scale invalid is due to the small 

size and that means companies can increase efficiency 

through the expansion of scale. When TENIRS = TE 

BC2¸it means unit being evaluated is in the decrease 

region of returns to scale, and the scale invalid is due to 

the overlarge size of decision-making unit, and that 

means companies can increase efficiency through 

narrowing the scale. In this paper, origin and evolution of 

the models are omitted, and the returns to scale status of 

companies are not listed in this paper. 

(a) C 2 R Model 

T − T + 

ρ = θ − ε ( l s + l s )] 

min[ 1 2 

n 

∑ λi 

xi 

i= 

1 

− 

+ s = 

n 

∑ λi 

yi 

i= 

1 

+ 

+ s = 

s .t. 

θx 

; 

(b) BC2 Model 

T − T 

ρ = θ − ε ( l s + l s 

min[ 1 2 

n 

∑ λi 

xi 

i= 

1 

− 

+ s = 

n 

∑ λi 

yi 

i= 

1 

+ 

− s = 

n 

∑ λi 

= 1 

i= 

1 

s .t. 

k 

θy 

; 

+ 

k 

k 

)] 

θx 

; 

(c) “Super Efficiency” Model 

T − T 

ρ = θ − ε ( l s + l s 

min[ 1 2 

n 

∑λ i xi 

i= 

1, 

j≠ 

k 

− 

+ s = 

n 

∑λ i yi 

i= 

1, 

j≠ 

k 

+ 

− s = 

s .t. 

(d) NIRS Model 

T − T 

ρ = θ − ε ( l s + l s 

min[ 1 2 

n 

∑ λi 

xi 

i= 

1 

− 

+ s = 

n 

∑ λi 

yi 

i= 

1 

+ 

− s = 

n 

∑ λi 

≤ 1 

i= 

1 

+ − 

i S S ≥ 

s .t. 

y 

k 

+ 

k 

; 

)] 

θx 

; 

y 

k 

+ 

k 

; 

)] 

θx 

; 

λ , 

In the model, 

, 0; ε 

are non-archimedean 

infinitesimal, and i λ 

is the weight of DMU decision- 

y 

k 

;


x = x , x , , 

x 

( 2 

making unit, and i 1i i mi is the input 

variable of DMU decision-making unit, and 

y = y , y , , 

y 

( 1i 2i 

si 

) 

i 

is the output variable of f DMU 

+ − 

decision-making unit, and 

S , S 

is the slack variable, 

and S − 

is the m-dimensional column vector variable, and 

S + 

the s-dimensional column vector variable, and ρ is 

the ration of narrowing input. If 

DMU i decision-making unit is DAE effective, and if 

ρ = 1 

DMU 

and there is non-zero value in S+, S- , i 

) 

ρ = 1 

, = = 0 

− + 

S S 

decision-making unit is DAE weekly effective, and 

ρ ≺ 1 DMU 

if , i decision-making unit is DAE invalid. 

ρ is the index of relative efficiency, and in 

T T 

l1 = (1,1,...,1) 1× m, l2 

= (1,1,...,1) 1× 

s , s is the output 

variable and m is the input variable. 

D. Selection and Application of Software 

(a) Characteristics and Application of LINDO 

Software. 

LINDO was developed by the Linnus Schrage and is a 

kind of software package that specially used to solve the 

mathematical programming problem. The software 

package contained a complete series since its inception 

including LINDO, GINO, LINGO and LINGO NL. As 

mentioned above, LINDO is mainly used to solve linear 

programming, integer programming and quadratic 

programming problems, and GINO can be used to solve 

nonlinear programming problem, and to solve linear and 

nonlinear equations, inequalities and the roots of 

algebraic equations, besides, GINO includes certain 

finance, probability and trigonometric functions and a 

variety of common mathematical functions which is 

available for user to invoke when creating the problem 

model, and LINGO can be used in solving linear and 

integer programming problem, and LINGO NL can be 

used for solving linear, nonlinear and integer 

programming problems. 

Because LINDO’s high speed on implementation and 

the convenience on inputting, solving and analyzing 

mathematical programming problems, LINDO is widely 

used in the fields of mathematics, scientific research and 

industry and LINDO has been developed several 

versions. Current versions of LINDO are powerful and 

are mainly used in solving linear, quadratic and integer 

programming problems. Interactive environment is 

available for beginners to set up and solve the 

optimization problem easily. On the other hand, it can 

also be used to solve some complex quadratic integer 

programming problems practically. Like on the largescale 

machine, it can be used to solve large-scale 

complex problems with more than 50,000 constraints and 

2,000,000,000 variables. Using LINDO software, this 

paper gets the value of DEA value of several insurance 

companies through selecting input and output variables. 


Entering the following procedure in LINDO6.1 

window, the technical efficiency value of the insurance 

company Pacific-Antai Life Insurance Company Limited 

(PALIC) is obtained, and the procedures of other 

efficiency values of insurance companies are similar, and 

they are omitted. Here just presents the following 

procedure in LINDO6.1 window: 

MINX26 

ST 

2)149987.00X1+54876.91X2+34618.16X3+18820.71X4 

+17674.05X5+6607.45X6+1245.22X7+4758.05X8+597. 

6X9+617.36X10+188.26X11+174.88X12+50.32X13+65 

0.98X14+75.69X15+333.9X16+213.04X17+116.74X18+ 

1206.1X19+120.13X20+83.48X21+3.66X22+53.59X23+ 

292.22X24+21.56X25>21.56 

3)3157X1+2384.7X2-1602.29X3-360.57X4+110.68X5- 

570.52X6-126.35X7-37.93X8-1.5X9-72.95X10- 

40.63X11-54.35X12-10.39X13-131.33X14-22.85X15- 

60.73X16-80.38X17-54.61X18-126.54X19-40.52X20- 

56.45X21-26.43X22-72.15X23+25.23X24-19.48X25>- 

19.48 

4)93294X1+34412.44X2+25341.38X3+14419.62X4+126 

50.46X5+5774.42X6+1213.97X7+3348.81X8+321.27X9 

+360.1X10+131.36X11+78.43X12+34.56X13+282.75X1 

4+43.62X15+281.293X16+113.75X17+82.08X18+1035. 

6X19+132.88X20+64.5X21+2.2X22+42.76X23+291.73 

X24+7.54X25>7.54 

5)3669X1+2848.06X2+1288.79X3+453.59X4+710.95X5 

+164.01X6+18.73X7+369.49X8+28.93X9+16.74X10+11 

.41X11+5.04X12+1.65X13- 

5.88X14+3.28X15+5.14X16+6.33X17+5.21X18+48.82X 

19+3.59X20+2.56X21+1.96X22- 

0.6X23+14.88X24+3.51X25>3.51 

6)12773.5X1+4331.735X2+1418.055X3+929.485X4+57 

5.3X5+240.775X6+54.58X7+127.4X8+25.27X9+21.57X 

10+10.79X11+10.43X12+0.6X13+13.5X14+5.955X15+7 

.622X16+22.89X17+14.695X18+36.97X19+17.99X20+1 

4.945X21+5.095X22+12.16X23+10.81X24+2.845X25- 

2.845X26


by COROLINA STATE UNIVERSICIY in 1966. The 

SAS INSTITUTE INC was established in 1976 and since 

then began the work of maintenance, development, 

marketing and training of SAS system. During the period, 

SAS had gone through many versions, and after several 

years’ improvement and development, SAS system has 

been valued as the international standard statistical 

analysis software and is widely used in various fields. 

SAS is a modular, integrated large-scale application 

software system. It consists of dozens of specialized 

modules and its functions include data access, data 

storage and management, application development, 

graphics processing, data analysis, report preparation, 

operation research approach, econometrics and 

forecasting etc. 

On one hand, SAS has characteristics of powerful 

functions and its statistical methods are abundant and 

new. SAS provides not only basic statistics calculation 

but also variance analysis, correlation and regression 

analysis and multivariate analysis of various statistical 

analysis processes of various experimental designs, and 

its technology of analysis is advanced and reliable. The 

analysis method is realized through the process call. 

Many processes also provide a variety of algorithms and 

options. For example, in the analysis of variance of 

multiple comparisons, more than 10 kinds of methods 

including LSD, DUNCAN, and TUKEY are provided. A 

choice of 9 various methods (such as STEPWISE, 

BACKWARD, FORWARD, RSQUARE etc) is provided 

in regression analysis. In the regression model, users can 

choose whether to include the intercept and can also predesignate 

some independent variable word groups 

(SUBSET) in the model. For the intermediate results, 

those can be all output, not output or selecting output and 

can also be stored to a file for further analysis procedure 

call. On the other hand, SAS is easy to use and flexible to 

operate. It yields data sets through a common data 

(DATA) and later complete various data analysis through 

different procedure calls. Its programming statements are 

concise and short, and generally a number of complex 

operations with satisfactory results can be completed by a 

only a few statements. Results are presented by concise 

English prompt, and statistical terminology is standard 

and easily understand, and it is available for preliminary 

Company Number 

TABLE1. EFFICIENCY OF VARIOUS DIFFERENT INSURANCE COMPANIES 1 

Super 

efficiency 

English and statistical basis. Users just tell SAS what to 

do without telling how to do. Design of SAS make users 

do not have to tell SAS something that can be “guessed” 

by SAS ( that is without setting), and SAS also can 

correct some minor errors automatically. Besides, SAS 

can give reasons and correction method of running-time 

errors. As a result, SAS organically combines the 

scientific, precise and accurate of statistics and the feature 

of easily use together, which greatly facilitates the users. 

In SAS9.0 window, entering the following procedure, 

main factors influencing technical efficiency value of 

insurance companies can be obtained. Procedure as 

followed: data A2;set A1;run;proc reg; model Y1=X1- 

X8/selection=stepwise sls=0.05 sle=0.2 r;run; A1: data 

files imported in SAS software, including the 8 

assumptive influencing factors and specific values of 

various efficiency, and other efficiency values are 

regression similar, and they are omitted in this paper. 

E. Efficiency Value and Evaluations of Company 

Table 1 shows us the super efficiency technical 

efficiency, pure technical efficiency and scale efficiency 

of different insurance companies. The top five insurance 

companies of super efficiency are Zhaoshangxinruo 

company, Ruitai life insurance company, 

Zhongbaokanglian v, PICC, Yangguang life insurance 

company, and the last five companies are Changcheng 

life insurance company, Hezhong life insurance 

company, Haier –NewYork company, Haikang life 

insurance company and Guangdianrisheng company, 

short term and long term companies are both include, and 

at the forefront only PICC is large scale company, while 

others are all small companies. The top five insurance 

companies of technical companies are Zhongbaokanglian 

company, Yangguang life insurance company, 

Yingdataihe company, Xingfu life insurance company 

and PICC, and the last five companies are Yingzhong life 

insurance company, Hezhong life insurance company, 

Haier-NewYork company, and Guangdianrensheng 

company. Pure technical efficiency and scale efficiency 

are similar, and this is because total value of technical 

efficiency is determined by pure technical efficiency and 

scale efficiency. 

Technical 

efficiency 

Average value 

Pure 

technical 

efficiency 

Scale 

efficiency 

Operating 

time: 

year 

China Life Insurance 1 1.8363 0.9771 1.0000 0.9771 2003-2008 

Ping An Life Insurance 2 

1.2635 0.9468 1.0000 0.9468 

2003-2008 

Pacific Life Insurance 3 1.0985 0.9571 1.0000 0.9571 2003-2008 

Xinhua Life Insurance 4 1.3150 0.9307 1.0000 0.9307 2003-2008 

Taikang Life Insurance 5 1.1445 0.9576 0.9933 0.9628 2003-2008 

Tai Ping Life Insurance 6 0.9604 0.7162 0.7979 0.8889 2003-2008 

Sino-Life Insurance 7 0.9095 0.6306 0.7971 0.7546 2003-2008 

AIA 8 1.0822 0.7964 0.9755 0.8200 2003-2008 



Zhonghong Life Insurance 9 1.4103 0.7883 0.8849 0.8417 2003-2008 

Pacific-Aetna Life Insurance 10 0.7050 0.5844 0.6307 0.8633 2003-2008 

Allianz Dazhong Life Insurance 11 0.5566 0.5566 0.8155 0.6497 2003-2008 

AXA-Minmetals Assurance 12 0.5001 0.4769 0.5607 0.7727 2003-2008 

China Life CMG 13 1.9438 1.0000 1.0000 1.0000 2003-2008 

Prudential Life Insurance 14 0.6829 0.5793 0.6557 0.7623 2003-2008 

John Hancock Tianan Life 

Insurance 

15 

0.5597 0.4378 0.8528 0.5012 

2003-2008 

Generali China Life Insurance 16 1.3654 0.9032 0.9162 0.9738 2003-2008 

Sun Life Everbright 17 0.5754 0.5163 0.6188 0.8224 2003-2008 

Haier New York Life 18 0.2565 0.2565 0.4327 0.5969 2003-2008 

Minsheng Life Insurance 19 0.6306 0.5326 0.6680 0.8020 2003-2008 

ING Insurance Company 20 0.8471 0.4018 0.5179 0.7364 2003-2008 

Sino-British Life Insurance 21 0.3992 0.3702 0.6227 0.5745 2003-2008 

Nissay-SVA Life Insurance 

Company 

22 

0.1805 0.1805 0.9527 0.1915 

2003-2008 

AEGON-CNOOC Insurance 23 0.2391 0.2391 0.4533 0.5286 2004-2008 

Heng An Standard Life 

Insurance 

24 

0.5092 0.5092 0.7096 0.7587 

2004-2008 

CIGNA and CMC Life Insurance 25 3.4377 0.6899 0.8611 0.7781 2004-2008 

China. MetLife 26 0.4016 0.3857 0.4340 0.6718 2005-2008 

Greatwall Life Insurance 27 0.3815 0.3815 0.6035 0.6295 2005-2008 

Cathay Life Insurance 28 0.4200 0.4200 0.6235 0.6692 2005-2008 

Winterthur Life 29 3.3269 0.7500 1.0000 0.7500 2006-2008 

United Metlife Insurance 30 0.6844 0.6492 0.7818 0.7918 2006-2008 

Union Life Insurance 31 0.2996 0.2996 0.4014 0.7529 2006-2008 

Huatai Life Insurance 32 0.6042 0.4977 0.7045 0.6348 2006-2008 

Jiahe Life Insurance 33 1.1850 0.6497 0.9100 0.6976 2007-2008 

Dragon Life Insurance 34 1.3203 0.6618 0.8649 0.8818 2007-2008 

Huaxia Life Insurance 35 0.6755 0.6755 0.8041 0.7096 2007-2008 

Sinatay Life Insurance 36 0.5879 0.5879 0.7880 0.7461 2008 

YingDaTaiHe Life Insurance 37 1.5471 1.0000 1.0000 1.0000 2008 

Happy Insurance 38 1.0442 1.0000 1.0000 1.0000 2008 

Sunshine Life Iunsurance 39 1.5893 1.0000 1.0000 1.0000 2008 

1 Data from the “China insurance Yearbook” from 2003 to 2008 and other relevant documents 

III. ANALYSIS OF FACORSIY OF INFLUENCING EFFICIENCY 

OF INSURANCE COMPANY 

A. Theory Analysis of Factors Influencing Efficiency of 

Insurance Company 

According to current domestic research on the 

efficiency of insurance companies, it is considered that 

size, ownership structure, human capital, proprietorship 

structure, operating time and business scope of insurance 

companies will affect the efficiency of insurance 

companies. Based on the domestic researches’ 

conclusions, this paper assumed the following factors 

influencing efficiency of insurance companies and did 

tests accordingly. 

a) X1Factor of the Capacity of Insurance Services: 

A key function of contracted business of insurance 

company is risking-sharing, and when insurers suffer 


losses, timely payments by insurance companies is one of 

the keys that insurance companies can get businesses. 

Therefore, the loss ratio is measured as an index of 

service competency of insurance companies. Low loss 

ratio will not only improve operational efficiency, but 

rather be in a disadvantage situation because of lack of 

appeal in the fierce competition. Insurance companies 

always pay attention to premium income but ignore the 

claims, and when the overall loss ratio is low, it is 

assumed that the higher the loss ratio is, the more 

premium and the better operating efficiency the insurance 

companies get. Loss ratio = current total amount of 

claim/current total amount of premium. 

b) X 2 Factor of Asset Scale: 

Insurance companies are enterprises operating risk 

business, and larger-scale insurance companies have 

higher ability of acceptance of risk, and small-scale 

insurance companies are disadvantage on both credibility


and popularity, and insurance industry in China is in the 

stage that development and improvement of 

competitiveness are relying on the growth of scale, so 

natural logarithm of asset amount of an insurance 

company is selected to present the scale of an insurance 

company. 

c) X 3Factor of Human Capital: 

Human resources play an important role in the 

development and competition of modern enterprise. An 

insurance company is composed of stuffs with different 

levels. It is generally believed that higher educated 

employees with stronger professional knowledge are 

good for the development of an insurance company. 

Therefore the ratio of well-educated stuff (number of 

employees who are undergraduate and over / total 

number of employees) is adopted as the human capital. 

d) X4 Factor of Productivity Per Labor Unit: 

Premium income per person is to measure the 

operating efficiency of an insurance company through the 

production efficiency per labor unit. Higher production 

efficiency per labor unit can produce better benefits and 

efficiency in an insurance company. Higher premium 

income per capita induces higher operating efficiency in a 

insurance company and vice versa. Premium income per 

capita = total premium income / total number of 

employees. 

e) X5 Factor of Operating Time of a Company: 

It takes a long time to manifest the operating 

performance of an insurance company, and companies of 

short operating time are at a disadvantage on business 

network, reputation and scale while companies of long 

operating time are at an advantage on business network, 

reputation and scale. Therefore, it is assumed that longer 

operating time, better operating efficiency. 

f) X6 Factor of Insurance Type: 

Business of life insurance companies can be divided 

into group insurance and individual insurance. Group 

insurance is better than individual insurance at terms of 

size and quality, so a higher proportion of individual 

insurance in the premium income of an insurance means 

lower operating efficiency and vice versa. The type of an 

insurance business is account to the individual 

proportion, and an individual proportion = an individual 

premium/ total amount of premium income. 

g) X7 Factor of Underwriting Quality: 

For an insurance company, more surrender will bring 

negative effects on normal operation and development. 

High surrender ratio stands for low efficiency in an 

insurance company. Surrender ratio = amount of 

surrender/ total amount of premium income. 

h) X8 Factor of Competence of Investment and 

Management: 

When an insurance company develops to a certain 

stage, the underwriting profit generally is low because 

competition increases. The insurance company enhances 

its competitiveness and developing ability mainly relying 

on high investment rate and good risk management. The 

insurance market is developing gradually, and the 

investment scope is expanding, and the investment risk is 

also expanding, and the rate of return on investment 


(ROI) influences greatly on the operating efficiency in an 

insurance company. Therefore, in this paper it is assumed 

that the higher rate of ORI the higher efficiency of an 

insurance company. ROI = net investment income/ total 

amount asset of the insurance company. 

B. Establishment of model 

The macroeconomic environment that influences the 

super efficiency, technical efficiency, pure technical 

efficiency, scale efficiency value of an insurance 

company includes insurance regulatory policy, 

macroeconomic conditions, as well as the operation 

situation of the enterprise itself such as ROI and 

premiums per capita. The enterprise can not alter the 

external factors like macroeconomic policy; however, the 

only changes that can be done by enterprise are to 

strengthen their management, improve operational 

efficiency. Coelli et al. (1998) proposed the famous “twostage” 

method, and its main thought is first calculates the 

efficiency value by DEA model, and then selecting 

appropriate environment variables to do regression 

analysis and then make sure of the factors influencing 

efficiency. In China, least squares regression and Tobit 

models are always used to estimate the influencing 

factors, and because the efficiency value of former 

model has a restrict range between 0-1, parameter 

estimation is biased and non-consistent. The technical 

efficiency value determined by DEA method can not 

distinguish the advantages and disadvantages among 

companies, and efficiency values of effective companies 

are all 1, and the restrict range of efficiency value which 

between 0-1 make the Tobit model no longer available in 

this situation. According to Hardwicketal’s method 

(2003), this paper did regression analysis on super 

efficiency and when did regression on technical 

efficiency, pure technical efficiency and scale efficiency 

value, convertible regression of efficiency value is 

adopted. Transform form is as below: 

Yi = Ln( 

TEi 

/ 1− 

TEi 

) 

TE i is technical efficiency, pure technical efficiency 

and scale efficiency value calculated in the DEA model, 

and the rang is 0-1, and Ln is natural logarithm, and in 

order to convert conveniently, all the efficiency value 

minus 0.0005, and regression model is established: 

Yi = β 0 + β1X 

1 + β2 

X 2 + + 

βn 

X n + εi 

After that, we can adopt the least squares method to do 

stepwise regression on dependent variable Y at different 

influencing factors and investigate the factors that have 

significant effect on the efficiency of insurance industry. 

C. Empirical Results Analysis 

From the regression results, we can find that the 

factors affecting the efficiency of an insurance company 

are loss ratio, human capital, premium income per capita 

and proportion of individual insurance and operating 

time. Loss ratio and proportion of individual insurance 

show significant difference at the level of 5% in t-test and 

others show significant difference at the level of 1% in ttest. 

Loss ratio, surrender rate, premium income per


capita and company scale plays a significant effect on 

influencing efficiency of an insurance company. Of these 

four factors, surrender rate shows significant difference at 

the level of 5% in t-test and the other three show 

significant difference at the level of 1%. Details are 

shown in Table 2. 

TABLE 2. REGRESSION RESULTS OF FACTORS INFLUENCING EFFICIENCY OF LIFE INSURANCE COMPANIES OF CHINA 

Variable Model 1 Model 2 Model 3 

Constant -2.0209(4.18)** 

Loss ratio 7.4912(4.33)** 7.3445(5.86)** 

Human capital 5.8667(6.89)*** 

Surrender ratio 6.2194(8.67)*** 

Company scale 0.7632(41.7)*** 0.2844(44.82)*** 

Premium per capita 0.2209(11.39)*** 0.1706(8.12)*** 

Individual insurance 1 

-0.3144(4.46)** 

Operating time 0.1923(18.34)*** 

Adjusted R 0.4795 0.1873 0.6231 

F statistics 32.79*** 41.70*** 73.99*** 

Number of observation 183 183 183 

***、**means difference at level of 1 and 5%,T statistics are in brackets. 

1 

Propotion of individual insurance 

Models are obtained according the results of regression these will be improved. There is no doubt that premium 

Y 

( 1 , Y2, 

Y3 

are technical efficiency, pure technical 

efficiency, scale efficiency value separately) 

Y1 

= 7. 

4912X 

1 + 5. 

8667X 

3 + 

0. 

2209X 

Model1 

4 + 0. 

1923X 

5 − 0. 

3144X 

6 

per capita is the index of output per unit, and low output 

unit will never bring high operational efficiency, thus as 

premium per capita increases 1%, technical efficiency 

increases 0.2209%. Compared with other indexes, the 

impact of premium per capita is less, and it is relative to 

with that insurance marketing of China only seeks the 

Model 2 

expanding in scale a few years ago, and it means that it is 

Y 2 = −2. 

0209 + 0. 

7632X 

2 

Model3 

Y3 

= 7. 

3445X 

1 + 0. 

2844X 

2 

not feasible that insurance companies of China increases 

efficiency by expanding scale. Proportion of individual 

insurance and technical efficiency are negatively 

correlated, and 1% increase in proportion of individual 

+ 0. 

1706X 

4 + 6. 

2194X 

7 

In model 1, the loss ratio, human capital, premium 

income per capita, company operating time is positively 

correlated with technical efficiency, and as loss ratio 

increased by 1 percentage, technical efficiency increases 

7.4912 percents, and this shows that insurance companies 

of China should strengthen management on claims and 

improve the function on claims and security. As human 

capital increases 1%, technical efficiency increases 

5.8667%. Human capital generally does not show 

significant difference in the previous researches and thus 

it is often removed. However, this paper shows that 

human capital begins to play an important role on 

efficiency of insurance companies according to the last 6 

years data, and it is inseparable from the practical 

environment that Chinese insurance industry has fully 

opened to foreign countries and insurance companies has 

enhanced competition and talents have began to play a 

great role on the development of insurance companies 

since 2003. Premiums per capita and established time 

have a positive impact on the technical efficiency of 

insurance companies, and this is consistent with previous 

analysis, and it means that the established time of the 

insurance company has a certain impact on efficiency of 

the insurance company, and this is mainly related to with 

the marketing channels of the insurance company. Newly 

established insurance companies are poor at brand 

influence and marketing networks, and as time goes on, 

insurance and 0.3144% decrease in technical efficiency 

and this is consistent with the current situation in China. 

China’s insurance market is relatively underdevelopment, 

and compared with individual insurance, group insurance 

has superiority on scale and quality. If the individual 

insurance proportion is high in the business of insurance 

companies, operating efficiency will be relatively low. 

Therefore, it is a better choice to increase the proportion 

of group insurance in the business of insurance 

companies. 

In model 2, scale of insurance companies affects the 

pure technical efficiency of insurance companies. If the 

scale of the company increases 1%, pure technical 

efficiency increases 0.7632%. Overall, compares with 

other factors, scale of companies has a lower influence on 

efficiency of insurance companies. 

In model 3, loss ratio, surrender rate, premium income 

per capita and company scale affect the scale efficiency 

of insurance companies. These four factors are positive 

correlation. Loss ratio increases 1%, and the scale 

efficiency increases 7.3445%. This is basically consistent 

with model 1. This means that loss ratio influences 

technical efficiency of the company through influencing 

scale efficiency. Increase of loss ratio contributes to the 

increase of scale, thus increase scale efficiency, and the 

increasing scale efficiency can help improving the 

technical efficiency. 1% decrease of surrender rate 

contributes 6.2194% increase of scale efficiency and 



which may be related to China's current surrender terms 

and that insured is in a weak position compares with 

insurance company. In other words, the insurance 

company constrains the insured through a relatively harsh 

terms of surrender, so the insured once surrenders, the 

insurance companies benefit. A 1 % increase of company 

scale and the premium per capita followed a crease of 

0.2844%, 0.1706% scale efficiency separately. As can be 

seen, it is very weak that insurance companies enhance 

technical efficiency through increasing scale efficiency. 

Overall, factors influencing technical efficiency of 

insurance companies include loss ratio, surrender rate, 

premiums per capita, company scale, company operating 

time and proportion of personal insurance in companies’ 

business. On the aspect of impact that ROI affects the 

operating efficiency, because significant issues in the 

regression analysis have been removed, and this is the 

disadvantage of the paper and it may be due to that when 

determining the DEA efficiency, return on investment is 

taken as the output item and thus make DEA efficiency 

highly related to ROI. 

IV. CONCLUSION AND SUGGESTION 

Statistical economics is playing a more and more 

important role in the modern society of economics and 

life. In order to grasp the pulse of the economy, 

government and enterprise collect and release large 

amounts of digital information every year, and in order to 

constitute the developing plan of society economy, 

several order differential equations, hundreds of 

simultaneous linear equations and solving large-scale 

matrix are processed, and it is inconceivable without the 

help of computer. The actual shapes of various curves in 

economics mainly come from the analysis of statistical 

data and knowledge of database, procedures, and systems 

etc. are needed in computer application science. It is not 

only because that economics gives us inspires to 

understand the complex economical society, but also that 

it makes the market economy go through smoothly and 

get better control. To accurately grasp the subtleties of 

economic and social development, we consider that we 

should first combine mathematics, economics and 

computer organically. Since we are familiar with our 

research object: the efficiency of life insurance 

companies, we should have a mathematical basis 

meanwhile we should also have to grasp the application 

of computer programs in some extent: LINDO software 

and SAS software, so we combine the two and carry out 

the thought of this article. 

In this paper, by the means of the software of LINDO 

and SAS, efficiency values of different insurance 

companies are fast calculated and it supplies a scientific 

tool for comparing business efficiency of each insurance 

company and it promotes the application of LINDO and 

SAS in the field of economic management. The 

appearance of LINDO and SAS well solve the 

shortcoming of lack of tools in the field of economic 

analysis. Especially in the problem of linear 

programming, LINDO software is simple, fast, and 


convenient to operate, and it is suitable for the users in 

the field of economic analysis. 

A. Improve the Underwriting Quality of Insurance 

Companies 

To the newly established insurance companies, it is 

pivotal that how to improve their popularity and get 

customers’ recognition, and only with the appropriate 

market scale, newly established companies can compete 

with the old famous state insurance companies; 

otherwise, everything is impossible. For the large scale 

insurance companies like PICC, it is an important issue to 

handle the relationship between scale and quality 

correctly. At the time of expanding scale, underwriting 

quality should be improved, and avoid ignoring 

underwrite quality because of the expansion of scale. 

B. Greatly Improve Loss Ratio and Service Quality of 

Insurance 

The appropriate loss ratio is a key indicator to attract 

interests of insured and is a basic function of insurance 

companies. Some insurance companies attempt to 

improve profits by deliberately suppressing loss ratio, set 

barriers of coverage and claims; however, these measures 

will not enhance the profitability, operating efficiency 

and competitiveness of insurance companies but will 

make themselves in a vicious circle and make themselves 

in a disadvantage situation in the competition of 

insurance. Appropriately increasing loss ratio will 

stimulate the enthusiasm of insured and also is 

conductive to the growth and maturity of the insurance 

market, and insurance companies also can deconcentrate 

the risk through the advantage of scale. 

C. Optimize Insurance Structure and Speed Up the 

Development of Insurance Products. 

In the current insurance market of China, group 

insurance is better than individual insurance on scale and 

quality; therefore, insurance companies should pay 

attention on promoting group insurance, increase the 

proportion of group insurance, and insurance companies 

should speed up the development of insurance products 

depending on the development of group insurance and 

especially focus on the market of individual insurance 

and promote various kinds of products to meet different 

individual need. As the development of economy, the 

individual insurance market will gradually develop and 

mature, and insurance companies will have a great space 

in developing individual insurance market. 

D. Improve the Investment and Management Capabilities 

of Insurance Companies and ROI, and Provide Supports 

For the Development of Insurance Companies. 

Modern insurance companies obtain a large number of 

insurance funds relying on the insurance market and then 

get high returns depending on the excellent operating on 

investment and management and thus it will support the 

development of insurance business. The traditional 

situation of getting profits relying on the underwriting are 

not existed anymore, and the fierce competition make the 

underwriting profits of insurance companies become


smaller and smaller, or even loss, so it is the inevitable 

choice for the insurance companies to improve the ROI. 

Facing the situation that foreign large insurance 

companies have joined the Chinese insurance market and 

the competition is increasing day by day, it is greatly 

effective to improve the insurance companies’ efficiency 

through strengthening the claim service, increasing the 

loss ratio, improving insurance structure, especially 

raising the proportion of group insurance. 

REFERENCES 

[1] Wei Quanling. DEA Data Packet Analysis[M].Beijing: 

Science Press.2004 

[2] Hou Jin, Zhu Lie.Non-life insurance empirical analysis of 

operating efficiency of China Insurance Company 

[J].Nakai Economic Studies.2004(4),108_112 

[3] Li Kecheng. Empirical analysis of operating efficiency of 

China Life Insurance Company[J].Insurance 

Studies.2005(2),37_41 

[4] Yao Shujie, Feng Genfu, Han Zhongwei. The Empirical 

Analysis of Efficiency of China's Insurance Industry 

[J].Economic Research Joural.2005(7),56_65 


[5] Hu Ying, Ye Yugang.An Empirical Study on the Factors 

Influencing the Efficiency of Insurers in China[J].Journal 

of Jinan University.2008(4),28_34 

[6] Yue Chaolong. SAS system and economic statistical 

analysis[M].Hefei: University of Science and Technology 

of China Press.2003 

Hongling WU, female, was born on December 31th, 1976 in 

Dingyuan Anhui Province. Education background: Master of 

Economics. Gain the Master of Quantitative Economics in 2008 

at Anhui University of Technology. Now, she applies herself to 

finance and transnational corporations. Work Experience: In 

1999-2000, Business College of East China University of 

Metallurgy. Teaching Assistant; In 2001-2010, School of 

Economics of Anhui University of Technology. Lecturer 

Xiaofei Zeng, man, was born on August in 1979 in Yinbin 

Sichuan Province. 

Education background: Master of Economics 

Major: Quantitative Economics


A Bayesian Belief Net Model for Evaluating 

Organizational Safety Risks 

Li Song 

School of Economy and Management, Anhui University of Science and Technology, Anhui Huainan, China 

Email: lilysong@ustc.edu 

Li Yang 

School of Economy and Management, Anhui University of Science and Technology, Anhui Huainan, China 

E-mail: yangli081003@163.com 

Jing Han 

Huainan Vocational & Technical College, Huainan, Anhui, China 

E-mail: hanjing623@163.com 

Abstract—A Bayesian Belief Network (BBN) is a valuable 

tool to represent the causal relationships that exist in a given 

set of variables. This paper presents a methodology for 

organizational risk analysis for safety management. 

Learning a BBN from data is a difficult and 

resource-consuming task, we presents the implementation of 

a greedy algorithm that automatically constructs a BBN 

from a dataset of cases obtained. The resulting BBN reflect 

installation specific factors respect to organizational factors 

and show the dependencies that exist among key variables 

that are associated to the trip generation process. 

Index Terms—Bayesian Belief Network; organizational risk 

factors; reliability analysis 


Complex socio-technical systems are comprised of 

physical system and human system. The performance of a 

complex socio-technical system is dependent on the 

interaction of technical, human, social, organizational, 

managerial and environmental factors [1]. Safety 

performances often depend on complex and distributed 

interactions between human operators and technical 

systems. In the present dynamic society, a very fast pace 

of change of technology is found at the operative level of 

society within many domains, and the rapid development 

of information and communication technology have 

leaded to a high degree of integration and coupling of 

systems and the effects of a single decision can have 

dramatic effects that propagate rapidly and widely 

through the global society. Living in a very aggressive 

and competitive environment, companies today would 

focus the incentives of decision makers on short term 

financial and survival criteria rather than long term 

criteria concerning safety. It is widely recognized that 


doi:10.4304/jcp.6.9.1842-1846 

Jinkai Li 

Guanghua Management School,Peking University; 

E-mail: lijinkai@sina.com 

accidents in which ‘human error’ plays a part are often 

not solely attributable to errors made by an operator but 

have deeper causes, arising from the behavior of many 

others within the organizational context of a system[2]. 

Investigations of accidents in complex systems have 

shown that events attributed to human error and blamed 

on an operator have systemic causes, such as procedural 

or organizational weaknesses. Many such failures and 

accidents do not have a simple explanation, particularly 

those that have significant contributions from human and 

organizational behaviors. Increasing interest over the past 

two decades in causal modeling of organizational safety 

behavior is in part motivated by the desire to understand 

the deeper more fundamental causes of accidents and 

incidents. Reason [8] describes the gradual relaxation of 

safety alertness following a period of safe operation, 

followed by increased alertness after an accident as 

‘currents in the safety space’. Rusmussen [3] stresses 

environmental pressure will cause the operation of a 

system to migrate towards the boundary of safety. To 

analyze the risk of accidents and to improve safety, 

organizational risk factors need to be understood and 

evaluated. Several frameworks for analyzing the 

organizational context of accidents have been proposed, 

but without the capability to assess risks numerically. 

Event trees are usually used to model accident process, 

while organizational weaknesses have only an indirect 

effect on the accident and are therefore not readily 

represented as events. We outline an alternative method, 

using Bayesian Networks to model accidents, with 

explicit representation of both events and root causes. 

II. APPLICATIONS OF BBN IN ORGANIZATIONAL RISK 

ANALYSIS


A. Bayesian Belief Network 

Bayesian probability has existed for many years. 

Implementation algorithms and software tools have 

become available in recent years. Bayesian Networks are 

a network-based framework for representing and 

analyzing models involving uncertainty. They handle 

uncertainty in a mathematically rigorous yet efficient and 

simple way compared with other knowledge-based 

systems. Due to its power to deal with the soft data in 

reliability, it has stimulated a strong interest [4]. A 

Bayesian Belief Network consists of a set of variables 

(causes and effects) and a set of directed edges between 

variables (paths of influence). Each variable has a finite 

set of mutually exclusive states. The variables together 

with the directed edges form a directed acyclic graph 

(DAG). Conditional probabilities carry the strength of the 

links between the causes and their potential effects. For 

example for a given state of a variable A with parents B1, 

…, Bn, we have the conditional probability of the state 

(A) occurring given the state of the contributing parent 

nodes: P(A|B1, …, Bn). Bayes' theorem in the subjective 

theory of probability is at the core of the inference engine 

of BBNs. In the definition of Bayesian Belief Networks, 

the DAG restriction is critical. Feedback cycles are 

difficult to model quantitatively and no calculus has been 

developed for causal networks that can cope with 

feedback cycles in a reasonably general way. 

Figure 1. Simple BBN 

B. Organizational risks factors 

The empirical studies of organizational safety 

performance have revealed a number of organizational 

factors in developing a predictive causal model of 

organizational risks. In Bella’s view, large organizations 

are complex systems, which adaptively change and 

self-organize, the global patterns of organizational 

behavior that tend to reduce the safety of systems are 

common to all systems. Biondi (1998) [5] states that the 

organization system can have an affect on the reliability 

through numerous interrelated ways, such as work 

overload, time pressure and systemic distortion of 

information. Many recent disasters happened not because 

of the way that safety was managed through the formal 

controls and procedures, but because of the safety culture 

in which safety management approaches were 

implemented [6]. Safety culture is a sub-facet of 

organizational culture and is defined as common safety 

value in organization [7]. Certain works on the 

organizational factors have been devoted mainly to the 


classification of such factors. Embrey(1992) [8]analyzes 

railway accidents in the United Kingdom, and points that 

organizational risks factors have three levels: Level 1 

includes latent, active, and recovery errors; Level 2 

includes error-inducing factors such as training, 

procedures, time pressure, responsibilities, etc.; and Level 

3 includes policy deficiencies such as project 

management, safety culture, training policy, etc. 

Davoudian (1994) [9] proves that organizational factors 

should include overall culture (communication, decision 

making, etc.) and certain attributes of decision making, 

communication, etc. Leveson (2004) [10] views safety as 

a control problem and managed by a control structure 

developed for a socio-technical system. 

III. BUILD BBN MODEL OF ORGANIZATIONAL SAFETY 

RISKS 

For risk management purpose, it is necessary to have a 

technique that is capable of assessing the impacts of 

potential changes. BBNs can be applied for predicting the 

effects of changes[11]. As a “probabilistic” technique 

rooted in Artificial Intelligence, BBN has the capability 

of utilizing subjective expert opinions. Adapting this 

technique makes the quantification of the organizational 

accident causation theory possible, even with a lack of 

actual data. Figure 2 shows a schematic process model 

representing hierarchical structure of the process system 

of an organization, different activities at different layers 

construct organizational safety activities. Two activities 

are either sequential or hierarchical. For example, A2 and 

A 3 

A 

are related sequentially, and 22 

and 2 A are 

related hierarchically ( 

A 22 is a sub-activity of 

A 2 ). But 

in realty, it is possible that a safety output performance of 

an organization is the result of two parallel activities, 

which are neither sequentially nor hierarchically related. 

The total safety performance 0 S 

can be broken down 

S that are the outputs of parallel 

into output 1 S to k 

activities 

A 11 to 

A1 K . Each of these activities have their 

own Resource ( R ), Input ( I ) and Control/Criteria ( C ). 

The second layer of activities, comprise those that 

have R , I , and C of the layer one as their outputs. For 

example, 

R 12 is the resource for activity 12 

A22 ; I12 is the input of activity 12 A 

output of activity R 

and the output of activity 

A and the 

A22 I C 

;and 12 is the control for 

A 

activity 12 

C 

and the output of activity 

A 22 . The same 

logic will hold for the activities from layer 1 to layer N 

(the layer where the modeler stops decomposition). 

The above schematic process model can be converted 

to BBN as shown in Figure 3. In Figure 3 the quality of 

safety output would be a function of the quality of 

A to K A . Knowing the state of 1 A to K A 

activities 1 

S 

given 1 A 

as well as the conditional probabilities for 0 

to A K , we can reach the probability of safety output with 

specific state. For example, considering a binary state for 

the factor (success and failure), by knowing the


probability of success and failure for activities in layer N, 

and also the conditional probabilities, one can find the 

According to scope of organizational behavior, we 

usually divide organizational behavior into three layers, 

from individual, team to organization. Shown in table 1, 

we select training and workload & time pressure to reflect 

individual level, communication, safety administration 

and safety decision to reflect team level, salary policy and 

safety culture reflect organizational level. Each node and 

its all possible conditions should be definite and give a 

conditional probability in order to analyze effect of each 

organizational factor on system reliability using BBN. 

First of all, we need consider all nodes that construct 

BBN comprehensively. We classify all nodes into two 


probability of total output being in the success state. 

Fig.2 Schematic process model of organizational process system 

Fig. 3. Bayesian belief network for the organizational process system 

categories: human error nodes and organizational factors 

nodes which result in human errors. Then calculate 

conditional probability of each node based on the data 

sample. Human error and severe loss are selected as 

accident nodes. Secondly, we separate organizational 

factors in human error accidents database, that is to say, 

each organizational factor is divided into several states, 

and each state corresponding to a discrete value. Table 1 

shows category and characteristics of human error data 

discretization. Table 2 shows samples of organizational 

factors data.


Table1 category and characteristics of organizational factors 

layer code Organizational factors characteristics 

Layer1 X 1 

safety culture 

Layer2 

X 2 

X 3 

X 4 

X 5 

Salary policy 

Communication 

Safety administration 

Safety decision 

good (1) 

average (2) 

Poor(3) 

Layer3 X 6 

training 

X 7 

Workload & time pressure 

Accident X 8 

X 9 

human errors 

severe loss 

Yes(1) No(2) Poor(3) 

Table 2 samples of organizational factors data 

X1 X2 X3 X4 X5 X6 X7 X8 X9 

1 1 2 1 2 1 2 2 2 

2 2 2 2 1 2 1 2 2 

1 1 3 1 2 2 2 2 2 

3 2 1 1 3 1 3 1 2 

3 2 2 2 3 2 2 1 2 

2 3 3 3 2 2 3 1 1 

Traditionally, BBNs were constructed from knowledge 

of human experts. However, during the last decade 

several methods had been developed to build them 

directly from databases. In order to ensure configuration 

of BBN, all variables 1 X , X 2 ,…, X 9 need to be ordered 

according to topology order . Father node set of each 

variable should be determined and partial probability of 

each state need to be assigned. For the purposes of this 

paper the K 2 algorithm was applied to database [12]. 

K 2 finds the optimal structure through a greedy search of 

a reduced space of possible networks. The greedy 

criterion is based on a scoring function that represents the 

probability of a structure given data. In K 2 algorithm, 

Let Z be a set of n discrete variables (nodes) , where a 

X i variable in Z has i r 

possible value assignments. 

Let D be a database of m cases, where each case 

contains a value assignment for each variable in Z . Let 

Bs denote a belief network structure containing just the 

variables in Z , and p B 

the conditional probabilities. 

X i B s 

Each variable in has a set of parents, which are 

W ij 

represented with a list of variables. Let denote the 

jth 

unique instantiation of i relative to D . Suppose there 

are i q 

such unique instantiations of i N ijk 

. Define to be 

the number of cases in D in which variable i X 

has the 

r ik value and i W ij 

is instantiated as . According to the 

following theorem, the K 2 algorithm determines the 

optimal structure through a greedy procedure that 

identifies if a node can increase the network probability 

by adding a new parent to it, the structure of BBNs can be 

calculated. 

The basic structure of the K 2 algorithm is described 

with the following pseudocode: 

1. for i = 1 to n 


π i = 

2. 

{ } 

3. 

gn [ i] 

= g( 

i, 

null) 

4. finish=false 

π i 

< u 

5. while not finish and 

gnnew 

= −∞ 

6. 

j = 1 

7. for to pred 

g( pred 

8. if 

[ j] 

, π i ) 

> [ ] i gn 

then 

gnnew = g( 

pred 

9. 

[ j] 

, π i ) 

z = j 

10. 

gnnew 

11. if > [ ] i gn 

then 

gn 

12. [ i] 

gnnew 

= 

π i = π i ∪ { z} 

13. 

14. else finish=true 

Where n is number of nodes in the network, i π 

is 

array of parents of node i , u is maximum allowable 

parents that any node can have, gn is array that stores the 

maximum values of ( ) g 

associated each node, pred is 

array of predecessor nodes to each node i , z is 

prospective parent with the highest probability. 

Figure 3 presents structure of BBNs applied above 

organizational factors database. The model indicates the 

value of safety-minded companies creating a safety 

culture that enhances communicating, decision and 

monitoring procedures, thereby reducing human error and 

severe loss. This direction will probably necessitate both 

restructuring the net in order to account for influences 

that were neglected at this stage and introducing other 

influences on organizational behaviour, such as training 

and salary policy, etc.


Fig. 4 structure of BBN of organizational factors of the sample database 

IV. CONCLUSION 

Bayesian belief networks provide a robust probabilistic 

method of reasoning with uncertainty and are is more 

suitable to represent complex dependencies among 

components and to include uncertainty in modeling. In 

this paper we have demonstrated in principle that BBNs 

can be used for evaluating accident probability of 

organizational factors. We have also shown how such a 

model can be used for practical applications. As stated in 

one of the assumptions of the K2 algorithm, an ordering 

of the nodes has to be established to define the structure 

of a BBN. The model succeeds in building a quantitative 

tier on top of the qualitative explanations of 

organizational risks. We do believe that such models can 

become a reliable tool for predicting influence of 

organizational risks changes and even in orienting safety 

investments at this level. 


This work is supported by the National Natural 

Science Foundation of China(71071003), the MOE 

Project of Youth Foundation of Humanities and Social 

Science (09YJC630004), and the Natural Science 

Foundation key project of Anhui University 

(KJ2009A59). 

REFERENCES 

[1] Gordon , R.P.E. The contribution of Human Factors to 

accidents in the Offshore Oil Industry[J]. Reliability 

Engineering and System Safety, 1998, 61: 95-108. 

[2] Reason J. A systems approach to organizational error 

[J].Ergonomics, 1995, 38(8): 1708-1721. 


[3] Rasmussen, J. Risk Management in a Dynamic Society: a 

modeling problem[J]. Safety Science, 1997,27: 183-213 

[4] Oien K. A framework for the establishment of 

organiza-tional risk indicators[J]. Reliability Engineering 

and System Safety, 2001, 7: 147-167 

[5] Zahra Mohaghegh, RezaKazemi,AliMosleh. Incorporating 

organizational factors into Probabilistic Risk Assessment 

(PRA) of complex socio-technical systems: A hybrid 

technique formalization[J]. Reliability Engineering and 

System Safety, 2009, 94:1000~1018 

[6] JEAN-CLAUDE ANDRE. Complexity and occupational 

safety and health prevention research. Theoretical Issues in 

Ergonomics Science, 2005, 6 (6): 483~507 

[7] Cooper G, Herskovits E.A Bayesian Method for the 

Induction of Probabilistic Networks from Data[J].Machine 

Learning, 1992,9: 309. 

[8] Embrey, D.E.. Incorporating management and 

organizational factors into probabilistic safety 

assessment[J]. Reliability Engineering and Systems Safety, 

1992(38): 199-208 

[9] Davoudian, K., et.al.. Incorporating Organizational Factors 

into, Risk Assessment through the analysis of Work 

Processes[J]. Reliability Engineering and System Safety, 

1994, 45: 85-91 

[10] Leveson, N. A new accident model for engineering safer 

systems[J]. Safety Science, 2004, 42(4): 237-270 

[11] M. Jaeger.Complex probabilistic modeling with recursive 

relational Bayesian networks[J]. Annals of Mathematics 

and Artificial Intelligence, 2001,32:179–220 

[12] Xiao qingkunl. theory and application of dynamic 

Bayesian networks learning [M]. National defence industry 

press, 2007 

Mrs. Li Song is an associate professor in School of Economics 

and Management, Anhui University of Science & Technology, 

Huainan, Anhui, China. Her major field of study includes safety 

evaluation, organizational behavior and risk management 

(E-mail: lilysong@ustc.edu)


Research and Application of J2EE and AJAX 

Technologies in Industry Report 

Min Hu 

School of Computer & Information /Hefei University of Technology, Hefei, China 

Email: uhnim@163.com 

Ding-ding Pan and Pei-en Zhou 

School of Computer & Information /Hefei University of Technology, Hefei, China 

Email: panding1986@sina.com, albertzpe@163.com 

Abstract—The traditional system of industry report is highly 

influenced by the speed of Internet and has low efficiency on 

report. In order to solve these problems, this paper studies 

J2EE and AJAX technologies, combine them and propose 

an industry report system which based on J2EE and AJAX 

technologies. The system which makes full advantages of 

both technologies, has solved the problems such as easily 

impacted by the bandwidth, reported in low efficiency, also 

increased the server’s load capacity. It obtains a good result 

in the practical application. 

Index Terms—Industry Report, J2EE, AJAX 


With the deepening of China’s economic reform, 

various economic types and operational forms of 

companies are emerging, number and size of enterprises 

are constantly expanding, and the traditional way of 

industry report encountered a series of problems and 

faced a serious challenge in practice. 

In this case, it is imperative to establish an online 

industry report system using computer and network 

technology. Enterprises could connect to data networks of 

management institutions and submit the industry reports 

directly through the Internet. The realization of online 

industry report system which has changed industrial data 

acquisition, is an inevitable reform trend of report 

method, also speeds up the construction of statistical 

information and achieves paperless report. The 

introduction of network-based work, brings a huge 

impact in industry report, and has greatly improved the 

capacity of data collection, analysis and aggregate while 

upgrading data quality and work efficiency. 

At present, despite that the online report system has 

got a certain application, most of them have many 

problems, such as slow access, system instability, report 

on low efficiency, poor server load capacity and so on. In 

view of these problems, after making detailed studies in 

J2EE and AJAX technologies, we apply them to the 

development of industry report system and achieve good 

results; both of them fully play respective advantages. For 

Project supported by the National Natural Science Foundation of 

China (No. 60773043). 


doi:10.4304/jcp.6.9.1847-1851 

example, J2EE technology holds high scalability and 

steady availability; and AJAX technology owns strong 

response capability between the client and server. 

II. THEORIES OF J2EE AND AJAX 

A. The Theory of J2EE 

J2EE [1] is a system structure which uses Java 2 

platform standard edition as the core to simplify the 

development of enterprise solutions and deploy and 

manage some complex issues. It not only consolidates the 

advantages of the standard such as “write once, run 

anywhere”, to facilitate database access JDBC API, 

CORBA technology, security model of protecting data in 

Internet applications and so on, but also provides full 

support of EJB, Java Servlets API, JSP and XML 

technology. J2EE has many technical advantages, as 

follows: 

a) Supporting heterogeneous environment: J2EE 

could develop transplantable program which deployed in 

heterogeneous environment, and the program that 

developed once can be deployed to a variety of platforms. 

b) Scalability: Applications based on J2EE platform 

could be deployed to a variety of operating system. The 

provider of J2EE field offers a wide range of load 

balancing strategies, allows integrated deployment of 

multiple servers, and achieves highly scalable system. 

c) Steady availability: A server-side platform must 

be able to run uninterrupted. J2EE supports long-term 

availability while being deployed to a reliable operating 

system. 

J2EE uses multi-tier distributed application model. 

Application logic is divided into components in light of 

the function. And all application components locate in 

different machines according to their location in different 

tiers. Now J2EE multi-tier enterprise applications divide 

different levels of two-tier model into many tiers. A 

multi-tiered application can provide an independent tier 

for each different service. The following is a typical 

four-tier structure of J2EE [2] (Shown in Fig. 1): 

a) Client tier components running on the client 

machine. 

b) Web tier components running on J2EE server.


c) Business logic tier components running on J2EE 

server. 

d) Enterprise information system tier components 

running on EIS server. 

B. The Theory of AJAX 

Figure 1. Four-layer Structure Diagram. 

AJAX [3] works is to use the XMLHttpRequest object 

to transfer requests and responses asynchronously 

between the client and server. Fig. 2 shows the process 

flow of communication between client and server. 

XMLHttpRequest object is the core of AJAX and has 

become the actual standard of asynchronous transfer for 

XML data via HTTP. Asynchronous interaction means 

that the browser could continue processing the events 

page while sending request. Data is transferred in the 

background, and automatically loaded to the page without 

refreshing. Using AJAX technology has the following 

advantages: 

a) No page refreshing, communicating with server 

within the page, and providing a good user experience. 

b) Communicating with server using asynchronous 

mode without interrupting the user’s operation, and 

holding a more rapid response capability. 

c) Passing some of the burden work from server to 

the client, using the client’s ability to deal with, reducing 

the burden on server and bandwidth, saving space and 

bandwidth rental costs. And AJAX reduces the burden on 

the principles of “on-demand access of data”, shows the 

greatest degree of reduction of redundant request and 

response on the server. 

d) Based on a standardized and widely supported 

technology, no need to download plug-ins or applets. 

Figure 2. The Process Flow Diagram of Communication between 

Client and Server 


C. The Relation of J2EE and AJAX 

J2EE and AJAX are two technologies of Java, or two 

frameworks. They can not communicate with each other 

and both have their own advantages. On J2EE, all tasks 

are on the server, because, on the one hand, the client is 

relatively simple and does not need to do complex logic; 

on the other hand, data processing on the server is securer 

than the client. However, if the server capacity is limited, 

to improve data bandwidth and processing capabilities are 

also limited, many customers can not bear the burden. On 

AJAX, almost all services are placed on the client and 

processing speed is fast, but the client would be so 

complex that leading to poor compatibility. In general, all 

operations are focusing on the client and server. In 

response, we combine the two technologies and give full 

play to both of superiority. The core part of the 

implementation will use J2EE on server, while a 

relatively minor operation will be implemented with 

AJAX on the client. 

III. SYSTEM DESIGN 

A. The Analysis of System 

The principle system design is to ensure stability, high 

reliability, security and scalability of the data, implement 

unified interface for data interchange, exchange standards 

and authentication. Industry report system accomplishes 

the task of data collection. In order to facilitate on-line 

industry data report and ensure the access to the Internet 

effectively, we need to create a unified plan of report 

platform and data centre, specify normative 

organizational structure and data exchange standards 

while providing data interface to other systems. 

Since the work of special report, industry report system 

uses three-tier B/S architecture. This structure fully 

accounts the special report, not only provides users with a 

simple operating environment, but also ensures a quick 

and easy transfer report effectively. Enterprises connect 

to higher authorities via the Internet; they are linked into 

a seamless system by WEB technology and database. 

During industry report system B/S structure, enterprise 

users and the authorities have always been at the client. 

Enterprise users could transmit the data to authorities 

through the IE browser. The authorities can also carry out 

audit, management, statistics and summary on real-time 

data reported and print out reports (Shown in Fig. 3). The 

client is responsible for user authentication, input, report 

and audit of the data; and server is responsible for data 

reception and management. 

B. The Architecture of System 

System architecture is three-layer structure achieved 

by Struts [4][5] framework (shown in Fig. 4). It includes 

three parts: Model layer, View layer and Control layer. 

Compared with four-layer structure of J2EE, the view 

layer corresponds to the client tier components and the 

WEB tier components; model layer corresponds to 

business logic tier. 

a) Model Layer


Figure 3. Flow Chart of Industry Report 

Model Layer is the main part of the system 

architecture. In the Struts framework, model layer is 

composed of ActionForm and JavaBean. ActionForm 

will encapsulate the user’s parameters ActionForm 

object, the object is forwarded to the Action by 

ActionServlet, and Action process client requests 

according to the request parameter in the ActionForm. 

JavaBean then encapsulates the underlying business logic 

like database access etc. The system uses DAO to access 

operations on database and protect the security of 

database. 

b) View Layer 

View layer which composed by the JSP page is the 

interactive interface and achieves development and 

design of the main page of each functional module. It 

could check operational status to model, synchronize and 

update the user interface. Struts framework provides a 

rich tag library which could reduce the use of scripts; 

custom tag library can achieve an effective interaction 

with the model layer. 

c) Control Layer 

Control layer is the core of system architecture. Struts 

uses built-in Servlet—ActionServlet—as a controller, 

which receives a request from the client, enables event 

scheduling mechanism, selects the model of the 

corresponding business logic layer upon request, and then 

sends the results of the response to the client. While there 

are more concurrent operations in the client, use data 

scheduling to reduce pressure on the client by load 

balancing access technology. 

Figure 4. Framework Diagram of Struts MVC. 


C. The Design of Module 

On the base of the demand, we finalize the system 

modules, namely, enterprise information, enterprise data 

reporting, enterprise data auditing, statistical summary, 

query analysis and system management, a total of six 

modules. In accordance with the functional requirements, 

each module contains several corresponding sub-module 

(Shown in Fig. 5). 

Figure 5. Diagram of System Module 

D. The Design of Database 

In the database design [6], using Power Designer [7] 

for the design of logical model and physical model, and 

automatically generate the SQL script of SQL Server 

2008. According to the system design and function 

modules, design database tables as follows: 

a) Enterprise basic information table: used to store 

basic information of enterprises. 

b) User table and permissions group table: used to 

store user names and distribution of user rights. 

c) Production table and benefit table: used to store 

the data of production and benefit which reported by 

enterprise. 

d) Summary table: used to store the summary results 

of the data of production and benefit. 

IV. SYSTEM IMPLEMENTATION 

A. Development Environment 

Based on J2EE platform, B/S structure is applied to 

achieve system’s cross-platform deployment and 

operation. And MS SQL Server 2008 is chosen as 

background database, Tomcat 5.5 as publishing tools, and 

Eclipse 3.1 as programming tools. 

B. Implementation of Persistence with DAO 

During the development of J2EE-based system, in 

order to take full advantage of object-oriented features of 

Java, developers often design the required Java classes to 

manipulate business data. Now databases used commonly 

are relational database rather than object database. 

Therefore, the data to add, delete, change and other 

operations in the Java class can not be directly persistent 

to the relationship table of database. In this paper we 

propose JDBC and DAO pattern to solve the problem 

using JDBC to establish a connection to the database and


DAO [8] to abstract and encapsulate all access on data 

sources. And DAO is also responsible for connection 

management and data sources to obtain and store the 

data. Figure 6 shows the principle diagram of DAO. 

Figure 6. Principle Diagram of DAO 

C. Implementation of J2EE and AJAX Technology 

Struts framework can achieve MVC model of J2EE 

easily, a clear division of application makes the 

application logic and display logic independent of each 

other. However, users need to interact with server while 

the server is running. In this process users need to fill a 

large number of forms, and these operations directly 

affect the response speed of user interface. To solve the 

problem, DWR [9] which is a kind of AJAX technology 

is introduced in this paper. Its biggest advantage is 

safeguarding data without updating the page, combination 

of asynchronous nature of AJAX and synchronous nature 

of normal Java method calls. In asynchronous mode, the 

resulting data could be accessed asynchronously after the 

call has been executed for a long time. 

DWR which is an open source library contains two 

main parts: First, JavaScript could get data from a Servlet 

which is in the WEB server and follows the principle of 

AJAX; second, a JavaScript library could help WEB 

developers to use the obtained data and change the 

content of page dynamically. In addition, DWR has 

adopted a new method which is similar to AJAX to 

dynamically generate JavaScript code which based on 

Java class, so the WEB developers could use the Java 

code in JavaScript. However, the Java code runs on 

server and is free to visit the WEB server resources. 

Finally, considering to the security, WEB developers 

must properly configure the Java class which can be used 

outside safely. 

V. EXAMPLES OF J2EE AND AJAX 

In this paper, we introduce the application of J2EE and 

AJAX technologies with the example: monthly report of 

economic benefits. 

A. The Implementation of J2EE Technology 

Struts framework is mainly used to implement the 

MVC pattern of J2EE technology. It provides the 

controller which is inherited HttpServlet class and 

intercepts all HTTP requests, then calls the model layer to 

complete the request upon the HTTP requests and passes 

the final result to the client. To achieve the functionality, 

we need configure the struts-config.xml file as follows: 

 

 


 

 

name=”success” 

 

 

 

…… 

 

 

 

name=”failure” 

And then implement the subclass which inherited the 

corresponding Action class. 

B. The Implementation of AJAX Technology 

The response of user interface is too slow while system 

is running. In order to solve the problem, this paper 

successfully introduced DWR which is a kind of AJAX 

technology. DWR allows passing a callback function 

which used to process the Java function call 

asynchronously. 

a) Configure dwr.xml file as follows: 

 

 

 

creator=”new” 

 

 

name=”class” 

 

…… 

 

 

b) Javascript Calls 

 

 

 

 

Function dosubmit(){ 

…… 

JjzbybbManager.submitJjzbybb(); 

…… 

} 

 

Then it enables the client to call the function in class 

JjzbybbDao to complete the operations of economic 

reports. 

VI. CONCLUSIONS 

Based on B/S architecture and relational database, in 

this paper we design and implement industry report 

system based on J2EE and AJAX technologies. During 

the implementation, we achieve the MVC pattern of J2EE 

with Struts framework, design the logical and physical 

models with Power Designer tool, and implement the 

development of data persistence with DAO technology. 

Also, with the combination of J2EE and AJAX, the


system has good scalability and security, fully embodies 

the design advantages of MVC pattern, and receives the 

desired results in practical application. 


The authors would like to thank the guest editors 

reviewer for their valuable comments and insightful 

suggestions. This research was supported by the National 

Natural Science Foundation of China (No. 60773043). 

REFERENCES 

[1] W. Grawford and J. Kaplan. J2EE Design Patterns, 

Beijing: China Electric Power Press, 2005. 

[2] Liu Yang, Gao Liansheng and Wang Bin. “Study and 

implement of distribution system based on J2EE and MVC 

design pattern”. Computer Engineering and Design, 2007, 

7:1655-1658. 

[3] D. Johnson, A. White and A. Charland. Enterprise AJAX 

strategies for building high performance applications, 

Beijing: People’s Posts and Telecommunications 

Publishing House, 2008. 

[4] Yang Shaobo. Struts framework technology, Beijing: 

Tsinghua University Press, 2008. 

[5] J. Carnell, R. Harrop and K. Mittal. Pro apache struts with 

ajax, Beijing: People’s Posts and Telecommunications 

Publishing House, 2008. 

[6] Wang Shan and Sa Shixuan. Introduction to Database 

System, 4 th ed., Beijing: Higher Education Press, 2006. 

[7] Jiang Xueying. Web Database Design and Development, 

Beijing: Tsinghua University Press, 2007. 

[8] D. Alur, J. Crupi and D. Malks. J2EE Core Model, Beijing: 

China Machine Press, 2002. 

[9] Frank W.Zammetti. Practical DWR 2 projects, Beijing: 

People’s Posts and Telecommunications Publishing House, 

2009. 


Min Hu received her bachelor's degree 

in Automation Engineering from 

Nanjing University of Technology, 

Nanjing, China (1988), master's degree 

in Automation Engineering from Hefei 

University of Technology, Hefei, China 

(1994) and Ph.D. in Computer 

Application Technology from Hefei 

University of Technology, Hefei, China (2004). Since 1994, she 

is teaching at the School of Computer and Information of Hefei 

University of Technology, where currently she is a Professor 

and actively participates in a number of national natural science 

foundation research projects concerning application and 

research of the theory and methods of multivariate rational 

interpolatory approximation in graphics, and research on the 

application of digital image processing based on continued 

fraction methods. Her current teaching and research interests 

include computer application, digital image processing and 

digital watermarking. She has authored more than 30 papers in 

international conferences and journals. 

Ding-ding Pan received his bachelor's degree in Computer 

Science and Technology from Anhui University of Architecture, 

Hefei, China (2008) and master's degree in Computer 

Application Technology from Hefei University of Technology, 

Hefei, China. His research interests include computer 

application, software engineering. 

Pei-en Zhou received his bachelor's degree in Computer 

Science and Technology from AnQing Teachers College, 

Anqing, China (2008) and master's degree in Computer 

Application Technology from Hefei University of Technology, 

Hefei, China. His research interests include computer aided 

design.


The Analysis of China New Energy Vehicle 

Industry Alliance Status based on UCINET 

Software 

Xiongfei Guo * 

School of Economics and Management/Beijing Jiaotong University, Beijing, China 

Yingqi Liu 

School of Economics and Management/Beijing Jiaotong University, Beijing, China 

Abstract—New energy vehicle industry acquires highly 

complex techniques. The new energy vehicle industry 

alliance is one of the most effective origination form and has 

been developed fast in China. This paper mainly use the 

software UCINET to draw up the picture of China’s new 

energy vehicle industry alliance network, and study the 

cooperation relationships within the alliances through 

analyzing their elements. The results find that the key factor 

that effects the development of China’s new energy vehicle 

industry alliance is automobile companies. The key point in 

future management and research of China’s new energy 

vehicle industry alliance are cooperation and management 

between automobile companies and other members in the 

alliance. 

Index Terms—new energy vehicle; industry alliance; 

element; automobile company 


The development of new energy vehicle industry not 

only can help solving problems like energy security, 

carbon dioxide emissions, but also increase companies’ 

ability of innovation, motivate industrial upgrade. In the 

past two decades, because of the complexity of new 

energy vehicle’s technologies and its research involves 

too many fields, there’re few companies that can master 

all technologies. Foreign new energy vehicle industries 

mainly adopted the organization of industry alliance, one 

of the most important organization forms to process the 

technical innovations. For example, in the year of 2010, 

Toyota and other Japanese automobile companies 

announced an alliance called CHAdeMo. Its group 

members include Toyota, Nissan, Mitsubishi, Fuji and 

Tokyo elec ∗ tric vehicle company. There’re 160 

companies joined the alliance in total, including foreign 

capital and government agencies (Japan Electric Vehicle 

Alliance, 2010). 

New energy vehicle industry was treated as a national 

strategic emerging industry, thus to gain fast 

development. From 2009, when China’s first self-owned 

brand new energy vehicle, Changan Jiexun, entered the 

market, to the end of 2010, many companies had carried 

∗ corresponding author of this article. 


doi:10.4304/jcp.6.9.1852-1856 

out their own new energy vehicle products. Inspired by 

the development of foreign new energy vehicle 

industries, China’s new energy vehicle industry alliance 

developed fast. Nearly 30 different types of new energy 

vehicle industry alliances had been established in China 

by the end of 2010 1 . 

This article mainly starts with the status of China’s 

new energy vehicle industry alliances, analyzing the 

elements within each alliance and seeks for related 

conclusions. 

II. ALLIANCE AND PARTNER TYPE 

Alliances are voluntary, cooperative agreements 

between two or more firms designed to achieve an 

advantage for the partners (Das and Teng, 2000). More 

generally, according to Gulati (1998), we define alliances 

are voluntary agreements between independent firms to 

develop and commercialize new products, technologies or 

services. Portfolio companies in the form of the formation 

of industry alliances and new product development to 

overcome the inherent risk associated with the control 

process of innovation and better results have reached a 

consensus (Jarillo, 1988; Gulati, Nohria, and Zaheer, 

2000). Hence, in the past two decades, Industry Alliance 

as an important form of industrial organization abounds 

in the automobile industry (Garcia-Pont and Nohria, 

2002). 

An entrepreneurial venture using R&D alliances within 

the new product development process has three distinct 

choices of partners differentiated by their position along 

the industry value chain (Baum et al., 2000). According 

to Rothaermel and Deeds (2006), partner can be divided 

into three different levels: upstream partners, horizontal 

partners and downstream partners. However, the division 

will be different to each other due to different types of 

alliances and industrial chains. 

In dimensions of new energy vehicle industry and 

division of work, industries’ upstream partners mainly 

involve in research developments, and horizontal partners 

mainly involve in products fabrication, while downstream 

1 Gathered by authors.


partner mainly involve in sell or services. Theoretically, 

upstream partners are research institutions, they basically 

do research developments from basis study to tapered 

technologies, and they all belong to institutional units. 

Horizontal partners should consist of automobile 

companies and automobile parts manufacturers. Their 

duty is to provide productivity. Although both automobile 

companies and automobile parts manufacturers are in 

same level, in fact, comparing to automobile assemble 

companies, automobile parts manufacturers are still a 

kind of upstream units. Downstream partners mainly 

consist of transportation companies and dealers etc., they 

mainly involve in sells or service providing. Both 

horizontal partners and downstream partners belong to 

enterprise units. 

However, not all alliance involve all these three levels, 

some alliance may consist of the companies come from 

the same level, which is called horizontal alliance. Others 

may be more integrated, its member may involve the 

most part of the industrial chain (Like SOE electric 

vehicle alliance or alliance in Beijing and Chongqing), 

which named vertical alliance. Factors that affect the 

formation of the alliance could be various. 

III. THE STATUS OF CHINA NEW ENERGY 

VEHICEL INDUSTRY ALLIANCE 

According to New Energy Automobile Manufacturing 

Companies Product Standards and Managing Rules 

Published by China Ministry of Industry, "New energy 

vehicles refers to the vehicles using unconventional 

vehicle fuels as a power source (or use the conventional 

vehicle fuel power plant using the new device), integrated 

power control and vehicle's advanced driving technology, 

adopting Advanced technological principles, and with 

new technology and structure." Include hybrid vehicles, 

electric vehicles (including solar power vehicles), fuel 

cell electric vehicles, hydrogen vehicles. And other new 

energy sources (such as high energy storage devices, 

DME) vehicles and products. In short, the new energy 

vehicles refers in the fuel or power systems differ from 

traditional internal combustion engine vehicle motor 

vehicles (Ministry of Industry, 2009). 

Currently, among China new energy vehicle industry 

alliances, in the geographical point of view, are divided 

into three categories: the first is established in their 

respective regions alliance; second is the alliance 

established at the national level; third category is 

international Union. The early established alliance in 

China is the regional alliance. Beijing new energy vehicle 

industry alliance as the first new energy vehicles industry 

alliance in China began its operation officially on March 

13, 2009. With Beijing's new energy vehicle industry 

alliance formed up, Chongqing, Hubei, Shanghai, 

Tianjin, Jilin, Zhejiang, Guangzhou, Anhui and Chengdu 

and other places to set up their own new energy vehicle 

industry alliance. The number of new energy vehicle 

industry is raised to 30. These alliances are categorized 

into national level, regional level and so-called 

international level. There’re both full of new energy 

vehicles alliances that covered entire industrial chain and 


alliances that only covered the chain of production and 

market services. The scale of these alliances ranges from 

60 to 2. There are both alliance with foreign enterprises 

and those involve only domestic firms. Meanwhile, the 

constitution of the various alliances has great differences 

among these alliances. Some requested the federal 

procurement, and some are not mandatory. 

Most of the new energy vehicle industry alliances have 

strong research and development abilities. According to 

the statistics data from 2009, the top 15 patent applicants 

in the main manufacturers are: Chery, BYD, FAW, 

Changan, and research institutions: Tsinghua University, 

Chongqing University, Shanghai Jiaotong University, 

Zhejiang University and Jilin University. Meanwhile, the 

core members of the industry alliances also invested 

strong R&D funding. For example, Chongqing Changan 

invested 2.52 billion yuan for R&D, FAW Group’s 

investment in new energy research and development in 

the Eleventh Five-Year period up to 3 million, receiving 

65 patents, 38 patents by the U.S. Patent Office (SAC, 

2010). 

IV. ELEMENTS ANALISIS IN INDUSTRY 

ALLIANCE 

To show the status of China’s new energy industry 

alliance better. We used UCINET (Borgatti, Everett, & 

Freeman, 2002) to build the network of China’s new 

energy vehicle industry alliance (Figure 1). We used the 

data to build an organization-organization binary matrix 

to define the relationship between one independent 

company and another. 

As Figure 1 shows, most Chinese new energy vehicle 

industry alliances are connected to each other except few 

ones. Members in these alliances, such as Tsinghua 

University, Beijing Institute of Technology, Wuhan 

Institute of Technology etc. are research institutions. 

There also have been relationships between automobile 

companies and its division companies, such as Dongfeng 

Automobile and Dongfeng Yunnan, FAW and Tianjin 

FAW etc.. There’re relationships between automobile 

parts manufacturers and division companies, such as 

Chunlan Electric and Chunlan clean-energy research 

institution etc.. Noted some companies have joined more 

than one alliance. Among our sample alliances, there’re 

only 3 alliances are not connected to others. Hence, these 

three alliances are lack of connection among the entire 

industry alliance. 

Our sample alliances consist of 183 companies or 

institutions. According to results, we divided them into 6 

groups: 1) Automobile parts manufacturer; 2) 

Automobile company; 3) Research institution; 4) 

Transportation company; 5) Government agency; 6) 

Financial Institution. Among this companies and 

institutions, automobile parts manufacturer has the largest 

number, 63; automobile company is second on the list, 

50. The third one is research institution, 48. Followed 

with, transportation company (9), government agency (9) 

and financial institution (5). (Table I) Some of the 

companies involved more than one business, like 

companies that both doing research and product


manufacturing, we distinguish them according its major 

role in the alliance. Other institutions looked like 

enterprise on surface, which mainly involve in 

coordination and management will be treated as 

government agencies. In fact there’s no place for 

government agencies and financial institutions in the 

industry chain, so we simply can’t put them into industry 

chain. 

Considering automobile parts manufacturer, 

automobile company, transportation company and 

26% 

research institution, according to the previews theory, 

there’re 48 upstream partners (Research Institution), 

26.2% in total; 113 horizontal partners (automobile parts 

manufacturer and automobile company), 61.7% in total; 9 

downstream partners (transportation company), 4.9% in 

total. Consequently, there’s more than half of the 

members in new energy vehicle industry alliance is in the 

position of horizontal level. 

Figure 1. China’s new energy vehicle industry alliance network 

Source: all the data gathered by authors 

5% 

5% 

3% 

27% 

34% 

Automobile Parts Manufacturer 

Automobile Company 

Research Institution 

Transportation Company 

Government Agency 

Financial Institution 

Figure 2. Members in China’s new energy vehicle industry alliance network 

Source: all the data get by authors 

From the point of view of cooperation, the total 

number of cooperative relationships involved in all 

alliances is 496. Among these cooperative relationships, 

partnerships between automobile companies and 

automobile parts manufacturers reached 126 (25.4%); 

partnerships between automobile companies reached 110 

(22.2%); partnerships between automobile companies and 

research institutions reached 96 (19.4%); partnerships 

between automobile companies and government agencies 

reached 39 (7.9%); partnerships between research 


institutions and automobile parts manufacturers reached 

23 (4.6%); partnerships between automobile parts 

manufacturers reached 21 (4.2%); partnerships between 

automobile parts manufacturers and government agencies 

reached 17 (3.4%); partnerships between research 

institutions reached 14 (2.8%); partnerships between 

research institution and government agencies reached 11 

(2.2%); partnerships between automobile companies and 

transportation companies reached 5 (1.0%); partnerships 

between government agencies and transportation


companies reached 2 (0.4%); partnerships between 

government agencies and financial institutions reached 2 

(0.4%); partnerships between government agencies 

reached 1 (0.2%). In general, the major types of 

partnerships are automobile company-automobile parts 

manufacturers, automobile company- automobile 

company and automobile company-research institution 

(66.9% in total). 

TABLE I. CHINA’S NEW ENERGY VEHICLE INDUSTRY ALLIANCE PARTNERSHIP 

Automobile Parts Manufacturer Partnership: Research Institution Partnership: 

Automobile Company 126 65.6% Automobile Company 96 64.0% 

Research Institution 23 12.0% Automobile Parts Manufacturer 23 15.3% 

Automobile Parts Manufacturer 21 10.9% Research Institution 14 9.3% 

Government Agency 17 8.9% Government Agency 11 7.3% 

Financial Institution 5 2.6% Financial Institution 6 4.0% 

Automobile Company Partnership: Government Agency Partnership: 

Automobile Parts Manufacturer 126 32.1% Automobile Company 39 54.2% 

Automobile Company 110 28.1% Automobile Parts Manufacturer 17 23.6% 

Research Institution 96 24.5% Research Institution 11 15.3% 

Government Agency 39 9.9% Financial Institution 2 2.8% 

Financial Institution 14 3.6% Transportation Company 2 2.8% 

Transportation Company 7 1.8% Government Agency 1 1.4% 

Transportation Company Partnership: Financial Institution Partnership: 

Automobile Company 7 77.8% Automobile Company 14 66.7% 

Government Agency 2 22.2% Automobile Parts Manufacturer 5 23.8% 

*. Source: all the data gathered by authors 

As TABLE I shows, we can generally find different 

traits of different types of partners. The major 

partnerships of automobile parts manufacturers are those 

with automobile companies; the major partnerships of 

automobile companies are those with automobile parts 

manufacturers, other automobile companies and research 

institutions; the major partnerships of Transportation 

companies are those with automobile companies; the 

major partnerships of research institutions are those with 

automobile companies; the major partnerships of 

government agencies are those with automobile 

companies; the major partnerships of financial 

institutions are those with automobile companies. In 

results, automobile company is the focal role of the entire 

new energy vehicle industry alliance. 

Generally, there’re four types of alliance formation we 

can find in Figure 1: 

(a) Dyadic alliance that automobile company acts as 

focal firm: This kind of alliance is formed in tree shape in 

Figure 2. Basically one company or more established 

partnership with focal firm, and no partnership existed 

among non-focal firm. 

(b) Dyadic alliance that government agency acts as 

focal firm: This kind of alliance is much similar to the 

first one, the only difference is that the focal firm is 

government agency. Meanwhile, no partnership existed 

among non-focal firm. 

(c) Alliance network: Partnerships existed between 

multiple companies and institutions, each alliance 


Government Agency 2 9.5% 

member established partnership to every other member, 

thus to form up a network. 

(d) Compound Alliance: Compound alliance refers to 

an alliance network with dyadic relationships, which 

means not all companies established partnerships to each 

other. 

We find Dyadic alliance that automobile company acts 

as focal firm is the main stream of alliance formation in 

Chinese new energy vehicle industry alliance. Others like 

Dyadic alliance that government agency acts as focal 

firm; Alliance network and Compound Alliance are lesser 

when compared with the first alliance formation. 

V. CONCLUSION 

After our analysis elements within China’s new energy 

vehicle industry alliance, our findings are summarized as 

follow: 

(a) Chinese new energy vehicle industry alliance 

formation can generally be divided into four types: 

Dyadic alliance that automobile company acts as focal 

firm, Dyadic alliance that government agency acts as 

focal firm, Alliance network and Compound Alliance. 

Dyadic alliance that automobile company acts as focal 

firm is the main stream of alliance formation in Chinese 

new energy vehicle industry alliance, while others are 

lesser when compared with the first alliance formation. 

(b) There’s more than half of the members in new 

energy vehicle industry alliance is in the position of 

horizontal level. In general, the major types of 

partnerships are automobile company-automobile parts


manufacturers, automobile company- automobile 

company and automobile company-research institution. 

(c) Automobile company is the focal role of the entire 

new energy vehicle industry alliance. The major 

partnerships of automobile parts manufacturers are those 

with automobile companies; the major partnerships of 

automobile companies are those with automobile parts 

manufacturers, other automobile companies and research 

institutions; the major partnerships of Transportation 

companies are those with automobile companies; the 

major partnerships of research institutions are those with 

automobile companies; the major partnerships of 

government agencies are those with automobile 

companies; the major partnerships of financial 

institutions are those with automobile companies. 

In summarize, we conclude that automobile company 

is the key factor of the entire China’s new energy vehicle 

industry alliance. Thus, improving and strengthening the 

cooperation between automobile companies and other 

members in the alliance is the essential to the 

development of alliance. This paper provides the basis for 

quantification study in the future. Further research will 

involve in theoretical and empirical studies on innovation 

model. 

REFERENCES 

[1] Baum, Joel A. C., Calabrese Tony. Silverman. Brian 

S.(2000). Don't go it alone: alliance network composition 

and startups' performance in Canadian biotechnology. 


Strategic Management Journal. Special Issue: Strategic 

Networks, 21(3): 267–294. 

[2] Das, T.K., Teng, B.-S. (2000). A resource-based theory of 

strategic alliances. Journal of Management, 26 (1):31–60. 

[3] Garcia-Pont, C. and N. Nohria (2002) “Local versus 

Global Mimetism: The Dynamics of Alliance Formation in 

the Automobile Industry”, Strategic Management Journal, 

23, 307-21. 

[4] Gulati Ranjay(1998).l Alliances and networks. Strategic 

Management Journal.Special Issue: Editor's Choice. 19(4): 

293-317. 

[5] Gulati, R., Nohria, N., Zaheer, A. (2000). Strategic 

Networks, Strategic Management Journal, 21: pp. 203-215. 

[6] Japan Electric Vehicle Alliance. 

www.CHAdeMOchademo.com. 

[7] Jarillo, J. Carlos(1998). On strategic networks. Strategic 

Management Journal. 9(1): 31–41. 

[8] Ministry of Industry . New Energy Automobile 

Manufacturing Companies Product Standards and 

Managing Rules [EB / 

OL].http://www.miit.gov.cn/n11293472/n11293832/nl129 

3907/n11368223112425871.html.2009-06-25. 

[9] Rothaermel, Frank T. & Deeds, David L., 2006. "Alliance 

type, alliance experience and alliance management 

capability in high-technology ventures," Journal of 

Business Venturing, Elsevier, vol. 21(4):429-460 

[10] State-owned Assets Supervision and Administration 

Commission. www.sasac.gov.cn/ 

[11] Stephen P Borgatti, M G Everett, Linton C Freeman(2002). 

Ucinet for Windows: Software for Social Network 

Analysis. Harvard Analytic Technologies, 2002.


Efficiency Evaluation Information System Based 

on Data Envelopment Analysis 

Jing Han 

Department of Economics and Management, Huainan Vocational & Technical College, Huainan, Anhui, China 


Malin Song 

School of Statistics and Applied Mathematics, Anhui University of Finance and Economics, Anhui Bengbu, China 

Email: songmartin@163.com 

Abstract—As data envelopment analysis (DEA) has been 

developed both in theory and application, the calculation of 

models become more and more important. Although many 

DEA software tools have been built for the calculation of the 

DEA models, there are some deficiencies in embedding them 

into enterprise management information system (MIS). As 

an extension of this work, an idea was generated in this 

paper, which could both calculate the DEA and further 

support the decision making for decision making units 

(DMUs), i.e., the organizations, in the information 

environment. This is an attempt to bridge between DEA and 

MIS. And we could demonstrate this approach for building 

efficiency evaluation information system. Furthermore, an 

efficiency evaluation information system of company A, 

which was built by ourselves, was shown to illustrate our 

purpose. 

Index Terms—Data envelopment analysis; Management 

information system; Efficiency evaluation information 

system; Decision support 


The efficiency evaluation is becoming more and more 

considerable in companies' daily management operating. 

By looking at the efficiency evaluation, the enterprises 

can be aware of their specific position, and find out the 

gap between them and their competitors, so as to 

determine how they could improve the quality of 

products on practical and scientific aspects. 

Data envelopment analysis (DEA), as a non-parametric 

programming technique, has becoming more and more 

popular in evaluating the performance efficiency of a set 

of homogenous decision making units (DMUs). It was 

first proposed by Charnes, Cooper and Rhodes in 1978 

[1] and extensively applied in multiple inputs and 

multiple outputs complex systems. Since the CCR model, 

there has been an impressive growth both in theoretical 

developments and applications of DEA. DEA researchers 

have developed a number of updated models, such as 

BCC model [2], additive model [3], multilevel models [4, 

5], super efficiency models [6] and so on. At the same 

time, DEA has also been extensively applied in 

performance evaluation and benchmarking of hospitals, 

universities, cities, courts, business firms, and others, 

including the performance of regions, countries etc [7]. 


doi:10.4304/jcp.6.9.1857-1861 

However, the applications of DEA in the enterprise 

management information system are few. 

There have been several DEA software tools in market. 

They can be divided into two groups: one group is 

professional, such as DEA Solver pro, DEAP, Efficiency 

Measurement System (EMS), DEA excel solver and so 

on [8]. For these DEA software tools, we can get the 

results just by inputting the DMUs’ data and choosing the 

appropriate model. The other is universal, such as Matlab, 

Lingo, Lindo and so on. Using these universal software 

tools, we must program the procedure by ourselves. 

However, all DEA software tools mentioned above can’t 

be embedded into management information system (MIS) 

perfectly. This limits their application strongly. Based on 

the theory of DEA, this paper tries to set up an evaluation 

information system for company A upon the platform of 

DEA in order to supply some useful management 

information for it. Company A’s MIS may contain a lot 

of sub-systems, such as staff information management, 

salary management, performance management and so on. 

We focused on the evaluation system and its relationship 

with others. For simplifying illustration, we construct the 

efficiency evaluation information system just based on 

CCR and BCC for real company A. 

This paper is aimed at evaluating DMUs and 

benchmarking by using efficiency evaluation information 

system. This approach has some applied advantages, 

especially in the information management. Section 2 

briefly reviews the traditional DEA models of CCR and 

BCC. Section 3 introduces the efficiency evaluation 

sub-system. In Section 4, we apply the idea to build MIS 

of a real company A which contains the efficiency 

evaluation sub-system. Finally, Conclusions are given in 

Section 5. 

II. DEA MODELS 

We assume that there are n DMUs to be evaluated, 

where each DMU contains s different outputs and m 

different inputs. We denote the ith input and rth output 

for DMUj( 

j = 1, 2,..., n 

) as ij x ( i = 1, 2,..., m ) and 

yrj 

( r = 1, 2,..., m ) respectively. We assume that 

x ≥ 0 ij , 

yrj ≥ 0 

and each DMU must has at least one


positive input and one positive output value. 

The CCR model, proposed by Charnes et al in 1978 

[1], for measuring the technical efficiency of the 0 jth 

DMU 

DMU ( 0 ) was first stated as follows. 

s m 

CCR 

e = max u y / vx 

s m 

∑ ∑ 

∑ ∑ 

r rj0 i ij0 

r= 1 i= 

1 

s.t. uy − vx ≤ 0, j= 1,2,..., n 

r rj i ij 

r= 1 i= 

1 

ur, vi ≥ 0, r = 1,2,..., s, i = 1,2,..., m. 

(1) 

Through the Charnes and Cooper transformation [9] 

for linear fractional programming yielded the equivalent 

programming problem as follows. 

s 

CCR 

e = max∑µ ryrj0 r= 

1 

s m 

∑ ∑ 

s.t. µ y − ω x ≤ 0, j = 1, 2,..., n 

r rj i ij 

r= 1 

m 

i= 

1 

∑ 

i= 

1 

ω x 

i ij0 

= 1 

µ r, ωi 

≥ 0, r = 1, 2,..., s, i = 1, 2,..., m. 

(2) 

for which the LP dual problem is 

CCR 

e = min θ 

n 

∑ 

s.t. λ x + s = θx 

, i = 1,2,..., m 

j= 

1 

n 

∑ 

j= 

1 

j ij 

− 

i i0 

λ y − s = y , r = 1,2,..., s 

j rj 

+ 

r r0 

λ j ≥ 0, j = 1,2,..., n. 

(3) 

Model (3) is sometimes referred to as the “Farrell 

model” because it is the one used in Farrell [10]. In the 

economics portion of the DEA literature, it is said to 

conform to the assumption of “strong disposal”, because 

it ignores the presence of non-zero slacks. Besides, it is 

also under the assumption of constant returns to scale 

(CRS). 

Then, based on CCR model, Banker, Charnes and 

Cooper built the BCC model as follows [2]. 

BCC 

e = min θ 

n 

∑ 

s.t. λ x + s = θx 

, i = 1,2,..., m 

j= 

1 

n 

∑ 

j= 

1 

n 

∑ 

j= 

1 

j ij 

− 

i i0 

λ y − s = y , r = 1,2,..., s 

j rj 

+ 

r r0 

λ = 1 

j 

λ ≥ 0, j = 1,2,..., n. 

j 


(4) 

In the economics portion of the DEA literature, the 

BCC model, that is (4), are under the assumption of 

variable returns to scale (VRS). BCC model could be 

used to determine the returns to scale, including 

decreasing, constant and increasing. 

DMU 0 is efficient if and only 

For model (3) and (4), 

− * + * 

* 

if θ = 1 s = s = 0 

DMU 

and i r for all i and r. 0 is 

* 

* 

weakly efficient if θ = 1 s 0 

and i 

− ≠ 

and (or) 

* 

s 0 r 

+ = 

for some i and r in some alternate optima. 

DMU * 

0 is inefficient if θ < 1 [11]. Assume the 

CCR and BCC scores of a DMU are CCR 

e and BCC 

e 

respectively. The scale efficiency is defined by 

scale CCR* BCC* 

e = e / e [12]. 

The following conditions identify the situation for 

returns to scale (RTS) for the CCR model given in (3). 

xˆ ˆ 

(i) Increasing RTS prevail at ( 0 : y 0) 

if and only if 

n 

∑ 

* 

λ < 1 j 

for all optimal solutions. 

xˆ ˆ 

(ii) Decreasing RTS prevail at ( 0 : y 0) 


j= 

1 

n 

∑ 

* 

λ > 1 j 

for all optimal solutions. 

xˆ ˆ 

(iii) Constant RTS prevail at ( 0 : y 0) 


j= 

1 

n 

∑ 

j= 

1 

* 

λ = 1 j 

for at least one optimal solution [13]. 

III. BASIC FUNCTION OF EFFICIENCY EVALUATION 

INFORMATION SYSTEM 

We thought that the efficiency evaluation sub-system 

should contain at least three parts. One is evaluation 

among it and its homogeneous DMUs, another is 

evaluation among its performance in different time, and 

the last is benchmarking. 

The first evaluation aims to determine the efficiency of 

DMU 0 when it compares with other DMUs from 

cross-sectional data. The 

DMU 0 can be aware of its 

location exactly among the same kind of DMUs by the 

results above. It is very useful for the DMU to understand 

the gap between itself and other DMUs. 

The second evaluation aims to determine the efficiency 

of 

DMU 0 in series times. Through the results, we can 

know whether the DMU’s performance has been 

improved. Here, we assume the DMU is a company. 

Through specifically understanding the developments and 

trends of the company, we can do better preparation for 

its future development, in order to avoid irreparable loss 

caused by the company management’s delaying.


An important part of organizational planning and 

control is the selection of proper performance 

benchmarks [14]. Benchmarking is a means to evaluate 

their own businesses and study other organizations. It 

takes the internal or external best practices in business 

enterprises as its own internal development goals, and 

then applies the goal to their business practice. Through 

the results, we can determine the practical and scientific 

path for efficiency improvement. 

Now, we began to introduce the efficiency evaluation 

sub-system in details. This sub-system contains three 

parts. The first part was evaluation among Company A 

and its homogeneous companies, the second part was 

evaluation among its performance in different time, and 

IV. REAL COMPANY A’S EFFICIENCY EVALUATION 

INFORMATION SYSTEM 

Before introducing the efficiency evaluation 

sub-system, we should introduce the management 

information system briefly which developed by ourselves. 

It contained five parts: system management, basic 

information management, sale data analysis, efficiency 

evaluation, data inquiry. The interface was designed as 

Fig.1. 

FIG.1 COMPANY A MANAGEMENT INFORMATION SYSTEM INTERFACE 

the last was benchmarking. The system’s interfaces 

were shown as Fig.2, 3, 4 respectively when they were 

running based on CCR model. 

FIGURE.2 EVALUATION OF OPERATIONAL EFFICIENCY THROUGH CROSS-SECTIONAL DATA 

From Figure.2, we could gain the companies’ 

efficiency value and their sizes intuitionally. 


From Figure.3, we could gain the company’s 

efficiency value and their trend in time series 

intuitionally.


FIGURE.3 EVALUATION OF OPERATIONAL EFFICIENCY OF COMPANY A OVER THE YEARS 

From Figure.3, we could gain the company’s 

efficiency improvement proposals for efficient and its 

returns to scale intuitionally. 

Part of the data used for DMUs’ evaluation was 

generated automatically by the information system. The 

others were inputted by an interface from outside. 

The calculations of the DEA models were operated at 

computer background. The results were stored in 

database. When we need the related data, we could call 

them by programming directly. This approach could 

reduce the running time apparently. For example, if we 

should evaluation the Company A’s efficiency among all 

DMUs, we just called the evaluation results in database 

by using SQL language instead of calculating the DEA 

models, that was, linear programming. If operations are 

frequent, the former’s advantages, which had low time 

and space complexity, would show out. 

V. CONCLUSION 

DEA has been used in many fields popularly. 

Nowadays, DEA has been used widely in many fields. 

There are several software tools for dealing with DEA 


FIGURE.4 BENCHMARKING BASED ON CCR 

model, including professional and universal tools. All 

these tools can deal with some DEA models. However, 

they can not become part of enterprise management 

information system perfectly, which is popularly used for 

management in our firms now. To build efficiency 

evaluation information system is very useful. This paper 

briefly introduces the DEA model and the parts which the 

system should contain. At last, for illustrating our idea, 

we take an efficiency evaluation information system of a 

real company A as an example. 

As one of the solutions, our proposed approach is only 

one way to integrate DEA into management information 

system. This will help managers grasp the state of their 

company among the same kind companies better. It is 

also useful to gain the company development trend 

during the time series. Last but by no means least, the 

system can make benchmarking and propose some useful 

suggestions for company too. However, our system is 

based on personal platform. This may limit its usage in 

some degree. Therefore some extensions can be studied 

in the future. The next work we will do is to build an 

efficiency evaluation information system based on Web.


REFERENCE 

[1] Charnes, A., Cooper, W.W., and Rhodes, E., “Measuring 

the efficiency of decision making units”, European Journal 

of Operational Research, vol. 2, No. 6, 1978, pp. 429–444. 

[2] Banker, R.D., Charnes, A., Cooper, W.W., “Some models 

for estimating technical and scale inefficiencies in data 

envelopment analysis”, Management Science, Vol. 30, No. 

9, 1984, pp. 1078-1092. 

[3] Charnes, A., Cooper, W.W., Golany, B., Seiford, L.M., 

Stutz, J., “Foundations of data envelopment analysis and 

Pareto-Koopmans empirical production functions”. Journal 

of Econometrics, Vol. 30, No. 9, 1985, pp. 91-107. 

[4] Fare, R., Grosskopf, S., Intertemporal Production 

Frontiers: With Dynamic DEA. Kluwer Academic, Boston, 

MA, 1996. 

[5] Liang, L., Yang, F., Cook, W. D., Zhu, Joe., DEA models 

for supply chain efficiency evaluation, Annals of 

Operations Research, Vol. 145, No. 1, 2006, pp. 35-49. 

[6] Andersen, P., Petersen, N.C., “A procedure for ranking 

efficient units in data envelopment analysis”, Management 

Science, Vol. 39, No. 10, 1993, pp. 1261–1264. 

[7] Cooper, W.W., Seiford L. M., Zhu Joe (Eds.), Data 

envelopment analysis, Kluwer Academic Publishers, 

London, 2004, pp. 1-2. 



London, 2004, pp. 539-564. 


[9] Charnes, A., Cooper, W.W., “Programming with linear 

fractional functionals”, Naval Research logistics quarterly, 

vol. 9, No. 3-4, 1962, pp. 181-185. 

[10] Farrell M.J., “The measurement of productive efficiency”, 

Journal of Royal Statistic Society. Series A, Vol. 120, No. 

3, 1957, 253-281. 



London, 2004, pp. 8-13. 

[12] Cooper W. W., Seiford L. M., Tone Kaoru (Eds.), Data 

envelopment analysis: A comprehensive text with models, 

Applications, references, and DEA-Solver Software, 

Kluwer academic publishers, Boston, 2000, pp. 136-138. 

[13] Banker, R. D., Cooper, W. W., Seiford, L. M., Robert, M. 

Thrall, Zhu Joe, ‘Returns to scale in different DEA 

models’, European Journal of Operational Research, vol. 

154, No. 2, 2004, pp. 345–362. 

[14] Thierry Post, Jaap Spronk, Performance benchmarking 

using interactive data envelopment analysis, European 

Journal of Operational Research, Vol. 115, No. 3, 1999, 

pp. 472-487. 

Jing Han is an associate professor in School of Economics and 

Management at Huainan Vocational & Technical College. Her 

major field of study includes electronic commerce, and 

enterprise management (E-mail: hanjing623@163.com).


An Optimal Inventory Control Model for a 

Supply Chain with Shortage Constraints 

Yinkuan Gu 

Management School/Anhui University of Technology, Maanshan, China 

Email: first.author@hostname1.org 

Hongxia Zhang 

Management School/Anhui University of Technology, Maanshan, China 

Abstract—The article studies, with the level constraints in 

short supply, the inventory decision model of the minimum 

total annual cost of the supply chain which, composed of a 

single supplier and multiple buyers, involving supplier’s 

lead time as a decision variable, replacing the cost of 

shortages with the level in short supply, and has solved the 

difficult problem in the practice. 

Index Terms—Level in short supply, supply chain, lead time 


To meet the needs of customers timely, businesses 

must maintain higher inventory levels to avoid shortages. 

However, high inventory levels are often associated with 

high inventory costs, many companies strive to reduce 

production or lead time cycle, and thus have a 

corresponding reduction in inventory. 

The current study on lead time and inventory decisions 

mostly focused on individual enterprises. Liao, C.J., and 

Shyu, C.H.(1992) , for the perpetual inventory system, 

divided the activities such as the procurement, order 

processing, production, transportation, storage test, of the 

lead time period into n-independent component parts and 

each part has its own different time limit as well as 

operating costs, to analyze the best lead time and reorder 

point. Ben-Daya, M, and Raouf, A. set the EOQ into Liao 

and Shyn’s study. Ouyang, LY., Yeh, NC., and Wu, KS., 

discussed the mixture inventory model with backorders 

and lost sales while some customers not wish to wait in 

the situation of out of stock. For being quite difficult to 

assess the unit costs for out of stock in practice, Ouyang, 

LY. and Chuang, BR.(2000) replaced the shortage cost 

with the level of out stock as the parameters of measuring 

shortage. However, the shortening of the lead time 

depends on the improvement and cooperation of the 

upstream and downstream of the supply chain. For this, 

Ben-Daya, M. and Hariga, M. (2004) studied the 

integrated single vendor single buyer model with 

stochastic demand and variable lead time to measure the 

optimal inventory for the perpetual inventory system. 

This paper studies, based on the aforementioned 

literature, with the level constraints in short supply, the 

inventory decision model of the minimum total annual 

cost of the supply chain which, being composed of a 


doi:10.4304/jcp.6.9.1862-1867 

single supplier and multiple buyers, involving supplier's 

lead time as a decision variable, replacing the cost of 

shortages with the level of short supply. 

II. ASSUMPTIONS AND PARAMETERS 

A. Assumptions 

Underlying assumptions of the proposed model in this 

paper as follows: 

(a) The buyers and suppliers are based on the periodic 

inventory system; 

(b) The suppliers take a batch production methods, 

common distribution strategy, and supply in the same 

cycle; 

(c)The production cycle of the suppliers is integral 

multiples to the same provision cycle above; 

(d)The production rate, delivery time, inventory 

holding costs, unit order number, transportation costs, 

order activity cost of the suppliers known as a fixed 

constant; 

(e)The amount of the buyer’s demand of the lead time 

is the same as the amount of the supplier’s requirement in 

the production cycle for a random variable, and following 

the normal distribution; 

(f)The target level of the suppliers and the buyers’ 

order are the average demand of the lead time plus the 

safety stock quantity. 

B. Parameters 

The parameters and their symbols of the model used as 

follows: 

n = The total number of the buyers; 

di = The demand per unit time of the buyer I, average 

2 

as 

d σ 

i 

di 

, variance as 

D = The demand per unit time of the suppler, average 

as 

n 

∑ di 

i= 

1 , variance as 

n 

∑ 

i= 

1 

σ 

2 

d i 

; 

P = Productivity of the suppler(P>D); 

T = The buyer’s common replenishment cycle; 

L = The supplier’s delivery time to the buyer’s orders; 

K = The shipping times of the suppliers in each 

production cycle;


xi = The demand of the buyer i of the warranty period 

(T+L),a random variable, following the probability 

density function fR+L(xi), average as 

2 

d i 

( T + L) 

σ d ( T + L) 

i 

variance as 

; 

y = The demand of a production cycle of the suppler, a 

random variable, following the probability density 

function fKT(xi),average as 

n 

KT ∑ 

i= 

1 

2 

d i 

n 

KT ∑ di 

i= 

1 ,variance as 

σ 

; 

Ri = The target level of replenishment of the buyer I; 

Zi = The safety factor of the of the lead time to the 

buyer i, set as a decision variable; 

Rv = The target level of the supplier’s production; 

Zv = The safety factor of the of the lead time to the 

supplier, set as a decision variable; 

αi = The shortage upper limit of the buyer i; 

αv = The shortage upper limit of the supplier; 

L0 = The buyer’s lead time at the time of the system; 

C (L) = The increased crashing cost of shortening the 

delivery time, a non-increasing function for L,and C (L0) 

=0; 

F = Basic ordering and transportation costs (USD / 

times); 

Fi = The ordering and transportation costs of the buyer 

i (USD / times); 

A = Suppliers’ batch adjustment costs(USD / times); 

hi = The holding costs of a unit of inventory of the 

buyer i (USD / times/unit product/unit time); 

hv = The holding costs of a unit of inventory of the 

suppler (USD / times/a unit product/unit time); 

ECi = The expected total inventory cost per unit time 

of the buyer I; 

ECv = The expected total inventory cost per unit time 

of the suppler; 

ETC = The expected total inventory cost per unit time 

of the supply chain; 

III. MODEL ANALYSIS AND SOLUTION 

A. Model analysis 

Based on the assumptions and parameter settings 

above, this paper establishes the following models: 

The shortage level of the order cycle of the buyer i 

(shortage probability) 

f T + L ( xi 

) dx = Φ( 

Z i ) 

= 

∫Ri ∞ 

The shortage level of the production cycle of the 

supplier (shortage probability) 

f ( y) 

dy = Φ( 

Z ) 

= ∫∞ Rv 

KT 

v 

By the previous assumptions, then there are 

Φ(Zi) αi,Φ(Zv) αv,i = 1,2,…,n 

Items of the related costs associated with the buyer, 

include the expected order cost, transport cost, expected 

holding cost. The order and transport cost of the buyer i is 


, 

Fi/T, the average inventory level of the buyer i can be 

estimated as: 

( Ri − d i L) 

+ ( R i − d i L − d iT 

) 

d iT 

= R i − d i L − 

2 

2 (1) 

Where, the target level of replenishment of the buyer i 

Ri 

= di 

( T + L) 

+ Ziσ 

d T + L 

i 

set as 

, then the expected 

stock holding cost per unit time of the buyer i can be 

⎛ 

⎞ 

⎜ 

d T 

h + Z T + L ⎟ 

⎜ 

i ⎟ 

estimated as ⎝ 

⎠ 

d i 

i 

i σ 

2 

. In addition, while 

the suppliers distribute in a joint way, the buyers share 

the cost of order and transportation per times, and then 

the cost of order and transportation per unit time any 

buyer burdened is F/T. 

To the related storage costs of the suppler, the 

adjustment costs per unit time is A/KT, according to the 

study of Ben-Daya and Hariga (2004)①, the average 

inventory level of the suppler can be estimated as: 

n 

n ⎡ 

⎤ 

∑ d iT 

⎢ ∑ d i 

⎥ 

n 

i = 1 

i = 1 ⎢ ( 2 − K ) + K − 1⎥ 

+ R v − KT ∑ d i 

2 ⎢ P 

⎥ 

i = 1 

⎢ 

⎣ 

⎥ 

⎦ 

(2) 

For the target stock level of the suppler is 

n 

n 

2 

v = KT∑ 

di 

+ Z v KT∑σ 

di 

i= 

1 

i= 

1 

R 

, then the expected 

inventory carrying cost per unit time of the suppler is: 

n 

n 

⎧ ⎡ 

⎤⎫ 

⎪∑ 

d iT 

∑ 

⎪ 

⎪ 

⎢ d i 

⎥ 

n 

i= 

1 i= 

1 

⎪ 

2 

h ⎨ ⎢ ( ) ⎥ 

v 2 − K + K − 1 ⎬ + Z v KT ∑ σ d i 

⎪ 2 ⎢ P 

⎥⎪ 

i= 

1 

⎪ ⎢ 

⎥ 

⎩ ⎣ 

⎦⎪⎭ 

Considering that the lead time of the suppler can be 

adjusted by the extra crashing cost, then the lead time can 

be regard as a decision variable, so the crash cost per unit 

time C(L)/T of shortening the lead time should be 

considered. To the C(L)/T, to be set as a step function in 

the paper, it is, the crash cost will not change in a certain 

range, if beyond the certain range, it will be higher, then 

the function can designed as: 

⎧r0 

L = L0 

⎪ 

r1 

L1 

≤ L < L0 

C( 

L) 

= ⎨ 

⎪M 

M 

⎪ 

⎩rb 

Lb 

≤ L < Lb 

−1 

Where, ri (i = 0,1,…,b) is a parameter for the crash 

cost, Lb is the shortest lead time. 

In summary, there are: the total cost of the system per 

unit time = the expected inventory total cost of the buyer 

per unit time + the expected inventory total cost of the 

suppler per unit time + the crash cost of shortening the 

lead time per unit time. It is: 

①Ben-Daya, M. and Hariga, M., Integrated single vendor single buyer 

model with stochastic demand and variable lead time[J], International 

Journal of Production Economics, 92, 1,2004: 75-80.


n 

C 

( ) 

( L) 

ETC K, 

T , L, 

Z, 

Z v = ∑ ECi 

+ EC v + 

T 

i= 

1 

⎧ n 

n 

n 

⎡ ⎛ ⎞ ⎛ ⎞⎤ 

⎫ 

⎪hv 

∑ d i ⎢ ⎜ ∑ d i ⎟ ⎜ 2∑ 

d ⎟ 

n 

i ⎥ n ⎪ 

1 ⎡ 

⎤ ⎪ i= 

1 

= 

= 

⎪ 

∑ ( ) ⎨ ⎢ ⎜ i 1 ⎟ + ⎜ i 1 

− ⎟ d 

⎥ 

ihi 

= ⎢F 

+ Fi 

+ C L ⎥ + T K 1 − 

1 + ∑ ⎬ 

T ⎣ = ⎦ ⎪ ⎢ ⎜ ⎟ ⎜ ⎟ 

i 1 

2 

P P ⎥ i= 

1 2 ⎪ 

⎪ ⎢ ⎜ ⎟ ⎜ ⎟⎥ 

⎩ ⎣ ⎝ ⎠ ⎝ ⎠⎦ 

⎪ 

⎭ 

n 

+ h Z 

n 

2 

+ hvZ 

v KT∑ 

σ d + ∑ hiZ 

iσ 

d T + L 

i 

i 

i= 

1 i= 

1 

(3) 

By the previous assumption, there are 

Φ(Zi)≦αi,Φ(Zv)≦αv,i = 1,2,…,n 

Equation (3) is the basic model established by this 

paper, the corresponding parameters to the minimum 

ETC are the optimal solution. 

To solve, the formula (3) is added with the slack 

variable Si2 (i= 1,2,…,n), then the following Lagrange 

function can be established as: 

n 

2 2 1 ⎡ A ⎤ 

ETC( 

K, 

T, 

L, 

Z, 

Zv 

, λ, 

λv, 

S , Sv 

) = ⎢F 

+ ∑Fi 

+ + C( 

L) 

T 

⎥ 

⎣ i= 

1 K ⎦ 

⎧ n 

n 

n 

⎡ ⎛ ⎞ ⎛ ⎞⎤ 

⎫ 

⎪hv∑ 

di 

⎢ ⎜ ∑di 

⎟ ⎜ 2∑di 

⎟⎥ 

n ⎪ 

⎪ i= 

1 

= 

= 

⎪ 

+ ⎨ ⎢ ⎜ i 1 

− ⎟ + ⎜ i 1 

− ⎟ d 

⎥ ihi 

T K 1 

1 + ∑ ⎬ 

⎪ 2 ⎢ ⎜ P ⎟ ⎜ P ⎟⎥ 

i= 

1 2 ⎪ 

⎪ ⎢ ⎜ ⎟ ⎜ ⎟⎥ 

⎩ ⎣ ⎝ ⎠ ⎝ ⎠⎦ 

⎪ 

⎭ 

v 

v 

KT 

n 

∑ 

i= 

1 

2 

σ + 

di 

n 

∑ 

h Z σ 

i i di 

i= 

1 

T + L 

n 

2 

2 

+ ∑λi 

[ αi 

−1+ 

Φ( 

Zi 

) + Si 

] + λv 

[ αv 

−1+ 

Φ( 

Zv 

) + Sv 

] 

i= 

1 

(4) 

Where, λ is the Lagrange multiplier, and λ= (λ1, 

λ2,…, λn) ≥ 0,λv ≥0 ,Z = ( Z1,Z2,…, Zn ),S = ( S12, 

S22,…, Sn2). 

Equation (4) is the single-stage supply chain 

inventory decision model with shortage constraints which 

the paper established. 

B. Solving process 

The solution process as follow: 

Take the formula (4) the second derivative for L, and 

get: 

n 

∑ 

i= 

1 

( ) 

( ) 3 

h 

2 

2 

iZ iσ 

di 

∂ ETC 1 ∂ C L 

= − 

2 

2 

∂L 

T ∂L 

4 T + L 

For the C(L) is a step function, to L in each L 

level range there are ∂2C(L)/∂L2 = 0, so to each L level 

range there are ∂2ETC/∂L2 < 0, this means that, with the 

specific K, T, the optimal solution to L is at the endpoint 

of the certain range. And for the slack variable Si2 is 0, 

the prerequisites of the minimum of the expected total 

cost per unit time ETC is that the first derivative is equal 

to zero, that is: 

∂ETC 

= α i −1 

+ Φ( 

Z i ) = 0 

∂λi 

(5) 

∂ETC 

= α v −1 

+ Φ( 

Z v ) = 0 

∂λv 

(6) 

∂ETC 

= hiσ 

d T + L − λiφ( 

Z i ) = 0 

i ∂Z 

i 

(7) 

n 

∂ETC 

2 

= hv 

KT∑ 

σ d − λvφ( 

Z v ) = 0 

i 

∂Z 

v 

i= 

1 

(8) 


By the formula (5) can get: 

Zi* = Φ-1(1-αi), i = 1,2,…,n (9) 


Zv* = Φ-1(1-αv) (10) 

Where, Φ-1(x)is a standard normal inverse function. 

By the formula (7) can get : 

hiσ 

d T + L 

* 

i λi 

= 

> 0, i = 1, 

2, 

⋅⋅ 

⋅, 

n 

φ( 

Zi 

) 

(11) 

* 

λ = 

n 

2 

hv KT∑ 

σ di 

i= 

1 

> 0 


v 

φ( 

ZV 

) (12) 

In addition, take the formula (4) the second derivative 

respectively for Zi, Zv, and can get: 

2 

∂ ETC ⎛ 

= −λ 

⎜ 

2 i − Zi 

∂Z 

⎜ 

i ⎝ 

2 

Zi 

1 − ⎞ 

2 e ⎟ = λiZ 

i 

2π 

⎟ 

⎠ 

2 

Zi 

1 − 

2 e > 0 

2π 

2 

∂ ETC ⎛ 

= −λ 

⎜ 

2 v − Zv 

∂Z 

⎜ 

v ⎝ 

2 

Zv 

1 − ⎞ 

2 e ⎟ = λvZv 

2π 

⎟ 

⎠ 

2 

Zv 

1 − 

2 e > 0 

2π 

Put the Z*, Zv*, λ*, λv* which get from the formula 

(9), (10), (11), (12) into the formula (4) can get: 

ETC 

* * * * 

( K, 

T Z , Z , λ , λ ) 

+ h Z 

* 

v v 

n 

* 

∑λi 

i= 

1 

v 

KT 

n 

∑ 

i= 

1 

v 

1 ⎛ 

= ⎜F 

+ 

T ⎝ 

2 

di 

+ 

∑ 

i= 

1 

n 

* 

∑hi 

Zi 

σ di 

i= 

1 

A ⎞ 

Fi 

+ ⎟ 

K ⎠ 

⎧ n 

n 

n 

⎡ ⎛ ⎞ ⎛ ⎞⎤ 

⎪hv 

∑d 

i ⎢ ⎜ ∑d 

i ⎟ ⎜ 2∑d 

i ⎟⎥ 

⎪ i= 

1 

= 

= 

⎨ ⎢ ⎜ i 1 ⎟ + ⎜ i 1 

+ T K 1− 

−1⎟⎥ 

+ 

⎪ 2 ⎢ ⎜ P ⎟ ⎜ P ⎟⎥ 

⎪ ⎢ ⎜ ⎟ ⎜ ⎟⎥ 

⎩ ⎣ ⎝ ⎠ ⎝ ⎠⎦ 

σ 

T + L 

* * 

* 

[ α −1+ 

Φ( 

Z ) ] + λ [ α −1+ 

Φ( 

Z ) ] 

n 

n 

∑ 

i= 

1 

⎫ 

⎪ 

dihi 

⎪ 

⎬ 

2 ⎪ 

⎪ 

⎭ 

+ 

i 

i v v 

v 

(13) 

With the specific K, take the formula (13) the first 

derivative, set it as 0 and can get: 

n 

n 

n 

⎡ ⎛ ⎞ ⎛ ⎞⎤ 

h 

2 

1 

⎢ ⎜ ⎟ ⎜ ⎟ 

∂ ⎛ ⎞ ∑d 

∑d 

∑d 

n 

v i 

i 

i 

ETC 

A 

⎥ 

i= 

1 ⎢ ⎜ i= 

1 

1 

1 ⎟+ 

⎜ i= 

= − ⎜F 

+ + ⎟+ 

− 

−1⎟⎥ 

2 ∑Fi 

K 

∂T 

T ⎝ 1 ⎠ 2 ⎢ ⎜ ⎟ ⎜ ⎟ 

i= 

K 

P P ⎥ 

⎢ ⎜ ⎟ ⎜ ⎟⎥ 

⎣ ⎝ ⎠ ⎝ ⎠⎦ 

F + 

n 

∑ 

T 

+ 

F + 

i 

i= 

1 

2 

n 

∑ 

i= 

1 

A 

K 

d Z ihi 

+ 

2 

* 

v 

n 

∑ 

n 

∑ 

2 

* 

K σd 

hiZ 

iσ 

i 

di 

i= 

1 i= 

1 + = 0 

2 T 2 T + L 

n ⎡ ⎛ 

hv∑ 

di 

⎢ ⎜ 

i= 

1 

= ⎢K⎜1 

− 

2 ⎢ ⎜ 

⎢ ⎜ 

⎣ ⎝ 

n 

∑ 

n 

∑ 

i= 

1 

* 

d h Z i i v 

i= 

1 

+ + 

2 

2 

K σ di 

i= 

1 

2 T 

+ 


n 

P 

∑ 

d 

i 

⎞ ⎛ 

⎟ ⎜ 2 

⎟ + ⎜ 

⎟ ⎜ 

⎟ ⎜ 

⎠ ⎝ 

n 

∑ 

i= 

1 

n 

∑ 

i= 

1 

P 

d 

* 

h Z σ 

i 

i 

i 

⎞⎤ 

⎟⎥ 

−1⎟⎥ 

⎟⎥ 

⎟⎥ 

⎠⎦ 

di 

2 T + L 

(14)


n 

⎧ n 

⎛ A ⎞ 

⎡ ⎛ 

2⎜ 

F + ∑ Fi 

+ ⎟ ⎪hv 

∑d 

i 

= 

⎪ 

⎢ ⎜ 

⎝ i 1 K ⎠ 2 i= 

1 = ⎨ ⎢K⎜1 

− 

3 

T T ⎪ 2 ⎢ ⎜ 

⎪ ⎢ ⎜ 

⎩ ⎣ ⎝ 

+ Z 

* 

v 

K 

n 

∑ 

i= 

1 

⎧ n ⎡ ⎛ 

⎪hv 

∑d 

i 

⎪ 

⎢ ⎜ 

2 i= 

1 ≥ ⎨ ⎢K⎜1 

− 

T ⎪ 2 ⎢ ⎜ 

⎪ ⎢ ⎜ 

⎩ ⎣ ⎝ 

n 

σ T 

2 

di 

∑ 

i= 

1 

3 

− 

2 

n ⎞ ⎛ ⎞⎤ 

di 

⎟ ⎜ 2∑di 

⎟⎥ 

⎟ ⎜ i= 

1 + −1⎟⎥ 

+ 

P ⎟ ⎜ P ⎟⎥ 

⎟ ⎜ ⎟ 

⎠ ⎝ ⎠ 

⎥ 

⎦ 

n 

∑ 

i= 

1 

+ 

n 

* 

∑hi 

Zi 

σ di 

i= 

1 

n ⎞ ⎛ ⎞⎤ 

di 

⎟ ⎜ 2∑d 

i ⎟⎥ 

⎟ ⎜ i= 

1 + −1⎟⎥ 

+ 

P ⎟ ⎜ P ⎟⎥ 

⎟ ⎜ ⎟ 

⎠ ⎝ ⎠ 

⎥ 

⎦ 

n 

3 

− 

n 

2 2 

* 

d T + ∑hi 

Z 

i 

i σ di 

i= 

1 

− ( T + L) 

n 

∑ 

i= 

1 

1 

2 

T 

⎫ 

dihi 

⎪ 

⎬ 

2 ⎪ 

⎪⎭ 

n 

∑ 

i= 

1 

−1 

⎫ 

dihi 

⎪ 

⎬ 

2 ⎪ 

⎪⎭ 

3 

* 

− 

+ Zv 

K∑σ 

( T + L) 

2 

i= 

1 

(15) 

Take the formula (13) the second derivative for T and 

can get: 

n 

n 

n 

⎛ A ⎞ * 

2 

* 

2⎜ 

F + 

2 ∑Fi 

+ ⎟ Zv 

K∑σ 

d T ∑Zi 

h 

i 

i di 

∂ ETC * * * * 

= 1 

= 1 

= 1 

, , , = 

⎝ i K ⎠ 

i 

i 

Z 

− 

− 

2 i Zv 

λ λv 

3 

∂T 

T 

4 

4 

3 

2 

σ ( T + L) 

Put the formula (16) into (15) then can get: 

⎧ n 

n 

n 

⎡ ⎛ ⎞ ⎛ ⎞⎤ 

2 

⎪hv 

2 

* * * * 2 ⎪ 

⎢ ⎜ ⎟ ⎜ ⎟ 

∂ 

∑di 

∑di 

∑di 

ETC 

⎥ 

i= 

1 

1 

1 

, , , 

1 

1 

2 

⎨ ⎢ ⎜ i= 

⎟+ 

⎜ i= 

Z = 

− 

− ⎟⎥ 

i Zv 

λ λv 

K 

+ 

∂T 

T ⎪ 2 ⎢ ⎜ P ⎟ ⎜ P ⎟⎥ 

⎪ ⎢ ⎜ ⎟ ⎜ ⎟⎥ 

⎩ ⎣ ⎝ ⎠ ⎝ ⎠⎦ 

3 

2 

(16) 

n 3 

− 

n 3 

3 

* 

2 2 

* 

− 

+ Z 

+ 

( + ) 2 

v K∑σ 

d T ∑hi 

Ziσ 

d T L > 0 

i 

i 

4 i= 

1 4 i= 

1 

This means that the solution T which get form the 

formula (14) must be a local minimum, furthermore, 

since the values of this second derivative is always 

positive, means for equation (13), with the particular K, 

the optimal solution to T get by formula (14) was the only 

solution. 

3 

n 

∑ 

i= 

1 

⎫ 

dihi 

⎪ 

⎬ 

2 ⎪ 

⎪⎭ 

In summary, the following algorithm can be taken to 

calculate the optimal solution to K, T, L, Z, Zv, λ, λv of 

the model: 

Step 1, set K=1, L=Li, i=0, 1,…,b. 

Step 2, get the solution to Z*, ZV* by the (9) and (10). 

Step 3, get the T*(K, Li) by the (14). 

Step 4, put the Z*, ZV*, T*(K, Li), Li into equation 

(11) and (12) and get λ*(K, Li), λv*(K, Li). 

Step 5, put the Z*, ZV*, T*(K, Li), Li, λ*(K, Li), 

λv*(K, Li) into the formula (13) to get 

( ) 

* 

* 

* 

* 

* 

ETCi ( K, 

Li 

) = ETCT 

( K, 

Li 

) , Z ( K, 

Li 

) , Zv 

( K, 

Li 

) , λ ( K, 

Li 

) , λv 

( K, 

Li 

) , Li 

, K 

Step 6, set ETCi(K, Li) = Min[ETCi(K, Li)], i=0,1,… 

,b. 

Step 7, if K=1, set ETCs = ETC(K, L), K=K+1, go 

back to step 3, then if ETC(K, L)< ETCs, ETCs = 

ETC(K, L), K=K+1, go back to step 3, otherwise, set 

K*=K-1, L*=L, T*=T*(K*, L), λ*=λ*(K*, L), 

λv*=λv*(K*, L), ETC*= ETC s, and the solution is over. 

IV. MODEL APPLICATION EXAMPLE 

Set the C(L) as: 

⎧0 

L = 0. 

02 

⎪ 

5 0. 

01 ≤ L < 0. 

02 

C( 

L) 

= ⎨ 

⎪11 

0. 

005 ≤ L < 0. 

01 

⎪ 

⎩18 

0.002 ≤ L < 0. 

005 

Other parameters shown in Table 1 and Table 2 as 

below: 

Table 1 The buyer’s parameters 

d = 6000 unit/year d = 5000 unit/year d = 10000 unit/year 

1 

2 

σ = 600 unit/year σ = 800 unit/year σ = 900 unit/year 

d1 

d 

2 

F1 = 100 USD/times F2 = 150 USD/times F3 = 80 USD/times 

h1 = 5 USD/unit/year h2 = 4 USD/unit/year h3 = 4.5 USD/unit/year 

α1 = 0.01 α2 = 0.01 α3 = 0.01 

F = 100 USD/times 

Table 2 The supplier’s parameters 

P = 28000 unit/year A = 200 USD/patch hv = 3 USD/unit/year αv = 0.01 

The results by calculation shown in Table 3 and Table 4 as below: 

K 

L * (K) 

T * (K) 

Z1 * 

Z2 * 

Z3 * 

Zv * 

λ1 * (K) 

λ2 * (K) 

λ3 * (K) 

λv * (K) 

ETC * (K) 


Table 3 The solution to the parameters 

1 2 3 4 5 

0.002 0.002 0.002 0.002 0.002 

0.07338 0.06390 0.05934 0.05626 0.05387 

2.326 2.326 2.326 2.326 2.326 

2.326 2.326 2.326 2.326 2.326 

2.326 2.326 2.326 2.326 2.326 

2.326 2.326 2.326 2.326 2.326 

30905.4 28896.3 27879.1 27169 26606.9 

32965.8 30822.7 29797.7 28980.3 28380.7 

41722.3 39010 37636.8 36678.2 35919.3 

41023.7 54138.3 63897.1 71838.6 78596.4 

23140.2 23101.9 23695.7 24392 25096.5 

3 

d 

3


Table 4 The optimal solutions to the model 

K 

* L * T * Buyer 1 

R1 S1 

Buyer 2 

R2 S2 

Buyer 3 

R3 S3 

supplier 

Rv Sv 

ETC 

2 0.002 0.0639 754 1 808 1 1197 1 3803 2 23101.9 

Note: S1, S2, S3 for the buyer’s expected shortages in the replenishment cycle, Sv for the supplier’s expected shortages in the 

delivery period. 

From the table 3 we can get the conclusion that the [8] Reve, T., and Johansen, E. Organizational Buying in the 

total cost curve shows the convex trends: The optimal Offshore Oil Industry. Industry Marketing Management, 

shipping times per production cycle for the manufacturer 

is 2 times. However, compared to the case of not 

considering the shortening delivery time (L*=L0), the 

best common replenishment cycle of the buyer T* is 

reduced by the 0.06414 years to 0.06390 years, the best 

replenishment lead time of the buyer L* is reduced to 

1982,(11):275-282. 

[9] Silver, E. and Peterson, P., Decision System for Inventory 

Management and Production Planning, 2nd ed., Wiley, 

New York, 1985. 

[10] Miller, D. and Dorge, C. Psychological and Traditional 

Determinants of Structure Administrative Science 

Quarterly, 1986, 31: 539-560. 

0.002 years. In addition, the table 4 shows that, all the [11] Malone, W., Yates, J. and Benjamin, I. Electronic Markets 

safety factors Z1*, Z2*, Z3* of the lead time of the buyer and Electronic Hierarchies. Communications of the ACM, 

1, 2, 3, are 2.326, 99% service level, that is, the buyers’ 1987,30(6):484-497. 

replenishment target level per cycle is 753, 808, 1197 

units, the expected shortages per cycle is 1 unit. The 

safety factor Zv* of the supplier’s lead time is 2.326 too, 

that is, the supplier’s replenishment target level per cycle 

is 3803 units, the expected shortages per cycle is 2 units. 

At the time, the optimal total cost of the supply chain per 

[12] Novack, A., and Simco, W. The Industrial Procurement 

Process: A Supply Chain Perspective. Journal of Business 

Logistics, 1991,12,(1):145-167. 

[13] Liao, C.J., and Shyu, C.H., An analytical determine of lead 

time with normal demand, International Journal of 

Operations and Production Management, 1991, l7, 4:115- 

124. 

unit time is $23101.9, go down 2.2% compared to case of [14] Ballou, R. H., Business Logistics Management, 3rd ed., 

the supplier not allow to shorten the delivery. 

Prentice-Hall, Englewood Cliffs, NJ, 1992. 

[15] Davis, T. Effective Supply Chain Management. Sloan 

V. CONCLUSION 

Man-agement Review (Summer), 1993: 35-46. 

[16] Lee, H. L. and C. Billington, Material management in 

To the supply chain management, the lead time has a decentralized supply chains. Operations 

great impact on inventory management performance. Research,1993,41(5): 835-847. 

However, the majority of supply chain inventory decision 

model regarded the lead time as being fixed, this does not 

correspond with the practice. Innovations of this paper is 

that, set the lead time as a variable, with the constraints of 

meeting different shortages, explores the inventory 

decision model of a supply chain which composed of a 

single supplier and multiple buyers. While the earlier 

[17] Gerwin, D. Manufacturing Flexibility: A Strategic 

Perspective. Management Science, 1993, 39(4): 395-410. 

[18] Li, K., Shyu, T. and Adiga, S. A Heuristic Rescheduling 

Algorithm for Computer-Based Production Scheduling 

Systems. International Journal of Production Research, 

1993, 31(8): 815- 1826. 

[19] Sethi, V. and Carraher, S.M., Developing measures for 

assessing the organizational impact of information 

calculations appear more complicated, if programmed, technology: a comment on Mahmood and Soon’s paper. 

the application will be very simple. The inadequacy of Decision Sciences, 1993, 24: 867-77. 

the study is that the model in this article are not compared 

with other models and sentenced to the merits, this is also 

the further research content for the author. 

[20] Banerjee, A. and Banerjee, S., A coordinated order-up-to 

inventory control policy for a single supplier and multiple 

buyers using electronic data interchange, International 

Journal of Production Economics, 1994, 35, 1-3: 85-91. 

REFERENCES 

[21] Ben-Daya, M, and Raouf, A., Inventory models involving 

lead time as a decision variable, The Journal of the 

[1] Cronbach, L. J., & Meehl, P. E..Construct validity in 

psychological tests. Psychological Bulletin, 

1955,52:281302. 

[2] Duncan, B. Characteristics of Organizational 

Environmental and Perceived Environmental Uncertainty. 

Administrative Science Quarterly,1972, 17: 313-327. 

[3] Lawshe, H. A Quantitative Approach to Contect Validity. 

Personnel Psychology, 1975, 28: 563-575. 

[4] Lawshe, H. A Quantitative Approach to Contect Validity. 

Personnel Psychology, 1975, 28: 563-575. 

[5] Whybark, C., and William, J. Material Requirement 

Planning under Uncertainty. Decision Science, 1976,17(4): 

595-606. 

[6] Mintzberg, T.Patterns in Strategy Formation. Management 

Science,1978,(24):934-948. 

[7] Churchill, A. A Paradigm for Developing Better Measures 

of Marketing Constructs. Journal of Marketing Research, 

1979,16(1):64-73. 

Operational Research Society, 1994,45(5): 579-582. 

[22] Tersine, R. J., Principles of Inventory and Materials 

Management, 4th ed., 1994. 

[23] Kotteaku, G., Laios, G. and Moschuris, J. The influence of 

product complexity on the purchasing structure. 

International Journal of Management Science, 1995, 

23(1):27-39. 

[24] Krajewski, L.J. and Ritzman, L.P., Operations 

Management: Strategy and Analysis. Addison Wesley, 

Reading, MA, 1996. 

[25] Ouyang, LY., Yeh, NC., and Wu, KS., Mixture inventory 

model with backorders and lost sales f or variable lead 

time, Journal of Operational Research Society, 1996,47: 

829-832. 

[26] Washington, D.C., Statistical Abstract of the United States 

U.S. Bureau of the Census: 1996, 116th ed., 443, 1996. 

[27] Fisher, M., What is the Right Supply Chain for Your 

Product . Harvard Business Review, 1997, 75(2): 105-116. 



[28] Calinescu, A., Efstathiou, J., Schim J., and Bermejo, J. 

Applying and Assessing Two Methods for Measuring 

Complexity in Manufacturing. The Journal of the 

Operational Research Society, 1998, 49: 723-733. 

[29] Hair, F., Anderson, E., Tatham, L., and Black, C. 

Multivariate Data Analysis. Prentice Hall, New Jersey, 

1998. 

[30] Shawnee, V., C. Roger and D. Cornelia. Supply chain 

flexibility: an empirical study. Journal of Supply Chain 

Management,1999,35(3):16-24. 

[31] Vickery, V., Roger, C. and Cornelia, D. Supply Chain 

Flexibility: An Empirical Study. Journal of Supply Chain 

Management,1999, 35(3):16-24. 

[32] Khurana, A. Managing Complex Production Processes. 

Sloan Management Review, 1999, 40(2): 85-97. 

[33] Zhu, L. and Soh, C. FMS Job-Shop scheduling under 

disruptions with consideration of time and sequence 

deviation. IEEE International Symposium on Intelligent 

Control/Intelligent Systems and Semiotics, 1999: 138-143. 

[34] Ouyang, LY. and Chuang, BR., Stochastic Inventory 

Model Involving Variable Lead Time with a Service Level, 

Yugoslav Journal of Operations Research, 2000, 10, 1: 81- 

98. 

[35] Chopra, S. and Meindl, P. Supply Chain Management: 

Strategy, Planning, and Operation. Prentice-Hall, New 

Jersey, 2001. 

[36] Subramaniam, C., and Shaw, M. A Study of the Value and 

the Impact of B2B E-Commerce: The Case of Web-Based 

Procurement. International Journal of Electronic 

Commerce, 2002, 6(4): 19-40. 

[37] Gatignon, H., Tushman, L., Smith, W., and Anderson, P.A 

Structural Approach to Assessing Innovation: Construct 


Development of Innovation Locus, Type, and 

Characteristics. Management Science,2002,48(9):1103- 

1122. 

[38] Taha, H. A., Operations Research, 2003. 

[39] Yang, B. Y., Watkins, K. E., & Marsick, V. J.The 

construct of the learning organization: Dimensions, 

measurement, and validation. Human Resource 

Development Quarterly, 2004, (15): 31-55. 

[40] Ben-Daya, M. and Hariga, M., Integrated single vendor 

single buyer model with stochastic demand and variable 

lead time, International Journal of Production Economics, 

2004, 92, 1:75-80. 

[41] Yao, Yuliang ;Dong, Yan ;Dresner, Martin, Managing 

supply chain backorders under vendor managed inventory: 

An incentive approach and empirical analysis, European 

Journal of Operational Research, 2010, 6,31(2):350-359. 

[42] Ahmadi Javid, Amir ;Azad, Nader, Incorporating location, 

routing and inventory decisions in supply chain network 

design, Transportation Research - Part E the Logistics and 

Transportation Review,2010,46,5:582-597. 

[43] Kogan, K ;Lou, S ;Tapiero, C S ;Shnaiderman, M, Supply 

Chain With Inventory Review and Dependent Demand 

Distributions: Dynamic Inventory Outsourcing, IEEE 

Transactions on Automation Science and Engineering, 

2010,7,2:197-207. 

[44] Battini, D. ;Gunasekaran, A. ;Faccio, M. ;Persona, A. 

;Sgarbossa, F., Consignment stock inventory model in an 

integrated supply chain, International Journal of 

Production Research, 2010, 119, 2: 477-500. 

[45] Harish Krishnan;Ralph A. Winter, Inventory Dynamics 

and Supply Chain Coordination, Management 

Science,2010,56,1:141.


Variable Selection for Credit Risk Model Using 

Data Mining Technique 

Kuangnan Fang 

Department of Planning and statistics/Xiamen University, Xiamen, China 

Email: ruiqwy@163.com 

Hong Huang * 

Economics Department/Hefei Normal University, Hefei, China 

Email: HH6@263.net 

Abstract—With the emergence of the current financial crisis, 

societies see the increasing importance of credit risks 

management in financial institutions. Four mainstream 

credit risk rating models have been developed, however, 

their applicability in the Taiwan market is yet to be 

evaluated. In this paper, six major credit risk models, 

including Merton Option Pricing Model,Discriminant 

Analysis Model, Logistic Regression (Logit) Model, Probit 

Model, Survival Analysis Model, and Artificial Neural 

Network Model were examined, in order to identify the 

common variables applicable to each model. The common 

variables were then applied to each respective model 

directly. Using Transition Matrix and mapping methods to 

estimate long term default probability, for developing 

appropriate credit risk model with the estimated default 

probability. 

Index Terms—Credit Default Risk; Logit; Logistic 

Regression Model 


In recent years, with the development of global credit 

portfolio management, continuous innovations in 

financial credit derivatives and financial statistical 

techniques, the growth on awareness of credit risks 

among financial institutions and regulatory authorities, 

both practical and theoretical research and development 

of credit risk evaluation models are given high 

importance and under vigorous progress. Seeing the 

vitality of considering credit risks in financial institutions, 

The New Basel Capital Accord focuses on strengthening 

the risk management mechanism of banks by requiring 

banks to establish a sound internal risk assessment 

mechanism and to increase the responsibility of the 

exte ∗ rnal supervisory bodies. The new accord encourages 

financial institutions to establish their own credit rating 

mechanisms; however, it has allowed flexibility in 

choosing which credit risk model to use. At present, there 

exists several developed credit risk models; each has its 

own theoretical basis and advantages. Further discussion 

is required to investigate whether a particular model is 

applicable to the Taiwan market, or, in other words, 

∗ Corresponding author of this article. 


doi:10.4304/jcp.6.9.1868-1874 

whether it is applicable globally or it should be adjusted 

according to local factors. 

With the flexibility towards credit risks allowed in the 

new accord, in this paper, we shall analyze the six major 

credit risk models, including Merton Option Pricing 

Model, Discriminant Analysis Model, Logit Model, 

Probit Model, Survival Analysis Model and Artificial 

Neural Network Model, to identify common variables 

applicable to each model based on the financial 

statements of companies in Taiwan and market data. The 

common variables can then be applied to each respective 

model directly, in order to establish an appropriate credit 

risk model with the estimated default probability. 

II. MAJOR CREDIT RISK MODELS 

A. Credit Metrics Model 

Credit Metrics Model was developed by J.P. Morgan 

in 1997. It mainly uses the technique of migration 

analysis and Value-at-Risk to look at the credit risks 

arising from credit ratings changes of credit assets in the 

investment portfolio. 

Credit Metrics Model mainly depends on historical 

average default rates and the credit rating transition 

matrix. First, it estimates the probability of transitions 

between risk groups based on historical data, and at the 

same time establishes the correlation between credit 

ratings and the value of a debtor company's asset, so as to 

determine the joint migration behavior of credit qualities 

among the asset portfolios. Then, portfolio default loss 

distribution can be generated by looking at the market 

value changes of asset portfolio in the Monte Carlo 

simulation of quality transitions. Eventually, the value of 

a single loan or loan portfolio can be calculated. The 

model has high applicability as it can be applied to a wide 

variety of financial products, such as bonds, loans, loan 

commitments, accounts receivable, letters of credit, as 

well as financial derivatives. However, it emphasizes the 

assumption that all counterparties within the same risk 

group have the same degree of credit risk. In addition, in 

determining the credit transition matrix probability, the 

model does not adjust properly according to the 

prevailing economic conditions. Therefore, there are


often gaps between estimation results and empirical 

results. 

B. KMV Model 

The KMV model is proposed by KMV Corporation 

based on the Merton Model. It defines the "distance to 

default" which indicates the distance between a 

company's asset value and the default point. The greater 

the distance, the smaller the default probability will be. 

On the other hand, the smaller the distance, the greater 

the default probability of the company's assets will be. In 

other words, default will occur when the company's asset 

value is lower than the default point. However, different 

from Merton model, KMV discovers that the company 

has refinancing abilities in real practices; therefore 

default may not necessarily occur when asset value is 

lower than the book value of liabilities. According to 

KMV, the real default point is usually somewhere 

between the value of total liabilities and the value of 

current liabilities. For normalization, the distance-todefault 

is indicated as the number of standard deviations 

between the company's asset value and the default point. 

Then, by mapping the distance-to-default to the Expected 

Default Frequency (EDF), the EDF can be calculated. 

KMV Corporation has accumulated a large database 

which is used to estimate correlations between default 

probabilities and corporate defaults. Based on these 

correlations, credit ratings transition matrix and default 

loss distribution of the debtor can then be further derived. 

Instead of relying on the credit ratings transition matrix, 

the KMV approach tracks the market conditions and 

incorporates the company's financial data and market data 

in the model to accurately grasp the credit risk changes of 

the asset components. In addition, the accuracy of the 

prediction from the model is enhanced by its ability of 

directly calculating the EDF of the company. However, 

the model assumes the company's asset value changes 

follow the normal distribution and does not consider the 

volatility of liabilities. 

C. Credit Risk+ Model 

Credit Risk+ is a default model proposed by Credit 

Suisse Financial Products (CSFP) in 1996. It is mainly 

based on an actuarial approach to derive the loss 

distribution of bonds or loans portfolio, and calculate the 

credit loss provision. The basic hypothesis is that default 

loss occurs when many debtors default, and each debtor's 

default probability is the same and very small. Therefore, 

the number of defaults in the asset portfolio can be 

estimated in accordance with the Poisson distribution, 

while the default probabilities depend upon a gammadistributed 

set of risk factors and will change over time. 

The model is based on a basic assumption that the 

number of defaults in the portfolio follows a Poisson 

distribution, and uses the volatility of default probabilities 

to reflect the influences of default correlation. Through 

statistical analysis of default rates and recovery rates of 

defaulted loans, loans of common default loss 

characteristics are put under same groups to derive the 

probability function of loss distribution. Then the future 

loss distribution of the portfolio will be estimated and 


eventually, the expected and non-expected losses of the 

portfolio can be obtained. The model makes no 

assumption for the reason of default risks and requires 

small amount of data. It has also taken into consideration 

of volatility of default probabilities in the process of 

calculation. However, the model assumes credit 

exposures are fixed and regarded as a constant. The 

model also does not take into account the risks of rating 

changes. 

D. Credit Portfolio View Model 

The basic theories of Credit Portfolio View were 

published by McKinsey & Company in 1997. The main 

characteristics of the model are that it assumes the 

probabilities of default occurrence and credit quality 

changes are closely related to the overall economic 

conditions. In general, many credit risk models assume 

that default occurrence is a result of individual financial 

health of the specific company. However, empirical 

findings show that the probabilities of default and rating 

migration of a company fluctuate with the business cycle. 

When economic conditions worsen, the default 

probability of a company default increases accordingly, 

and vice versa. In other words, credit cycles and 

economic cycles are closely correlated. The model 

mainly uses the following process to assess the credit risk 

of a company: set up a multi-factor model which measure 

systematic risks to determine the economic conditions; 

then evaluate the default probability of a company with 

the Logit Model. By modeling the relationship between 

credit ratings transition matrix and macroeconomic 

factors such as economic growth rate, default loss 

distribution is derived. The model assumes that default 

probabilities are related to the overall economic 

conditions, which is in line with the reality. In the 

process of calculation of credit risk, it uses the actual 

discrete distribution of the portfolio, which is more 

accurate than using continuous distribution, and is able to 

assess the credit risks of liquid and non-liquid assets at 

the same time. However, the selection of economicfinancial 

factors may be subjectively influenced, and 

important economic factors could be missed out in the 

evaluation process, resulting in overestimation or 

underestimation. 

III. RESEARCH METHODOLOGY 

For the purpose of accuracy and applicability, we first 

use six models, namely Merton Option Pricing Model, 

Discriminant Analysis Model, Regression Analysis 

Model, Logit Model and Probit Model, Survival Analysis 

Model, Artificial Neural Network Model, to establish a 

credit risk scoring model with the best variables set and 

common variables set. Among the six models, only 

Merton Option Pricing Model uses the market approach, 

while the other five models uses the actual approach. The 

common characteristic of actual approach models is that 

they require historical financial data for modeling. The 

selection of variables to be used in the model is another 

concern. We shall first select the variables that can be 

input into the model, then, among these selected ones,


choose the best variables set using statistical methods, 

and apply the common variables to each model for 

comparing their results of differences. 

Having derived results from the above evaluation 

model, in order to find out a reasonable default 

probability, a bank will usually use a quantitative 

approach to modeling. When using a quantitative 

approach for modeling, attention should first be made to 

whether the selected variables are suitable for estimating 

default probabilities. The bank must prove that the 

selected estimation variables have significant correlation 

with default probabilities. It should adopt a statistical 

method to prove if the selected variables have significant 

explanatory power of default probabilities. To this end, 

the most common statistical method is to build a scoring 

system based on the regression approach. After the 

scoring system is established, the bank must rank and 

grade the rating of each exposure of its investment or 

loan portfolio. According to the New Accord, there 

should be at least seven grades of rating so as to prevent 

over-concentration of risks. In this paper, we first 

establish the required scoring model, then quantified the 

ratings by mapping method to derive default 

probabilities; the results are validated with benchmarking. 

IV. EMPIRICAL ANALYSIS 

A. Sampling 

In this study, sample selection criterion is that the 

company has to be publicly listed as of December 2010. 

Accordingly, credit clients' data between January 2001 

and December 2010 have been collected from banks in 

Taiwan as samples. The financial information used in this 

study are mainly combined statements, supplemented by 

individual statements. 10,032 observations have been 

collected, excluding data with omission,which include 

285 default cases and 9747 normal companies. In order to 

apply to the model, we classify the samples into training 

samples and valid samples. The sample distribution is 

summarized in table 1. 

Table 1: Sample distribution 

training 

samples 

Valid 

samples 

Total 

number 

normal 

companies 

5,604 4,143 9,747 

default cases 153 132 285 

Total number 5,757 4,275 10,032 

B. Selection of Variables 

(a).Selecting Common Variables 

There are more than a hundred variables generated 

from a company's financial statement analysis; however, 

it is doubtful if each variable can be used to explain the 

default occurrence of the company. Therefore, we will 

first make reference to the variables selection of famous 

research institutions in Taiwan and around the world, as 

well as those adopted by representative papers. 

The industry characteristics and sampling quantities 

adopted by the Taiwan Corporate Credit Risk Index 

(TCRI), which evaluates public companies (non- 


financial), conform to the research requirements of this 

paper. Therefore, we have made reference to the 

variables used in the TCRI rating. According to the 

TCRI, a good company should be profitable, with asset 

management efficiency, sound financial planning and a 

market leader. Accordingly, we use the following four 

dimensions of financial indicators: profitability, 

efficiency, security and size. 

The Falkenstein (2000) model uses variables from six 

dimensions (profitability, security, size, liquidity, 

efficiency and growth ability) and compares the 

correlation between financial ratios and default 

probability under each dimension to choose the most 

suitable variables to be used in the model. Finally, 10 

financial ratios are chosen to build the statistical model. 

In 1968, Edward I Altman used a number of variables 

to conduct estimation for company failures. There were 

22 financial ratios used for validation, including liquidity, 

profitability, financial leverage, repayment capability and 

efficiency. Eventually, the five ratios with best predictive 

power were selected for the statistical modeling. 

According to the empirical experience of TCRI in 

credit rating, the credit risks of a company are not 

completely reflected in the financial ratios and many risks 

are actually reflected in non-financial data. Therefore, in 

this paper we will also consider the following variables: 

opinions of accountants, related-parties purchase-sales 

ratio, directors' pledge ratio, P/E ratio, P/B ratio and 

compound Return on Equity. 

Due to various reasons, a company may adopt financial 

ratios to make its books look better. If this is the case, we 

may not find out the real situation about the company by 

judging the financial ratios only. For example, a 

company may borrow in the name of its subsidiary by 

endorsing the loan. In light of the consideration that 

financial ratios may not reflect the real stories, in this 

research, we also calculate the "adjusted" financial ratios, 

including: recurring net profit, debt-to-equity ratio, longterm 

profitability indicators. 

(b). Selecting best variables for each model 

Even after the above variables screening, we still come 

up with a large number of variables. Each of these 

variables may not necessarily has explanatory power 

about our sample companies; besides, if there are too 

many variables included in the model, the model will 

become too complicated, and collinearity problem among 

variables may arise, leading to unreasonable estimation of 

parameters. In addition, the variables that can explain 

default probabilities may vary among different models. 

Therefore, in the following, we will evaluate the variables 

suitable for each model so that models with the best 

statistical explanatory power can be built. 

Regarding Logit Model, Probit Model and Survival 

Analysis, in the process of variables selection, first we 

put the independent variables into two groups, depending 

on whether the company is defaulting or not; then we use 

the SPSS (software for quantitative data analysis) to carry 

out two-tailed T-tests for independent samples. At the 

confidence coefficient of 0.05, mean differences are 

tested and variables with significant differences (P-


Value


Table 3: Table of Common Variables 

Dimension Selected Variance 

ROA (Before Interest 

Profitability 

and Tax) 

Compound ROA 

Repayment Capability Quick Ratio 

Activity Total Asset Turnover 

Growth Revenue Growth 

Financial Structure Debt / Equity 

A good independent variable shall have significant 

explanatory power on the dependent variables. We want 

to conduct a statistical analysis to validate if the selected 

variables have good explanatory power over the default 

probability. The most common method used for such 

purpose is regression; therefore, we will use regression to 

Selected Variables 

validate the variables. As to the estimation of default 

probabilities, we have made reference to the research by 

Xue Ren-rui, Liu Ying-feng.After we have obtained the 

default probabilities, we have conducted simple 

regression analysis to the variables using the weighted 

least squares method, so as to find out if each variable has 

significant explanatory power to the default probabilities. 

Table 4 is the analysis results. From Table 4, we can see 

that other than revenue growth, all variables can 

significantly explain the default probabilities. Through 

the above selection of variables, from the variables that 

we have selected arbitrarily, we can find out those that 

are proven by statistical inference to be closely related to 

default probabilities. 

Table 4: Analysis of simple regression of variables 

Coefficie 

nt 

T- 

value 

P- 

value 

Includ 

/Exclud 

ROA 

(Before Interest and Tax) 

-0.0066 -3.77 0.00 Includ 

Compound ROA -0.0069 -2.70 -0.01 Includ 

Quick Ratio -0.0012 -5.81 -0.00 Includ 

Total Asset Turnover -0.0066 -3.77 -0.00 Includ 

Revenue Growth -0.0004 -1.49 0.14 Exclud 

Debt / Equity 0.0012 4.11 0.00 Includ 

As "Industry" is a dummy variable, regression cannot 

be used to verify its correlation with the default 

probabilities. In this research we assume it is a 

significant variable without testing. 

In order to test if collinearity exists among the selected 

variables, we also look at the VIF values. The VIF values 

of the variables are calculated and listed in Table 5. 

Table 5: Table of VIF of Selected Variables 

Selected Variables 

ROA 

VIF Value Included/Excluded 

(Before Interest and 

Tax) 

2.3122 Included 

Compound ROA 2.1243 Included 

Quick Ratio 1.3848 Included 

Total Asset Turnover 1.2623 Included 

Debt / Equity 1.3717 Included 

From Table 5, only Return on Assets Ratio (Before 

Interest and Tax) and Compound Return on Assets Ratio 

generate higher VIF values. Based on the usual standard 

that VIF less than 10 will not jeopardize the parameter 

estimates, we infer that there are no multi-collinearity 

issues with the above variables. 

(d). Validation 

We have conducted regression to the best variables and 

common variables of each model ,we can define the 

accuracy ratio of each model as the numbers of 

companies whose regression results are in conformity 

with physical facts divided the numbers of observations 

(10,032) .In order to verify the efficiency of the variables 

in the models, we have listed out the accuracy ratios of 

the models with the best variables set and common 

variables set respectively in Table 6 and Table 7: 

Table 6: Accuracy Ratio of Each Model with the Best Variables Set 

Model 

Merton 

Artificial 

Accuracy 

Logit Probit Discriminant Hazard 

Option 

Neural 

Ratio 

Regression Regression Analysis Ratio 

Pricing 

Network 

In-Sample N/A* 0.9566 0.9556 1.0000 0.8044 1.0000 

Out-Sample 0.4667 0.9111 0.9289 0.9378 0.7778 0.8756 

Note: As the Merton model does not require In-Sample parameters, such data is not available 

Model 

Accuracy Ratio 

Table 7: Accuracy Ratio of Each Model with the Common Variables Set 

Merton 

Option 

Pricing 

Logit 

Regression 

Probit 

Regression 

Discriminant 

Analysis 

Hazard Ratio 

Artificial 

Neural 

Network 

In-Sample N/A* 0.9578 0.9378 0.9589 0.7956 1.0000 

Out-Sample 0.5467 0.9200 0.9211 0.9289 0.7889 0.8067 



From the validation results, we can see that the 

accuracy ratios of using common variables set are similar 

to the accuracy ratios using best variables set. This 

demonstrates it is feasible to establish a common 

variables set which can be applied to different models. 

C.Credit default probabilities 

During validation, we have obtained the default 

probabilities of each sample company under different 

models. Once we have a representative probability rate 

(score) for each sample, we can carried out the grading 

with the scores. 

Regarding the estimation of default probabilities, the 

most direct approach is to acquire the ratings of the 

samples in different periods, and obtain the long run 

average estimated default probabilities under each grade 

by building a transition matrix. 

Take Logit Model as an example; based on the rating 

results in 2005 & 2006 and information of default 

occurrence of the companies in 2006, we can build a 

transition matrix as in Table 8. 

Table 8: 2005-2006 Transition Matrixes 

1 2 3 4 5 6 7 8 9 10 Default 

1 72% 16% 8% 0% 0% 3% 0% 1% 0% 0% 0% 

2 9% 53% 26% 3% 5% 2% 2% 2% 0% 0% 0% 

3 12% 10% 39% 25% 14% 2% 0% 3% 3% 2% 0% 

4 0% 4% 19% 33% 23% 17% 4% 0% 0% 0% 0% 

5 0% 0% 5% 15% 41% 26% 8% 5% 0% 0% 0% 

6 0% 2% 2% 3% 14% 34% 29% 14% 2% 0% 2% 

7 0% 0% 0% 0% 3% 24% 31% 29% 3% 0% 9% 

8 0% 0% 0% 2% 0% 7% 11% 40% 28% 7% 4% 

9 0% 0% 0% 0% 0% 1% 1% 22% 46% 26% 5% 

10 0% 0% 0% 0% 0% 2% 0% 0% 12% 65% 22% 

The data in transition matrix represents the probability 

that credit quality migrate from some rate to another rate 

after a year. This probability can be calculated as follows: 

P 

i, j 

= 

n 

n 

1, j 

0, j 

n 0, j is the number of companies with credit rating i at 

n 1, j 

is the number of companies with credit rating j 

If data deficiencies result in banks failing to estimate 

default probabilities, we need quantify the internal ratings 

by use of other approaches. By Carey、Hrycay (2001), 

Mapping method can be used to calculatedefault 

probabilities. 

Based on the above mentioned internal rating results, 

using the rating results of Logit Model as indicators, we 

find out the default probabilities from internal ratings by 

various mapping methods including judgmental, 

mechanical and weighted average mappings. 

t=0, 

at t=1. 

Table 9: Simulation Mapping Results 

Actual Data Mapping 

Internal 

Rating 

No. of 

Companies 


Defaulting 

Companies 

Default 

Probability 

(PD) 

Median of 

Logit 

Regression 

Rating 

PD that 

corresponds 

to the median 

Average of 

Logit 

Model 

Ratings 

Weighted 

Average 

PD 

1 1089 0 0.00% 1 0.00% 1.27 0.04% 

2 978 0 0.00% 2 0.00% 2.39 0.29% 

3 942 3 0.31% 3 0.00% 3.30 0.55% 

4 882 3 0.34% 4 0.67% 4.16 0.66% 

5 948 12 1.27% 5 0.98% 5.09 0.81% 

6 1032 15 1.45% 6 0.58% 5.77 1.22% 

7 936 30 3.21% 7 1.76% 6.66 2.58% 

8 1116 48 4.30% 8 5.36% 7.55 4.58% 

9 1044 69 6.61% 9 7.80% 8.69 7.77% 

10 1065 105 14.37% 10 14.53% 9.81 13.39% 

Under Judgmental mapping, at first we map internal 

grades to external grades subjectively, then take default 

probabilities from external ratings as probabilities from 

internal ratings. Under mechanical mapping, we sort the 

company in each internal grade by corresponding external 

grades, and then take the median of average default 

probabilities in external grades as default probabilities in 


each internal grade. Under weighted average mapping, 

we take the weighted average of average default 

probabilities in external grades corresponded to internal 

grades as default probabilities in each internal grade. Due 

to Judgmental mapping is lack of logical basis, we just 

calculate the default probabilities with mechanical


mapping and weighted average mapping respectively (see 

Table9). 

We can see that it is more suitable to use median to 

compute the actual default probability for the safe grades 

Logit 

Probit 


Co. 

(1~4), while it is better to use weighted average to 

estimate actual default probability for the risky grades 

(8~10). 

Table 10: Table of Benchmark Ratings 

1 2 3 4 5 6 7 8 9 10 

1 363 95% 3% 2% 0% 0% 0% 0% 0% 0% 0% 

2 326 5% 91% 4% 0% 0% 0% 0% 0% 0% 0% 

3 314 0% 4% 93% 3% 0% 0% 0% 0% 0% 0% 

4 294 0% 0% 5% 90% 4% 1% 0% 0% 0% 0% 

5 316 0% 0% 0% 3% 92% 5% 0% 0% 0% 0% 

6 344 0% 0% 0% 1% 5% 91% 3% 0% 0% 0% 

7 312 0% 0% 0% 0% 3% 4% 89% 3% 1% 0% 

8 372 0% 0% 0% 0% 0% 0% 1% 94% 5% 0% 

9 348 0% 0% 0% 0% 0% 0% 0% 5% 90% 5% 

10 355 0% 0% 0% 0% 0% 0% 0% 1% 11% 88% 

To validate the default probabilities, we adopt 

benchmark comparisons for empirical explanations. 

Continue with the above analysis using the results 

corresponding to the medians under mapping, we use 

Logit Model as benchmarks for ratings comparison. 

Then carry out the benchmarks comparison to the Probit 

Model. The data represents the distribution of the number 

of some rate under logit model. For example, when the 

rate is 1 under logit model, 95%number of company 

belongs to 1 under probit model; only 5% number of 

company is other rate. The results are shown in Table 10. 

From Table 10, we can see that the comparison results 

for each grade are acceptable. Consequently, we can 

build a credit scoring model based on these findings. 

V. CONCLUSION 

The effectiveness of credit risks management relies on 

whether it can operate with the local environment. 

Therefore, choosing the variable set that fits in with the 

local conditions is critical to the performance of a credit 

rating model. But most of researchers always choose 

corresponding variables for different model. In this paper, 

we have adopted Merton Option Pricing Model, 

Discriminant Analysis Model, Regression Analysis 

Models (Logit Model and Probit Model), Survival 

Analysis Model and Artificial Neural Network Model in 

finding the common variables set for application in credit 

risks management models.Our findings show that five 

variables, namely Return on Assets Ratio (Before Interest 

and Tax), Compound Return on Assets Ratio, Quick 

Ratio,Total Asset Turnover, Debt-to-EquityRatio, are 

applicable to different credit risk rating 

models.Moreover, this paper estimates long term average 

default probability by using Transition Matrix and 

Mapping method. Validation findings also show that they 

have good forecasting abilities. Such findings help to 

simplify the application of credit risks management 

models and better adapt the models to the local conditions 

in China. We believe that the research methods presented 

in this paper can also be applied to other countries or 

regions around the worlds. It can serve as a good 


reference for establishing credit risks management 

models that fit with the local conditions. 


This work was supported by the Fundamental Research 

Funds for the Central Universities (2010221040), China 

National Social Science Fund (09AZD045), Ministry of 

Education for Humanities and Social Sciences 

(08JA630004), Anhui Provincial Natural Science 

Research Project for Universities (KJ2010A072) and 

China National Bureau of Statistics Fund (2009LZ045). 

We would like to thank the editor, associate editor, and 

referees for careful review and insightful comments, 

which have led to significant improvement of the article. 

REFERENCES 

[1] Basel Committee on Banking Supervision (1999), Credit 

Risk Modeling: Current Practice and Application, Bank for 

International Settlements. 

[2] Basel Committee on Banking Supervision (2003), 

Consultative Document: The New Basel Capital Accord, 

Bank for International Settlements. 

[3] Carey, Mark and M. Hrycay (2001), Parameterzing credit 

risk models with rating data, Journal of Banking & Finance 

25, p. 197-270 

[4] Division of Banking Supervision and Regulation (1998), 

Bank Holding Company Supervision Manual, Board of 

Governors of the Federal Reserve System 

[5] Division of Banking Supervision and Regulation(2003), 

Draft Supervisory Guidance on Internal Ratings-Based 

Systems for Corporate Credit, Board of Governors of the 

Federal Reserve System 

[6] Engelmann, Bernd, E. Hayden and D. Tasche (2003), 

Testing Rating Accuracy, Risk Falkenstein, E., A. Boral, 

and L. Carty (2000), RiskCalc Private Model: Moodys 

Default Model for Private Firms, Moodys Investors 

Service. 

[7] Ferguson Jr., Roger W. (2003), Basel II: Some Issues for 

Implementation,Basel Sessions 2003 Speech, Institute of 

International Finance, New York . 

[8] Keenan, S. C. and J. R. Sobehart (2001), Performance 

Measures for Credit Risk Models, Moody’s Risk 

Management Services Research Report.


Corporate-, Product-, and User-Image 

Dimensions and Purchase Intentions 

The Mediating Role of Cognitive and Affective Attitudes 

Xian Guo Li*, Xia Wang, Yu Juan Cai 

Department of Marketing 

School of Business, Renmin University of China, Beijing, China 

Abstract—This study investigates the effects of corporate-, 

product- and user image dimensions on purchase intentions, 

with cognitive and affective attitudes as mediator. A 

questionnaire survey was conducted with convenience 

sample. The results demonstrate significant effects of three 

brand image dimensions on purchase intention. In addition, 

the cognitive and affective attitudes fully or partially 

account for the relationship. This study contributes to the 

understanding of the assessment of the relationship between 

brand image dimensions and purchasing behavior. 

Implications for brand management are also discussed. 

Index Terms—Corporate Image, Product Image, User 

Image, Purchase Intention 


Brand image has been an important concept in 

consumer behavior research since the early 1950s. Both 

marketing researchers and marketers have long advocated 

the use of a clearly defined brand image as a basis for 

market success. A well-communicated brand image 

enables consumers to identify the needs satisfied by the 

brand and thereby differentiate the brand from its 

competitors [1, 2]. In fact, developing a brand image 

strategy has been described as the first and most vital step 

in positioning a brand and driving brand equity in the 

marketplace [3-5].As the growing importance of brand 

image strategy in marketing, a research issue evolved that 

how the brand image perceptions affect consumers 

purchasing behavior. 

Brand image is the most efficient way to talk to 

consumers via translating the different benefits about a 

brand. One common mistake brand strategists make is 

having too narrowed a view of the brand and only 

focusing some attributes when creating a brand’s image 

[6]. Given consumers’ perceptions may not be product 

specific; brand image is a multi-dimensional construct. 

The image of a brand can be described as having three 

contributing sub-images, the image of the provider of the 

product/service, or corporate image; the image of the 

user; and the image of the product/service itself [4]. Thus, 

two questions arise: what makes an effective brand and, 

Corresponding author: Xian Guo Li is with the Department of 

Marketing, School of Business, Renmin University of China (Phone: 

86-13910602316; Email: rdlxg@126.com) 


doi:10.4304/jcp.6.9.1875-1879 

second, how the company can effectively communicate to 

consumers with different brand image strategy? 

In the aforementioned studies, however, relatively little 

empirical evidence has been provided for the effects of 

these dimensions on purchase intention. Especially for 

the Chinese markets, as brand image perception varies 

across culture [7], the effects of brand image dimensions 

need to be examined further. Thus, the purpose of this 

study is to investigate the predicting roles of corporate-, 

product-, and user-image on purchase intention in the 

context of Chinese mobile-phone market, and the 

mediating role of cognitive and affective attitude are also 

examined in this study. 

This paper is structured as follows. We first review the 

literature on key issues involving brand image, 

associations with purchasing behavior, and the mediating 

role of attitude. The data and methods of the study follow 

in the section. Empirical evidence on the effects of brand 

image dimensions on purchasing intention is provided, 

and the mediating role of cognitive and affective attitude 

is also demonstrated. The paper ends with conclusions 

and implications. It is expected that this study will 

provide a more thorough understanding of building a 

company’s brand image strategy focusing on three brand 

image dimensions in Chinese mobile-phone industry. 


A. Brand Image 

There has been general agreement that brands—at least 

some brands—do have images, defined as the 

associations linked to a brand [4], or perceptions about a 

brand as reflected by the brand associations held in 

consumer memory [3]. When consumers see a particular 

brand, the brand association is any idea caused by that 

certain brand, including feelings, experiences, appraisals, 

and brand positioning [3]. The brand image perception 

varies across categories, brands [4] and culture [7], thus 

need to be investigated in multi-cultures, especially for 

Chinese markets. 

Brand image is a complex constructs and can be made 

of several dimensions [8]. Brand association is the mutual 

combination of informational nodes and come from all 

possible forms, and may reflect product characteristics or 

independent characteristics outside the product [3]. Biel


[4] suggests brand image has three components: corporate 

image, image of the user and image of the product. While 

Hsieh et al [9] extends the product image with corporate 

image and country image, and inspect the relationship 

between product-, corporate-, country- image and 

purchase behavior, which was also verified in multicultures 

[10]. 

Park et al [2] states brand image incorporates the 

functional, experiential and symbolic benefits to the 

consumer; many brands offer a mixture of symbolic, 

functional, and experiential benefits. A brand with 

functional benefits is one designed to satisfy consumers' 

needs to solve consumption-related problems. A brand 

with symbolic (or social) benefits is one designed to 

fulfill consumers' desires for self-enhancement, role 

position, group membership, or ego identification. 

Finally, a brand with experiential benefits is one designed 

to fulfill consumers' desires for sensory pleasure, variety, 

or cognitive stimulation. 

The current study extends Hsieh’s et al [9] study by 

following Biel’s [4] definition and adopts the three 

previously mentioned brand concepts as corporate image, 

product image, and user image. The brand theorists 

suggest that what a person knows about a company can 

influence perceptions of the company's products, e.g. the 

corporate ability associations and corporate social 

responsibility associations will influence consumers’ 

beliefs about and attitudes towards the products of that 

company [11], thus corporate brand image may affect the 

product evaluations, and the relationship is moderated by 

perceived risk [12]. The product image is related to the 

benefits attached to the products. As the symbolic, 

functional, and experiential benefits of the products have 

been proved to lead to brand preference [6], the product 

image will also influence the product evaluations. The 

user image refers to whether the brand personality is 

congruent with the consumers [13]. If the brand 

personality fit the consumers’ self-concept, the product 

may receive a high evaluation. 

With regard to the performance of brand image, Aaker 

[1] claims that brand association aid in acquiring or 

handling information, creating positive attitudes or 

feelings, positioning brand and differentiating it from 

competitors as well as creating value for the company. 

Empirical evidence suggests that brand image has 

positive influence on brand-extension attitude [8]. 

Moreover, Krishnan [5] demonstrates that compared to 

brands with low equity, high equity brands will have 

greater number of positive associations, more unique 

associations from competing brands, fewer unique 

associations from the category, and more associations 

from direct experiences and word-of-mouth, which has 

directly verified the relationship between association 

pattern of brand image and brand equity. 

B. Mediating Role of Consumers Attitude 

Brand associations in the study of Keller [3] are 

classified into three major categories with respect to their 

level of abstraction (i.e., attribute, benefit, and overall 

brand attitude). Here, attribute refers to descriptive 

features that characterize a product or service, benefit is 


the personal value that consumers attach to the product or 

service, and brand attitude is consumers' overall 

evaluation of the brand [9]. Ideally, in consumers' 

memory, brand-image perception should encompass all 

three types of brand associations. However, given the 

entailed complexity, most of the studies incorporate only 

benefit associations as the key elements [6, 9]. The 

corporate-, product-, and user image of this study also 

embrace only the benefit associations of brand image. 

Thus, overall brand attitude as an important part of brand 

association should be investigated further. 

Brand attitudes are important because they often form 

the basis for consumer behavior (e.g., brand choice). 

Though different models of brand attitudes have been 

proposed, one widely accepted approach is based on a 

multi-attribute formulation in which brand attitudes are a 

function of the associated attributes and benefits that are 

salient for the brand [3]. According to the theory of 

planned behavior, there are three conceptually 

independent determinants of intention: attitude toward the 

behavior, subjective norm and perceived behavioral 

control. As a general rule, the more favorable the attitude, 

the stronger should be an individual’s intention to 

perform the behavior under consideration [14]. According 

to the planned behavior theory, attitudes develop 

reasonably from the beliefs people hold about the object 

[14]. Thus the attitude may mediate the relationship 

between brand image beliefs and purchase intention. 

The basic theory of planned behavior model was 

expanded to include the separation of affective and 

cognitive predictors of attitude towards purchase 

intention. The majority of social psychology literature 

suggests that attitudes are composed of cognitive, 

affective, and behavioral parts. This multidimensional 

view of attitude implies that consumers’ willingness to 

buy may be influenced by cognitive and affective 

antecedents [15]. Following this, we propose that 

purchase intention can be predicted by cognitive and 

affective attitudes. 

C. Research Questions and Hypotheses 

Drawing from the previous literature and field 

observation, we set up the research questions in an 

attempt to explore the relationship between brand image 

dimensions and purchase intention. Generally, the social 

responsibility and consumer concern of an enterprise will 

increase the consumers’ willingness to buy their products, 

and a well product or service image may increase the 

consumers’ brand usage. Meanwhile, consumer would 

like to buy the products with congruent personality. Thus, 

we hypothesize that: 

H1a: Corporate image has positive influence on 

purchase intentions. 

H1b: Product image has positive influence on purchase 

intentions. 

H1c: User image has positive influence on purchase 

intentions. 

According to the planned behavior model, the attitude 

may mediate the relationship between beliefs and 

intention. Thus with regard to this study, we separate the


two components of cognitive attitude and affective 

attitude, and hypothesize that: 

H2a: Cognitive attitude will mediate the relationship 

between corporate image and purchase intention. 

H2b: Cognitive attitude will mediate the relationship 

between product image and purchase intention. 

H2c: Cognitive attitude will mediate the relationship 

between user image and purchase intention. 

H3a: Affective attitude will mediate the relationship 

between corporate image and purchase intention. 

H3b: Affective attitude will mediate the relationship 

between product image and purchase intention. 

H3c: Affective attitude will mediate the relationship 

between user image and purchase intention. 

III. DATA AND METHOD 

A. Instruments 

All our measures employ items from multiple-item 

scales that have been tested and used in previous studies. 

The dependent variables of purchase intention gathered 

from the work of Dodds et al [16]. The purchase intention 

was measured on four items as “The likelihood of buying 

products of this brand is very high”, “I would consider 

buying products of this brand”, “The probability that I 

would like to buy products of this brand is very high”, 

and “My willingness to buy this product is very high”. 

The predictor variables of three brand image 

dimensions were measured on multi-items modified from 

Zhuohao et al [17] and Xiucheng and Jie [18]. The 

corporate image was measured as “The innovation and 

update of the products of this corporate is strong”, “The 

corporate of this brand care for customer very much”, and 

“The corporate of this brand have a well impression”. 

The product image was measured with the following 

indicators: Function, Style, Durability, and Quality. And 

the user image was measured with the following 

statements. “I can easily imagine this brand as a person”, 

“This brand have a strong personality”, and “The 

personality of this brand matches with mine”. 

The mediator variables of cognitive attitude and 

affective attitude were adapted from the study of 

Verplanken et al [19]. The affective attitudes were: 

favorable, pleasant, comfortable, exciting, and attractive. 

The cognitive attitudes were good, wise, positive, useful, 

and worthy. All of the items were evaluated on a sevenpoint 

Likert scale ranging from “strongly disagree” to 

“strong degree”. 

B. Participations 

The study was investigated in mobile-phone market, 

because the products have the following characteristic: 

(1) It is a hedonic and utilitarian products; (2) It is a highinvolvement 

characteristic; and (3) It is widely adopted in 

Chinese markets. Data was collected from three 

universities in Beijing. The final effective sample size 

was 268. 


IV. RESULTS 

A. The Effect of Brand Image Dimensions on Purchase 

Intention 

We expect that the brand image dimensions have 

positive dimensions have positive influence on purchase 

intention. The coefficients of brand image dimensions 

and purchase intention were estimated and presented in 

table 1. All of the coefficients for corporate image, 

product image and user image were significantly positive, 

indicating positive relationships between brand image 

dimensions and purchase intention. H1 to H3 were 

supported. The standardized path coefficients of product 

image leading to purchasing intention showed a relative 

stronger relationship than the two other dimensions, 

indicating the product image should be given more 

attention in the context of this study. 

TABLE I. REGRESSION RESULTS OF BRAND IMAGE 

DIMENSIONS ON PURCHASE INTENTION 

B SE B β 

Constant 1.044*** .195 

Corporate Image .094* .047 .104 

Product Image .423*** .046 .451 

User Image .202*** .052 .212 

*p


partially explain the relationship between product image 

and purchase intention. 

TABLE II. THE MEDIATING EFFECT OF COGNITIVE 

ATTITUDE ON THE RELATIONSHIP BETWEEN CORPORATE 

IMAGE AND PURCHASE INTENTION 

Step Predictor: 

Corporate Image 

B SE B β 

1 Corporate Image .411*** .043 .456 

2 Corporate Image .434*** .036 .540 

3 Corporate Image .035 .035 .039 

Cognitive Attitude .867*** .043 .772 

4 Sobel Z 10.32*** 

*p


previous studies, which suppose corporate image, product 

image, and user image will affect the consumers’ 

willingness to purchase. 

Results of this study indicates that building brand 

image of mobile-phone industry should focus more on the 

product image to leading consumers’ buying decisions, 

however, other dimensions of corporate and user image 

should not be ignored as they have positive relationship 

with consumers’ purchasing behavior. 

Besides that, the mediating roles of cognitive and 

affective attitudes were also examined in the study and 

reveal that attitudes can fully or partially account for the 

relationships between brand image dimensions and 

purchase intentions. 

The contribution of this study is to empirically 

investigate the effects of brand image dimensions on 

purchasing behavior, and examine the influence route 

deeply from the view of planned behavior model, which 

may contribute to understanding the relationships 

between brand image dimensions and performance, and 

make a supplementary for the planned behavior model 

further. The managerial implication is that help the 

enterprises comprehend the three dimensions of brand 

image, and make appropriate marketing campaigns. 

Limitation of this study includes the lack of category 

specific investigation as the contribution of the three 

dimensions of brand image varies by product category 

and by brand [4]. Sampling frames is coming from the 

students in the university. The convenience sample may 

limit the generalizability of this study. Other variables 

such as subjective norm should also be controlled in the 

study, as they may affect the consumers’ willingness to 

buy [16]. Direction of further research is to conduct 

research in other categories and increase the 

generalizability of the study. 

REFERENCE 

[1] Aaker, D. (1991). Managing Brand Equity. Ontario: The 

Free Press. 

[2] Park, C.W., Jaworski, B.J., Maclnnis, D.J. (1986). 

“Strategic Brand Concept-Image Management”, Journal 

of Marketing, 50(4): 135-145. 

[3] Keller, K.L. (1993). “Conceptualizing, measuring, and 

managing customer based brand equity”, Journal of 

Marketing, 57 (1), 1-22. 

[4] Biel, A. (1992). “How Brand Image Drives Brand Equity”, 

Journal of Advertising Research, 32(6): 6-12. 

[5] Krishnan, H. S. (1996). “Characteristics of memory 

associations: A consumer-based brand equity perspective”, 

International Journal of Research in Marketing, 3, 389- 

405. 

[6] Salciuviene, L., Lee, K., Yu, C. (2007). “The Impact of 


Brand Image Dimensions on Brand Preference”, 

Economic and Management, 12, 464-469. 

[7] Park, H., Rabolt, N.J. (2009). “Cultural Value, 

Consumption Value, and Global Brand Image: A Cross- 

National Study”, Psychology & Marketing, 26(8): 714- 

735. 

[8] Salinas, E.M., and Perez, J.M.P. (2009). “Modeling the 

brand extensions' influence on brand image”, Journal of 

Business Research, 62: 50-60. 

[9] Hsieh, M., Pan, S., Setiono, R. (2004). “Product-, 

Corporate-, and Country Image Dimensions and Purchase 

Behavior: A Multicountry Analysis”, Journal of the 

Academy of Marketing Science, 32(3): 251-270. 

[10] Chung, J., Pysarchik, D.T., and Hwang, S. (2009), 

“Effects of Country-of-Manufacture and Brand Image on 

Korean Consumers’ Purchase Intention”, Journal of 

Global Marketing, 22, 21-41. 

[11] Brown, T.J., Dacin, P.A. (1997). “The Company and the 

Product: Corporate Associations and Consumer Product 

Responses”, Journal of Marketing, 61(1): 68-84. 

[12] Gurhan-Canli, Z., and Batra, R. (2004). “When Corporate 

Image Affects Product Evaluations: The Moderating Role 

of Perceived Risk”, Journal of Marketing Research, 41(2): 

197-205. 

[13] Sirgy, J. (1982). “Self-concept in Consumer Behavior: A 

Critical Review”, Journal of Consumer Research, 9, 287- 

300. 

[14] Ajzen, I. (1991), “The Theory of Planned Behavior, ” 

Organizational Behavior and Human Decision Processes, 

50, 179-211. 

[15] Zajonc, R.B., Markus, H. (1982), “Affective and 

Cognitive Factors in Preferences”, Journal of Consumer 

Research, 9(2): 123-131. 

[16] Dodds, W., Monroe, K.B., Grewal, D. (1991), “Effects of 

Price, Brand, and Store Information on Buyers' Product 

Evaluations”, Journal of Marketing Research, 28: 307-319. 

[17] Zhuohao, C., Zhi, L., and Qingyun, J. (2006), “How Does 

Brand Personality Influence Consumer’s Attitudes ? A 

Study from the Perspective of Consumer Brand 

Cognition”, Journal of Marketing Science, 2(2): 103-116 

(In Chinese). 

[18] Xiucheng, F., and Jie, C. (2002), “Measurement of Brand 

Image: A Brand Identity-Based Integrated Model and 

Empirical Study”, Nankai Journal (Philosophy and Social 

Science Edition), 3, 65-71. 

[19] Verplanken, B., Hofstee, G., Janssen, H.J.W. (1998), 

“Accessibility of Affective versus Cognitive Components 

of Attitudes”, European Journal of Social Psychology, 28, 

23-35. 

[20] Baron, R.M., & Kenny, D. A. (1986). “The Moderator - 

Mediator Variable Distinction in Social Psychological 

Research: Conceptual, Strategic, and Statistical 

Considerations”, Journal of Personality and Social 

Psychology, 51(6):1173-1182.


A Microcomputer-Based Predictive Digital 

Current Programmed Control System for Threephase 

PWM Rectifier 

Zhongjiu Zheng 

College of Electrical Engineering, Dalian University of Technology, Dalian, 116024, P .R. China 

zhengzhongjiu@163.com 

Guofeng Li, and Ninghui Wang 

College of Electrical Engineering, Dalian University of Technology, Dalian, 116024, P .R. China 

guofenli@dlut.edu.cn, ninghuiw@263.net 

Abstract—The paper describes a microcomputer control 

system, which uses the floating-point digital signal processor 

TMS320LF2407 from Texas Instruments, for three-phase 

PWM rectifier. It could effectively eliminate harmonic 

distortion of line currents and provides power factor 

correction. Moreover, it can be save electrical energy and 

reduction of production cost. In the control system, the 

predictive current control in two-dimensional (α-β) 

stationary frame, makes the input current following the 

phase voltage in phase to get unity power factor; and space 

vector pulse wide modulation (SVPWM) generates the 

modulation wave. Finally, the three-phase PWM rectifier 

using the proposed control system is designed in 

Simulink/Matlab and executed in laboratory prototype, and 

the results are provided to verify the proposed control 

system in the end of the paper. 

Index Terms—PWM rectifier; Predictive digital current 

control; Space vector pulse wide modulation; Unity power 

factor 


In knowledge economy era, research in the field of 

power electronics has taken a great interest in the power 

quality, such as power supply efficiency, saving electrical 

energy, economical, reliability, volume, and weight. 

Traditional uncontrolled three-phase rectifiers have been 

widely used in the industrial complexes, but the 

disadvantages are severity energy losses, high cost, big 

volume and weight, and introducing massive harmonic 

currents into the grid that does not fulfill the new 

standards for the electric grid. 

With the development of digital signal processors 

(DSP) control devices and the IGBT power devices, DSPbased 

controller for three-phase PWM rectifiers have 

been proposed in some papers [1]-[6] in which general 

purpose and floating-point DSPs are used. This technique 

uses a floating-point DSP to effectively eliminate system 

harmonics and it also provides power factor correction. 

Moreover, it can be save electrical energy and reduction 

of production cost. 

PWM rectifiers [7]-[11] as a non-polluting and 

economical equipment are going to be more popular 

because of several advantages described as: 


doi:10.4304/jcp.6.9.1880-1885 

- reduction of production cost; 

- saving electrical energy; 

-Low harmonic distortion of line currents; 

- Regulation of input power factor to unity; 

- Adjustment and stabilization of output DC voltage; 

-Bi-directional power flow; 

The objective of this paper is to present a economical 

predictive digital current control strategy of three-phase 

PWM rectifier based on modern floating-point digital 

signal processor (DSP) which facilitates the work on 

software development. The proposed predictive digital 

current control system operates with constant switching 

frequency using space-vector modulation (SVM).The 

control system include predictive current algorithm, 

SVPWM control algorithm, proportional integral (PI) 

algorithm and so on. The predictive current control make 

the input current following the phase voltage in phase to 

get unity power factor; SVPWM generates the six via 

modulation wave; and the PI regulator keep the output 

voltage constant . In this way, the whole system of PWM 

rectifiers is obtained when the control algorithm and 

PWM generation are carried out using a digital signal 

processor (DSP) with minimal external hardware. 

II. MODELING FOR THE THREE-PHASE PWM RECTIFIER 

The three-phase boost PFC rectifiers is consisted of six 

switches with anti-paralleled diodes as shown in Figure 1 

.This topologies is ideally applicable to DC-linked AC 

Figure 1. The main circuit of three-phase PWM rectifier.


motor drives since it draws sinusoidal input current ,and 

controls the DC bus voltage .Moreover ,its capability of 

bi-directional power flow allows reverse operation ,which 

is especially advantageous for three-phase high power 

factor PWM rectifier. 

Assume that the three-phase voltage is symmetrical, 

stable and interior resistance is zero; three-phase loop 

resistance R and L are the same value respectively; 

switching loss and on-state voltage is neglectable; 

affection of distributing parameters is neglectable 

;switching frequency of the rectifier is high enough . 

The parameters in Figure 1 are listed below. 

e , e , e Phase voltage; 

a b c 

i , i , i Phase current; 

a b c 

U , U , U Voltage between leg midpoint and N 

AN BN CN 

point; 

L Input inductance; 

R Equivalent resistance of the loop; 

C Capacitance of the dc bus; 

S 

V dc Output DC voltage; 

i o Load current; 

R Load resistance; 

L 

We can define switch function as follows, 

S i 

⎧1 

= ⎨ 

⎩0 

i phase upper switch is on 

i phase bottom switch is on 

(1) 

,i=a,b,c 

Hence, the mathematic model of PWM rectifier is: 

⎡ 

⎢ 

− R 

⎢ 

⎢ 0 

A = ⎢ 

⎢ 

⎢ 0 

⎢ 

⎢ 

⎣ sa 

• 

Z X = A X + U 

0 

− R 

0 

s 

b 

0 

0 

− R 

s 

c 

− ( s 

a 

− ( s 

b 

− ( s 

[ L, 

L, 

L C ] 

Z , 

S 

c 

1 

− 

3 

1 

− 

3 

1 

− 

3 

0 

∑ 

k 

k = a, 

b, 

c 

∑ 

k 

k = a, 

b, 

c 

∑ 

k 

k = a, 

b, 

c 

(2) 

⎤ 

s ) 

⎥ 

⎥ 

⎥ (3) 

s ) 

⎥ 

⎥ 

s ) ⎥ 

⎥ 

⎥ 

⎦ 

= (4) 

[ ] T 

e e , e i 

U = , − 

a 

, (5) 

b 

[ ] T 

i , i , i , 

X V 

a 

b 

c 

c 

dc 

o 

= . (6) 

III. CONTROL SYSTEM FOR THREE-PHASE PWM RECTIFIER 

The control system schematic diagram of the threephase 

PWM rectifier is shown in figure 2. The control 

system adopts predictive current control in twodimensional 

(α-β) stationary frame ,pulse wide 


modulation mode is based on space vector ,DC voltage 

control adopts conventional PI controller. This method 

keeps the fast response merit. Real current can follow 

reference current in one switching period and switching 

frequency keeps constant .In addition ,parameters 

selection is simple for there is only one PI controller in 

the system . 

A. Principle of Predictive Digital Current Control 

Write mathematical model of three-phase PWM 

rectifier in three-dimensional stationary (a-b-c) frame as 

eα 

eβ 

iβ 

iα * 

uαN * 

uβN * 

I amp 

Figure 2. Predictive current control configuration of three-phase PWM 

rectifier based on (α-β) stationary frame. 

follow . 

⎧ 

⎪U 

⎪ 

⎨U 

⎪ 

⎪ 

U 

⎪⎩ 

AN 

BN 

CN 

= e 

a 

= e 

b 

= e 

c 

dia 

− ( L + Ria 

) 

dt 

dib 

− ( L + Rib 

) 

dt 

dic 

− ( L + Ric 

) 

dt 

The mathematic model in the two-dimensional 

stationary (α-β) frame can be obtained by applying the 

following α-β transformation as seen in equation (8). 

Vdc 

* 

Vdc 

(7) 

⎡ 1 1 ⎤ 

⎢ 

1 − − 

2 

⎥ 

T = 

2 2 

αβ / abc ⎢ 

⎥ (8) 

3 ⎢ 

3 3 

0 − ⎥ 

⎢⎣ 

2 2 ⎥⎦ 

Expression (9) is the mathematic model in stationary 

(α-β) frame: 

⎧ 

⎪U 

⎨ 

⎪U 

⎪⎩ 

αN 

βN 

= e 

= e 

α 

β 

di 

− ( L 

dt 

di 

− ( L 

dt 

α 

β 

+ Ri 

α 

+ Ri 

β 

) 

) 

(9)


Judging from (9), Uα N , U β N are the only variable to 

control AC current i α , i β respectively. Average (9) from 

t k to k + 1 

t , it derives 

∫ + 1 tk 

1 diα 

U α N = eα 

− ( L + Riα 

) dt 

T tk 

dt 

S 

L 

R t 

e − [ iα 

( tk 

+ 1) 

− iα 

( tk 

)] − iα 

dt 

T 

T tk 

= ∫ + k 1 

α (10) 

S 

∫ + 1 tk 

1 diβ 

U β N = eβ 

− ( L + Riβ 

) dt 

T tk 

dt 

S 

L 

R t 

e − [ iβ 

( tk 

+ 1) 

− iβ 

( tk 

)] − iβ 

dt 

T 

T tk 

= ∫ + k 1 

β (11) 

S 

Here, ( U α N , U β N ) , ( e α , e β ) stand for 

average value of( Uα N , U β N ),( e α , e β )in one 

control period respectively. S 

T = k + 1 

S 

S 

t - t k . Assume U αN 

, U β N are the same with the reference voltage 

U α , U β in each period and omit R, it can derive 

* 

N 

* 

N 

⎧ 

⎪ 

U 

⎨ 

⎪U 

⎪⎩ 

* 

αN 

* 

βN 

= e 

= e 

α 

β 

L 

− [ i ( t α 

TS 

L 

− [ i ( t β 

T 

S 

k + 1 

k + 1 

) − i ( t 

α 

) − i ( t 

β 

k 

k 

)] 

)] 

(12) 

Assume grid current can track reference current in one 

period that means iα t ) = 

* 

iα t ) and 

( k + 1 

( k + 1 

* 

iβ t ) = iβ ( t + 1) 

, thus (13) can be written as 

( k + 1 

k 

⎧ 

⎪ 

U 

⎨ 

⎪U 

⎪⎩ 

* 

αN 

* 

βN 

= e 

= e 

α 

β 

L 

− [ i 

TS 

L 

− [ i 

T 

S 

α 

β 

* 

* 

( t 

( t 

k+ 

1 

k+ 

1 

) − i ( t 

α 

k 

) − i ( t 

β 

k 

) 

) 

(13) 

From the expression (13) ,we can see that the same 

variables have been decoupled .And the variable 

* * 

( U α N , U β N ) is the reference voltage vector U ref which 

is the input value of the SVPWM algorithm. 

B. SVPWM ALGORITHM 

As shown in Figure 1 ,there are eight possible 

combinations of on and off states of the upper power 

transistors .So there are six kinds of active state ,i.e. ,nonzero 

vectors ,and two kinds of zero vector ( U and 000 


U ) .The eight basic space vectors defined by the 

111 

combination of the switches are shown in Figure 3 .In 

order to make the input current phase track the input 

voltage phase ,and keep output DC voltage constant ,the 

SVPWM technique is used to approximate the reference 

voltage vector U ref .The following presents how to use 

β 

* 

U βN 

Figure 3. Basic space vectors. 

d 

U0 

U 

0 

U60(110) 

° 

60 

d 

U60 

U 

60 

* 

U αN 

U ref 

α 

U0(100) 

Figure 4. Projection of the Reference Voltage Vector in Sector Ⅰ. 

the fast algorithm to obtain the SVPWM signal according 

to the reference voltage vector based on DSP . 

(a). Determination of the sector 

* 

* 

U and U are converted to a balanced three-phase 

α β 

N 

N 

quantities V ref 1 , V ref 2 and ref 3 

V according to the 

following inverse CLARKE transformation : 

⎧ 

⎪V 

⎪ 

⎨V 

⎪ 

⎪ 

⎪V 

⎩ 

ref 1 

ref 2 

ref 3 

= u 

rβ 

− u 

= 

− u 

= 

rβ 

rβ 

+ 

2 

− 3 * u 

2 

3 * 

u 

rα 

rα 

(14) 

From (14), the following decisions can be made on the 

variable N information: 

V > 0 then a=1, else a=0 

If ref 1 

If V ref 2 > 0 then b=1, else b=0 

If V ref 3 > 0 then c=1, else c=0 

The variable N is defined as : N = 4*c + 2*b + a ;


The sector in which U ref is depends on variable N 

.The corresponding relations between variable N and 

sector are shown in table Ⅰ. 

The Figure 4 show the projection of the reference 

voltage vector in sector Ⅰ. 

TABLE I. 

CORRESPONDING RELATIONS BETWEEN VARIABLE N AND 

SECTOR 

N 1 2 3 4 5 6 

Sector Ⅱ Ⅵ Ⅰ Ⅳ Ⅲ Ⅴ 

(b) Calculation of the durations 

* 

uα N 

* 

and uβ N represent the normalized (α , β) 

components of U ref with respect to the maximum phase 

voltage( V dc 

3 

). We can obtain the duty ratio by the 

following formula: 

⎡ 

⎡d 

⎤ 0 

x ⎢ 

⎢ ⎥ 

⎢ 

⎢ ⎥ 3 

= ⎢ 

⎢ 

d y ⎥ 

⎢ 2 

⎢ ⎥ 

⎢ 3 

⎢⎣ 

d ⎥⎦ 

⎢− 

z 

⎣ 2 

1 

⎤ 

⎥ * 

⎥ ⎡u 

α N ⎤ 

1 

(15) 

⎥ ⎢ ⎥ 

2 ⎥ ⎢ * ⎥ 

1 ⎥ ⎣u 

βN 

⎦ 

2 

⎥ 

⎦ 

For different sectors the value for duty ratio 

( d , d ) in terms of ( d x , d y , d z ) are listed in 

U k 

Table Ⅱ. 

U k + 60 

TABLE II. 

CORRESPONDING RELATIONS BETWEEN ( d U 

, d k U 

) AND 

k + 60 

( d x , d y , d z ) IN DIFFERENT SECTORS 

Sector Ⅰ Ⅱ Ⅲ Ⅳ Ⅴ Ⅵ 

d - z d z d d x -d x 

- d y d y 

U k 

d U k + 60 

d x y 

According to the duty ratio, we can get the 

corresponding SVPWM signal for the reference voltage 

U by DSP .Form the above motioned , this simplified 

ref 

d -d y 

and fast algorithm for SVPWM avoid the nonlinear 

operations and improve the calculation speed and 

accuracy. 

IV. SIMULATION AND EXPERIMENTAL RESULTS 

In order to evaluate the PWM rectifier performances, 

using the proposed predictive digital current control 

system operates with space-vector modulation (SVM), 

simulation and prototype model have been carried out 

using the following parameters: output power is 900W, 


d z - z d - d x 

input phase voltage e =50V , output DC voltage 

V dc =150V, equivalent resistance of the loop R =0.5Ω, 

input inductance L =8mH, DC-link capacitor 

C =4700µF, and the switching frequency 

S 

f =5KHz. 

s 

Simulation has been completed by MATLAB software 

.The predictive current control make the input current 

following the phase voltage in phase to get unity power 

factor .And the PI regulator to keep the output voltage 

constant .According to simulation model ,A three-phase 

boost-type PWM rectifier with DSP control system is 

implemented and tested in the laboratory .In the prototype 

model ,DSP micro-controller (TMS320LF2407A) and 

IPM (pm100cla60) are employed ;the AD conversion 

,CLARKE transformation ,PI regulator , predictive 

current controller ,and the fast algorithm for SVPWM are 

implemented in the software procedures . 

Both the simulation and experiment ,we have got the 

desired results .A sinusoidal input current in same phase 

with the corresponding input phase voltage is obtained as 

shown in Figure 5 and Figure 6.We can see that the 

power factor is nearly unity .Figure 7 and Figure 8 show 

the input current of phase A ,B ,C which are phase 

separation of 2π/3 and the THD


Figure 7. The simulation waveforms input current of phase A,B,C. 

Figure 8. The experimental waveforms of input current of phase A,B,C 

(current : 10A/div). 

V. CONCLUSIONS 

This paper has presented a economical three-phase 

PWM rectifier, which use the predictive digital current 

programmed control system. The proposed predictive 

current control strategy operates with constant switching 

frequency using SVM. The simulation and experimental 

results have proved excellent performance of the 

proposed predictive digital current programmed control 

system. Below is the features and advantages: 

1) It will reduce the production cost of the three-phase 

PWM rectifier; 

2) It is vital importance in non-polluting and energy 

conservation; 

3) The PWM rectifier realizes the low harmonic and 

unit power factor ; 

4) The utilization efficiency of DC voltage will be 

higher close to 1 ; 

REFERENCES 

[1] B.R. Lin and T.Y. Yang, “Three-phase AC/DC converter 

with high power factor,” IEE Proc.-Electr. Power Appl., 

vol.152, no. 3,pp.757-764, May. 2005. 

[2] C.-T. Pan and Y-H. Liao, “Modeling and coordinate 

control of circulating currents in parallel three-phase boost 

rectifiers,” IEEE Trans. Ind. Electron., vol. 54, no. 2, pp. 

825–838, Apr. 2007. 

[3] Soo-Bin Han, Nam-Sep Choi, and Gyu-Hyeong, 

“Modeling and analysis of static and dynamic 

characteristics for buck-type three-phase PWM rectifier by 

circuit DQ transformation,” IEEE Transactions on Power 

electronics, vol. 13, no. 2, pp. 323–336, Mar.1998. 


Figure 9. The simulation waveforms of output DC voltage. 

Figure 10. The experimental waveforms of output DC voltage 

(DC voltage : 25V/div ). 

[4] Yuri Shtessel, Simon Baev, Haik Biglari, “Unity Power 

Factor Control in Three-Phase AC/DC Boost Converter 

Using Sliding Modes,” IEEE TRANSACTIONS ON 

INDUSTRIAL ELECTRONICS, vol. 55, no. 11, pp.3874- 

3882,Nov.2008. 

[5] Ivo Barbi and Flabio Alberto Bardemaker Batista, “Space 

Vector Modulation for Two-Level Unidirectional PWM 

Rectifiers,” IEEE TRANSACTIONS ON POWER 

ELECTRONICS, vol. 25, no. 1, pp.178-187,Jan. 2010. 

[6] T. Jin and K. M. Smedley, “A universal vector controller 

for four-quadrant three-phase power converters,” IEEE 

Trans. Circuits Syst. I, Reg. Papers, vol. 54, no. 2, pp. 377– 

390, Feb. 2007. 

[7] Bhim Singh, BrijN.Singh, Ambrish Chandra,Kamal Al- 

Haddad, Ashish Pandey, Dwarka P. Kothari, “A Review of 

Three-Phase Improved Power Quality AC–DC 

Converters,” IEEE Transactions on Power electronics, vol. 

51, no. 3, pp.641-660,Jun.2004. 

[8] Ana Vladan Stankovic and Ke Chen, “A New Control 

Method for Input–Output Harmonic Elimination of the 

PWM Boost-Type Rectifier Under Extreme Unbalanced 

Operating Conditions,” IEEE TRANSACTIONS ON 

INDUSTRIAL ELECTRONICS, vol.56, no.7, pp.2420- 

2430,Jul .2009. 

[9] Sergio Vazquez, Juan Antonio Sanchez, Juan Manuel 

Carrasco, Jose Ignacio Leon, and Eduardo Galvan, “A 

Model-Based Direct Power Control for Three-Phase Power 

Converters,” IEEE TRANSACTIONS ON INDUSTRIAL 

ELECTRONICS, vol.55, No.4,pp1647-1657, Apr. 2008. 

[10] Yongsug Suh and Thomas A. Lipo, “Control Scheme in 

Hybrid Synchronous Stationary Frame for PWM AC/DC 

Converter Under Generalized Unbalanced Operating 

Conditions,” IEEE TRANSACTIONS ON INDUSTRY


APPLICATIONS, vol. 42, no. 3,pp.825-835, May/Jun. 

2006. 

[11] Wang Jiuhe, Li Huade, Wang Liming, “Direct Power 

Control System of Three Phase Boost Type PWM 

Rectifiers” Proceedings of the CSEE, vol.26 ,no.18,pp.54- 

60, Sep. 2006. 

Zhongjiu Zheng was born in 

Heilongjiang, China, in 1981. He 

received the B.S. degree in Information 

and Communication Engineering from 

Dalian University of Technology 

(DUT), Dalian, China, in 2003, where 

he is currently working toward the 

combined M.S./Ph.D. degrees in the 

area of Electrical and Electronics 

Engineering in DUT. 

His research interests include three-phase power factor 

correction, digital control of switching power converters, power 

converter topologies, and uninterrupted power supply system. 

Guofeng Li was born in Heilongjiang, 

China, in 1968. He received the B.S. 

degree in Physics from Harbin Normal 

University, Harbin, China, in 1990, the 

M.S degree from Northeast Normal 

University, Shenyang, China, in 1993, 

and the Ph.D. degree from Dalian 

University of Technology, Dalian, 

China, in 2000. 

From 1993 to 1997, he was a lecturer 

with the Physics Department, Northeast Normal University. 

Since 2000, he was a lecturer in Dalian University of 

Technology. Currently, he is a Professor and Director of the 

Special Power Supplies Research Institute, Dalian University of 

Technology. His research interests include special power 

supply, environmental engineering, static electricity, and ship 

electrified transmission automation. 

Ninghui Wang was born in Jilin, China, 

in 1954. He received the B.S. degree in 

Physics from Northeast Normal 

University, Shenyang, China, in 1981. 

From 1972 to 1991,he was a engineer in 

Northeast Normal University. Since 

1991, he works in Dalian University of 

Technology (DUT). Currently, he is a 

Professor of DUT, and Standing 

Director of China Power Supply Society. 

His research interests include Theory and new technology of 

electrical engineering, mechano-electronic, and preparation of 

magnesium oxide. 



Supply Chain Coordination under Return Policy 

with Asymmetric Information about Cost of 

Reverse Logistics Operations 


Institute of Economics and Management/Anhui Normal University, WuHu, China 

saztl@mail.ustc.edu.cn 

Abstract—In this paper, we study return policy and supply 

chain coordination in a channel of one supplier and one 

retailer. The paper assumes that unsold merchandise should 

been refunded to the supplier by the retailer. The retailer 

knows the cost of reverse logistics operations but the 

supplier has to estimate it. The contract menu under 

asymmetric reverse logistics cost information between 

supply chain members was designed and discussed. The goal 

of the supplier’s contract is to coordinate the channel and 

then get more profit. The problem is analyzed as a 

Stackelberg game in which the supplier declares a contract 

menu with return price and wholesale price to the retailer 

and requires the retailer report the cost of reverse logistics. 

Then the retailer reports the cost and gets the 

corresponding contract. The optimal solutions of the 

contract menu are derived, and numerical examples are 

presented to illustrate the properties of the contract menu. 

Index Terms—-supply chain; return policy; reverse 

logistics; asymmetric information 


Because of the existence of multiple decision makers 

in supply chain, the decisions that are locally optimal can 

be globally inefficient. It is well documented in marketing 

and economics literature that uncoordinated decisions lead 

to “double marginalization”, which is one of the causes of 

channel inefficiency [1],[2]. Coordination among 

suppliers and retailers is a very important strategic issue in 

supply chain management. Coordination between 

independent firms in a supply chain relationship has 

gained much attention recently and many studies have 

been presented. In order to provide compatible incentives 

to improve the supply chain performance and achieve the 

win-win solution, some types of supply chain contracts 

have been discussed. For instance, see return policies [3], 

revenue sharing[4], quantity discount [5], quantity 

flexibility [6], sales rebate[7]. See[8]for excellent reviews. 

The goal of these contracts is to coordinate supply chain, 

which means that the total profit of the decentralized 

supply chain will be equal to that achieved under a 

centralized system. 

In this paper, the focus is on combined contract of 

wholesale price and return policy. Wholesale price is a 

fundamental decision for supply chain coordination in 

distribution channel. The “Quantity discount” is a popular 


doi:10.4304/jcp.6.9.1886-1890 

method used to stimulate the retailer to order [9]. [10] 

shows that in complex supply chain linear quantity 

discount alone cannot coordinate supply chain. 

The return policy is also called as buyback policy for 

many cases and in many researches. This is 

understandable since almost all return policies incur 

buyback price. That is why vast researches pay more 

attention to buyback price. [3]demonstrates that a policy 

to offer full credit to the buyer for a partial return of goods 

may achieve channel coordination and the supplier can get 

any percentage of channel profit by setting proper 

wholesale price and buyback price. There are many 

restrictions about setting. For example, when the retail 

price which affects is endogenous, the buyback contract 

no longer coordinates the supply chain [8]. For more 

complex setting, the other contract will with buyback to 

improve the performance[7],[10]. 

Though both buyback policy and return policy decide 

buyback price, there are remarkable difference between 

them. That is, return policy incurs reverse logistics, but 

buyback policy may not incur reverse logistics. Most 

studies about buyback policy or return policy do not 

consider reverse logistics. Recently, more attention is 

devoted to the logistics of return policies. [11] shows that 

the development of e-commerce in electronic market 

increases the value of surplus products. [12] reviews 

return contract and illustrates the decision of contract 

when consider cost of the return good. [13] investigates a 

supply chain consider the forward logistics and reverse 

logistics simultaneously. 

Most studies to date on return policy have assumed that 

the salvage value is same for all member of supply chain. 

But for some products, such as electronic products and 

books, the excess products should been reprocessed or 

delivered to another substitute channel if resend them to 

the supplier. Therefore, the value of excess goods may be 

higher for the supplier than for the retailer. This is a 

foundation assumption for return policy. On the other 

hand, the paper investigating buyback contract considers 

the logistics is rare. Furthermore, the majority of supply 

chain coordination researches assume a symmetric 

information situation. Because of the variety and 

complexity of logistics activities, the accurate cost 

accounting is troublesome. It is difficult to estimate the 

expenses of the returned purchase when the return


activities are managed by the retailer. This paper considers 

an asymmetric information about the cost of returned 

goods. 

The paper proceeds as follows. The next section 

presents the assumptions and notations. In Section 3, the 

integrated model is discussed firstly. In Section 4, the 

return policy under symmetric information situation is 

investigated. Section 5 focuses on the return policy for an 

asymmetric information relationship. Section 6 gives the 

numerical analysis. Section 7 summarizes the findings. 

II. ASSUMPTIONS AND NOTATIONS 

The demand D is a random within [0, b ] . We denote 

by f, F , µ the density function, distribution function 

y 

of D , respectively. Let E( y) = ∫ xf( x) dx . The retail 

−∞ 

price p and the supplier cost c are exogenous variable 

and the wholesale price of the supplier w is endogenous 

variable. The salvage values of the supplier and the 

retailer are different and denoted by m v and v r , 

respectively. In this paper, we assume the retailer takes 

back work and pays the logistic cost, denoted by l . 

If vr ≥vm − l , from the supply chain point of view, 

returning goods is unreasonable. This paper 

assumes vr < vm −l and considers an asymmetric 

information about the cost of reverse logistics. We 

assume the real value of l is the retailer’s private 

knowledge and we call this retailer l -type retailer for 

convenience in presentation. The supplier does not make 

sure the type of the retailer, but he deems the value of l is 

either l with probability of ρ or l with probability 

of1− ρ . The buyback contract is a practical method for 

the supplier to share risks and losses of the retailer. We 

denote r as the buyback price, which is the decision 

variable of the supplier as well as w . In asymmetric 

information situation, the supplier should offer retailers a 

menu of returns policies trading off l -type retailer 

with l -type retailer. The one goal of the supplier’s 

contract is to coordinate the supply chain and the other is 

to maximize the supplier profit. 

Let l -type retailer’s ordering size is Ql () , the 

expected surplus and sale are Ol () and Sl () . Simply 

calculating gives 

Ql () 

Ol () = ∫ F( xdx ) , Sl () = Ql () −Ol 

() (1) 

0 

The total expected profit of the channel is 

∏ m+ r() l = ( p−c) Q() l −( p− vm+ l) O() l (2) 

The profit of l -type retailer is 

∏ r () l = ( p−w) Q() l −( p− r+ l) O() l 

(3) 

The profit of the supplier is 

∏ () l = ( w−c) Q() l −( r− v ) O() l 

(4) 

m m 

III. THE INTEGRATED MODEL 

The goal of this paper is to develop a return policy to 

coordinate the supply chain. The coordination of supply 


chain means the decision in decentralized enable the 

channel to obtain the same profits as a vertically 

integrated firm’s. In order to give a benchmark for follows 

discussion, in this section, we first focus on an integrated 

structure in which both the supplier and the retailer agree 

to take decisions to maximize the total channel profits 

(joint profit maximization). 

We denote the optimal order size and the maximum 

expected profit of the channel by * * 

Q () l , m r() l ∏ + . Using 

Leibniz’s rule to obtain the first and second derivatives 

shows that m r() l ∏ + is concave. The sufficient optimality 

condition is the well-known formula: 

* 

F( Q ( l)) = ( p− c)/( p+ l− vm). 

(5) 

Using the relationship 

Q 

∞ 

xf( x) dx = µ − xf( x) dx 

∫ ∫ 

0 

and substituting from (5) into (2) and simplifying gives 

the optimal expected profit: 

* * 

∏ m+ r() l = ( p+ l− vm) E( Q ()). l 

(6) 

* * 

Proposition 1. Q () l and m r() l ∏ + decrease as l increases 

* 

Proof. From (5), we have ∂Q ()/ l ∂ l < 0. 

Taking the first- 

* 

order derivative of m r() l ∏ + , one has 

* * * * 

∂ ∏m+ r()/ l ∂ l = E( Q ()) l − Q () l F( Q ()) l < 0. 

The higher l means the higher the cost, thus this 

conclusion is intuitional. 

For the convenience in presentation is follows 

subsections, let 

* 

Q () l 

* 

O () l = ∫ F( x) dx. 

(7) 

0 

IV. THE RETURN POLICY UNDER SYMMETRIC 

INFORMATION SITUATION 

For the sake of comparing, before investigate the 

asymmetric information situation, now we discuss the 

problem of channel coordination by return policy with 

common knowledge about l . When the supplier know 

the retailer’s cost l , the supplier first declares the 

wholesale price w and buyback price r . The retailer, as 

s 

the follower sets the decision of ordering size Q () l . It is 

straightforward to find that only if r > vm+ l , then the 

retailer sends back the excess goods. 

Using the same method gives 

s 

F( Q ( l)) = ( p− w)/( p+ l− r). 

(8) 

s 

l -type retailer’s expected profit, denoted by r () l ∏ , is 

s s 

∏ r ( l) = ( p+ l− r) E( Q ( l)) 

(9) 

From (5) and (8), we get the observation as in Proposition 

2. 

Proposition 2. If wrsatisfy , 

w= βc+ (1 − β) p, r = (1 − β)( p+ l) + βvm 

, (10) 

where l /( l + c −vm) ≤β≤1 the combined contract of wholesale price and buyback 

policy can coordinate the supply chain and has the 

follows properties: 

Q


(i) vm≤r ≤ w, 

∂r/ ∂ w> 0, ∂r/ ∂ vm> 0, ∂r/ ∂ l > 0; 

s * 

(ii) ∏ () l = β∏ s 

* 

(), l ∏ () l = (1 −β) ∏ (); l 

r m+ r m m+ r 

(iii) 0≤1 −β≤( c− vm)/( l+ c− vm) 

Proposition 2 shows that when the supplier takes the 

contract (10) to coordinate the supply chain, the buyback 

price the supplier offers is higher than his savage value. 

Furthermore, the buyback price should increase with the 

increase of wholesale price, the supplier’s savage value 

and the retailer’s logistics cost. If this buyback does not 

incur reverse transportation, the dominant supplier can 

get any percentage of the channel profit. However, when 

the retailer return the excess products, the logistics cost 

will impose restrictions on the supplier’s profit, and the 

higher l leads to the lower profit. 

V. THE RETURN POLICY UNDER ASYMMETRIC 

INFORMATION SITUATION 

Assume now the supplier cannot confirm the retailer 

is l -type or l -type. That is, the supplier has only the 

retailer’s reported cost value, denoted by l f . There is no 

reason to assume that the retailer will report honestly 

“ l f ” as the same as the real value. In this section, we 

discuss how to develop a policy to coordinate the supply 

chain. The key factor for success in coordinating the 

supply chain is to make the l -type retailer report 

honestly, i.e. lf= l Accordance with “the revelation 

principle”[14], the contract should satisfies “incentive 

compatibility constraint” and “participate constraint”, 

simultaneously. 

From Proposition 2, consider the contract 

* * 

menu{( 

w, r, Q ( l)),( w, r, Q ( l ))} : 

w= βc+ (1 − β) p, r = (1 − β)( p+ l) + βvm, 

where l /( l − vm + c) 

≤ β ≤1 

. (11) 

w= βc+ (1 − β) p, r = (1 − β)( p+ l) + βv 

where l /( l − vm + c) 

≤ β ≤1 

The supplier declares the contract menu and requires the 

retailer report her cost of logistics l . He should offer l - 

* 

type retailer the contract ( wrQ , , ( l )) and l -type retailer 

the contract 

that 

by 

* 

( wrQ , , ( l )) . It is straightforward to find 

* * 

{( w, r, Q ( l)),( w, r, Q ( l))} can be substituted 

* * 

{( , Q ( l)),( , Q ( l))} 

β β . On base of analysis in 

subsection 4, we get the retailer’s profit. 

For l -type retailer, if she reports honestly, i.e. lf= l, 

as 

then her profit, denoted by ∏ ( β, 

l) 

, is 

r 

as 

* 

r m r 

∏ ( β, l) = β ∏ + ( l) 

; (12) 

if she reports dishonestly, i.e. lf= l , then her profit, 

as 

denoted by ∏ ( β, 

l) 

, is 


r 

∏ = ∏ +∆ . (13) 

as 

* * 

r ( β, l) β m+ r( 

l) lO ( l) 

m 

For l - type retailer, if she reports honestly, i.e. lf= l , 

as 

then her profit, denoted by ∏ ( β, 

l) 

, is 

r 

as 

* 

r ( , l) m r( 

l) 

∏ β = β ∏ + , (14) 

if she reports dishonestly, i.e. lf as 

denoted by ∏ ( β, 

l) 

, is 

= l , then her profit, 

r 

as 

* * 

∏ r ( β, l) = β ∏m+ r( 

l) −∆ lO ( l). 

(15) 

The incentive compatibility constraint is 

as as as as 

∏r ( β, l) −∏ r ( β, l) > 0, ∏r ( β, l) −∏ r ( β, 

l) 

> 0. 

(16) 

Let 

* 

∏m+ r() 

l ∆lO() l ∆lO() 

l 

all (,) = , bll (,) = , cll (,) = . 

* * * 

∏m+ r() l ∏m+ r() l ∏m+ 

r() 

l 

(17) 

The condition of (16) is same as 

βall (,) + bll (,) ≥ β ≥ βall 

(,) + cll (,) . (18) 

We assume that the reserved profit of the retailer is the 

profit which she can get in decentralized setting without 

the supplier’s buyback policy. In this model, the supplier 

set the optimal wholesale price which maximizes his 

0 

profit. We denote the retailer profit by ∏ r . Hence, the 

participate constraint is 

as 

0 as 

0 

∏r ( β, 

l) 

≥∏r and ∏r ( β , l) 

≥∏ r . (19) 

From (11) and (19), let 

0 ⎧ ∏r 

l ⎫ 

β = Max , 

0 ⎨ * 

⎬ 

⎩∏m+ r() 

l l+ c−vm⎭ . (20) 

0 ⎧⎪ ∏r 

l ⎫⎪ 

β 0 = Max ⎨ , 

* 

⎬ 

⎩⎪∏ m+ r() l l+ c−vm⎭⎪ Therefore, the constraint condition of (11) and (19) can 

been simplified as 

0 0 , β ≥ β β ≥ β . (21) 

Let 

β = all (,) β 

1 0 + cll (,), β = all (,) β 

2 

0 + bll (,) . (22) 

After making clear the conditions which ensures the 

retailer report honestly and achieves the coordination of 

the supply chain, we now discuss the supplier’s decisions 

of β, β which optimizes the supplier’s profit. The 

problem of the supplier is 

as 

* * 

Max ∏ m = ρ(1 −β) ∏ m+ r( l) + ( 1 −ρ) (1 −β) ∏m+ 

r( 

l) 

. (23) 

s..(19) t and (21) 

0≤β≤ 1 and l ≥ l, 

then 

Lemma. If 0 

* * * 

m+ r m+ r 

∏ () l +∆lO () l ≤∏ () l and β ≤ 1 

2 

Proof. Given l , the derivative of 

* 

∏ 

* * 

() l +∆lO () l ≤∏ () l with respect to l is 

m+ r m+ r 

[ 

* 

m+ r( l) * 

lO ( l)]/ l 

* * * * * * 

∂∏ +∆ ∂ 

= E( Q ()) l −Q () l F( Q ()) l −[ E( Q ()) l −Q 

() l F( Q ())] l


and the derivative of 

respect to l is 

* * * 

EQ ( ( l)) − Q( lFQ ) ( ( l)) 

with 

* * * * 

∂[ E( Q ()) l −Q () l F( Q ())]/ l ∂ l =− F( Q ()) l < 0. 

Thus, 

∂∏ +∆ ∂ ≤ , which means 

* * 

[ m+ r( l) lO ( l)]/ l 0 

∏ () l +∆lO() l ≤∏ () l 

.Because 

* * 

m+ r m+ r 

* * * 

β all (,) bll (,) [ () ()]/ () 

2 

m+ rl lO l m+ rl 

2 

≤ + < ∏ +∆ ∏ , so 

β ≤ 1 . 

Eq.(23) shows that the supplier’s problem is a Liner 

Program (LP) and all value coefficient are negative. If the 

restricts of β ≥ β is omitted, the feasible region 

0 

for β , β is illustrated as in Fig.1. Due to the properties of 

LP, the nearest point to origin of coordinates is the 

solution of (23). 

Proposition 3. The solution of the supplier’s problem, 

* * 

denoted by β , β , is 

(i) if β ≤ β , then 

0 1 

* * 

= and β = β 0 

β β 

1 

(as in Fig.1), 

(ii) if β ≤ β ≤ β , then β = β 

1 0 2 

0 and β = β 0 (as in 

Fig.2), 

* 

(3) if β < β ≤ all (,) + bll (,) , then β = β 

2 0 

0 and 

* 

β = [ β − bll ( , )]/ all ( , ) (as in Fig.3), 

0 

* 

(4) if β > all (,) + bll (,) , then (23) has no feasible 

0 

solution. 

0 * 

If ∏r / ∏m+ r( l) ≥ l/( l+ c− vm) 

, it is straightforward to get 

β () l −T 

(,) l l 

0 

1 

0 0 * 

∏r ⎪⎧ ∏r ∏m+ 

r() 

l l ⎪⎫ 

∆lO() 

l 

= −Max , 

* ⎨ * * ⎬− 

* 

∏m+ r() l ⎪⎩∏m+ r() l ∏m+ r() l l+ c−v ∏m 

r() 

l 

m ⎪⎭ 

+ 

. 

0 * 

Proposition 4. If ∏ / ∏ ( l) ≥ l/( l+ c− v ) , then 

β ≤ β and β = β β = β 

0 

1 

* * 

1, 

r m+ r m 

0 

I. NUMERICAL ANALYSIS 

In this section, we assume D is uniform distribution 

within [ 0,b ] . 

By simply operation, one gets 

p+ l−vm all (,) = 

p+ l+∆l− v 

∆l 

, bll (,) = 

p+ l−v , 

p+ l−vm cll (,) = 

( p+ l+∆l−v ) 

m m 

m 

2 

⎧p+ l−vm l ⎫ 

β = Max , 

0 ⎨ ⎬ 

⎩4( p− vr) c+ l−vm ⎭ 

. 

⎧p+ l+∆l− vm l+∆l ⎫ 

β 0 = Max ⎨ , 

⎬ 

⎩ 4( p− vr) c+ l+∆l−vm ⎭ 


* 

β 

β 

1 

β 

0 

β 

β 

2 

β 

1 

β 

all (,) + bll (,) 

β 

0 

β 

2 

β 

1 

C 

A 

B 

β 0 

A 

C 

β 0 

β = all (,) β + bll (,) 

β = all (,) β + cll (,) 

Figure 1. β ≤ β 

β = all (,) β + bll (,) 

A 

C 

β 0 

0 1 

β = all (,) β + cll (,) 

Figure 2. β ≤ β ≤ β 

1 0 2 

β = all (,) β + bll (,) 

D 

β = all (,) β + cll (,) 

β − bll (,) 

0 β = 

all (,) 

Figure 3. β < β ≤ al (,) l + b(,) l l 

2 0 

The sensitivity analysis of β β are listed in Tables 1- 

* 

3. Tables 1-3 show that β , β increase as ∆l increase. 

The greater ∆ l means more uncertainty of the supplier 

about the retailer’s type and it is disadvantageous for the 

* * 

supplier. Table 2 shows that β , β decrease as c 

increases. The higher cost make the supplier ask for the 

higher percentage to ensure enough profit. The 

* , 

* 

* 

E 

1 

E 

1 

E 

1 

β 

β 

β


higher c and the higher ∆l should decrease the supplier’s 

* 

profit. Table 4 shows that β , β increase as p increases, 

this is easy to understand because the higher retailer price 

leads to the higher reserved profit. 

(i) c = 40, p = 80, v = 25, v = v −∆ v, l = 2, l = l+∆ l 

* 

* 

* 

m r m 

Table 1. β , β vary with ∆v and ∆ l 

∆ l 

* 

β 

* 

β 

* 

β 

* 

β 

* 

β 

∆ v = 12 

∆ v = 16 

∆ v = 20 

2 0.225 0.214 0.219 0.210 0.219 0.210 

4 0.282 0.285 0.282 0.285 0.282 0.285 

6 0.329 0.347 0.329 0.347 0.329 0.347 

8 0.364 0.4 0.364 0.4 0.364 0.4 

10 0.390 0.444 0.390 0.444 0.390 0.444 

(ii) c = 40,50,60, p = 80, v = 25, v = 10, l = 2, l = l+∆ l 

* 

* 

Table 2. β , β vary with c and ∆ l 

∆ l 

* 

β 

* 

β 

m r 

* 

β 

c = 40 

c = 50 

c = 60 

2 0.219 0.210 0.219 0.210 0.219 0.210 

4 0.283 0.285 0.218 0.217 0.218 0.217 

6 0.329 0.347 0.233 0.242 0.217 0.225 

8 0.364 0.4 0.264 0.285 0.217 0.232 

10 0.390 0.444 0.288 0.324 0.229 0.255 

12 0.410 0.482 0.308 0.358 0.247 0.285 

(iii) 

c = 40, p = 70,80,90,100, v = 25, v = 10, l = 2, l = l+∆l * 

* 

Table 3 β , β vary with p and ∆ l 

∆ l 

* 

β 

* 

β 

* 

β 

* 

β 

m r 

p = 70 

p = 80 

p = 90 

2 0.221 0.210 0.225 0.214 0.226 0.215 

4 0.281 0.285 0.281 0.285 0.282 0.285 

6 0.325 0.347 0.329 0.347 0.331 0.347 

8 0.357 0.4 0.364 0.4 0.369 0.4 

10 0.380 0.444 0.390 0.444 0.398 0.444 

* 

β 

II. CONCLUSIONS 

This paper has formulated a supply chain coordination 

problem with asymmetric information between one 

supplier and one retailer for a single-period product. This 

paper assumes that the salvage value of unsold products 

is higher for the supplier than for the retailer. The 

supplier wants to coordinate by proper return contract. In 

this return policy we assume that the excess goods 

refunded by the retailer and the cost of reverse logistics is 

asymmetric information. This paper formulates a contract 

menu with return price and wholesale price. The 

observations are developed and show that this contract 

menu enables the retailer report the logistics cost honestly 

and can achieve the coordination. The solution of this 

contract menu is derived, and the numerical examples 

illustrate that the greater variation of the supplier’s 

estimation about the logistics cost is disadvantageous for 

the supplier. The greater variation will produce more 

harm to the supplier who has the higher cost. 


* 

β 

* 

β 

* 

β 

* 

β 

* 

β 

There a number of possible extensions of our study 

that can constitute future research endeavors in this area. 

One immediate extension is to consider the cooperating 

reverse logistics between the channel members. 

Developing better contract menu to deal with the 

asymmetric information is another interesting theme. 


The authors would like to acknowledge the supports of 

research Grants from National Natural Science 

Foundation of China (Project No. 70901001) and Anhui 

Provincial Natural Science Foundation (Project No. 

090416244). 

REFERENCES 

[1] Spengler, J., 1950. Vertical integration and anti-trust 

policy. Journal of Political Economy 58, 347-552. 

[2] Tirole, J., 1989. The Theory of Industrial Organization. 

MIT Press, Cambridge MA. 

[3] Pasternack, B.A., 1985. Optimal pricing and return policies 

for perishable commodities. Marketing Science, 4 (2), 166- 

176. 

[4] Cachon, G., Lariviere, M., 2005. Supply chain 

coordination with revenue sharing: Strengths and 

limitations. Management Science, 51 (1), 30-44. 

[5] Weng, Z.K., 1995. Channel coordination and quantity 

discounts. Management Science, 41 (9), 1509-1521. 

[6] Tsay, A.A., 1999. The quantity flexibility contract and 

supplier–customer incentives. Management Science, 45 

(10), 1339-1358. 

[7] Taylor, T.A., 2002. Supply chain coordination under 

channel rebates with sales effort effects. Management 

Science, 48 (8), 992-1007. 

[8] Cachon, G., 2003. Supply chain coordination with 

contracts. In: Graves, Steve, de Kok, Ton (Eds.), 

Handbooks in Operations Research and Management 

Science: Supply Chain Management. North Holland, 

Amsterdam. 

[9] Munson, C.L., Rosenblatt, M.J., 2001. Coordinating a 

three-level supply chain with quantity discounts. IIE 

Transactions 33, 371–384. 

[10] Krishnan, H., Kapuscinski, R.K., Butz, D.A., 2004. 

Coordinating contracts for decentralized supply chain with 

retailer promotional effect. Management science, 50 (1), 

48-62. 

[11] Choi,T.M., Li, D., Yan,H., 2004. Optimal returns policy 

for supply chain with e-marketplace. International Journal 

of Production Economics, (88), 205-227. 

[12] Tsay, A., 2001. Managing retail channel overstock: 

markdown money and return policies. Journal of Retailing 

77, 457–492. 

[13] Ferguson, M., Jr, V. G., Souza, G., 2004. Supply chain 

coordination for false failure returns.Georgia Institute of 

Technology ,Working paper. 

[14] Fudenberg, D., Tirole, J., 1991. Game Theory, The MIT 

Press, Cambridge, Massachusetts, London, England. 

Ting Long Zhang, Ph.D., Associate Professor of Institute of 

Economics and Management, Anhui Normal University. Field 

of Research: Management Science, Supply Chain Management, 

Logistics.


Economic Development and Financial Support 

for Coal Resource Cities 

— A Panel Data Analysis 

Zuhuai Yuan 

School of Management/China University of Mining and Technology, Xuzhou, China 

E-mail: yzhhn@tom.com 

Li Yang 

School of Management/Anhui University of Science & Technology, Huainan, China 

E-mail: yangli081003@163.com 

Jing Han 

Huainan Vocational & Technical College, Huainan, china 


Keliang Wang 

School of Management/Anhui University of Science & Technology, Huainan, China 

E-mail: klwang@163.com 

Abstract — This paper uses measurement methods and 

selects relevant indicators from both quantitative and 

structural aspects, empirically analyses the relationship 

between financial development and economic development 

of 2000-2008 in more than 18 coal- resourced cities in China. 

The results show that financial development takes a 

significant role in the economic development of 

coal-resourced cities. However, the high industry 

concentration of financial resources leads to a decline in 

financial resource allocation efficiency. 

Index Terms—Resource-based city; Economic development; 

Financial support 


Coal City is an important component of the urban 

system in China. According to the survey, there are 68 

coal cities nationwide, accounting for 38.2% of the total 

number of mining cities, 10.3% of the total number of 

cities; supplying for 93.6% of the coal to national 

economic construction [1]. Relying on coal resources, 

coal cities make huge contributions in national 

urbanization process, in promoting national economic 

development, and expansion of social employment. 

Meanwhile,a group of highly coal-relied cities formed 

accordingly. With the depletion and reduction of 

resources, coal cities, like many other mining cities, will 

face with the "mine dry up, city fall" threat. Finance is the 

core of modern economy. Between financial development 

and economic development, there is an inherent 

mechanism. Coal-based cities can’t develop sustainably 

without effective financial support. Many scholars at 

home and abroad used empirical analysis methods, 


doi:10.4304/jcp.6.9.1891-1895 

verified the quite significant relationship between 

financial development and economic development. For 

instance, Goldsmith (1969), who started the earliest 

quantitative research on financial and economic relations, 

found out simultaneous development of finance and 

economy, a period of rapid economic growth always went 

along with the ultra level financial development [2]. After 

introducing other factors that affect economic 

development, King and Levine studied the relevant data 

during 1960-1989 of 80 countries. It shows that the 

financial development and economic growth in a positive 

correlation [3]. Han Tingchun uses causality tests on the 

relevant data of 1981-2002 in China. It shows that from 

1981 to 1991 financial development and economic 

growth causality is not obvious, however, during 

1992-2002, financial development is the direct cause of 

economic growth [4]. Tan Ruyong, after studied by OLS 

on the quater data of China in 1993-1998, concluded that 

China's financial intermediary development and 

economic growth have a significant posotive correlation 

between each other [5]. Cao Tingqiu and Wang Xihang 

studied the panel data of 1995-2001 in each region in 

Shandong, among sample area, both finance and 

economy growth in an obvious trend, the relations among 

areas certain difference[6]. The above studies objects are 

mainly nations and provinces. According to the data that 

the author has, as now, coal-based city development is 

also limited to the capital, labor, technology integration 

and other traditional elements [7]. Coal-based urban 

development studies which put into the relationship 

between financial development and economic 

development are mostly qualitative and case studies, 

empirical research literature are still few. Therefore, the 

empirical study of economic development of coal


resources city's financial support issues, has certain 

practical and theoretical significance. 

This assay will target on 18 prefecture-level coal-based 

cities1, select panel data of these cities in 2000-2008, use 

Eviews5.0 software, empirically analyze quantitative and 

structural support issues in the economic development of 

coal cities, to provide some ideas for exploring financial 

support for the sustainable development of coal resource 

cities . 

The research data come mainly from China's 

economic statistical databases and related city 

government statistics website. 

II. EMPIRICAL ANALYSIS OF FINANCIAL 

SUPPORT TO COAL-BASED URBAN 

ECONOMIC DEVELOPMENT 

A regional financial development matching with local 

economic development includes two aspects. In 

quantitative aspect, financial element input adjust to 

regional economic development requirement, push the 

arise of gross regional economy. In structural aspect, 

internal economic factor delivery structure is compatible 

with regional economic structure, which promotes the 

adjustment of economic structure optimizing and 

upgrading. Therefore, this paper will be about empirical 

analysis on the relationship between financial 

development and economic development of coal resource 

cities from quantitative and structural support aspects. 

A. Quantitative support analysis (quantity effect) 

a. Theoretical model construction and variable 

description 

Economic development is inseparable from human, 

finance and material. Production function is the most 

common used model in the quantitative study about 

economic development. Traditional production function 

mainly inspects on the relationship between production 

element labor and material input and output variables 

among each other. The above literatures have proved the 

close relationship between financial development and 

economic growth and the regional financial resource 

increase can push regional economic development. This 

paper uses Cobb-Douglas production function as the 

basic model, through the introduction of variable 

financial scale which influences economic development, 

to study the quantitative match relationship between 

financial development and economic expansion: 

Q=AKαLβFγ (1) 

Take on both sides of the model number: 

LNQ = LNA+αLNK + βLNL+ γLNF (2) 

Here, Q for economic development: as the economic 

development indicators, the existing papers have chosen 

GDP, GDP growth rate or per capita GDP, this article 

chose GDP as economic growth rate indicator variables, 

and the unit is ten thousand Yuan. 

A for integrated function coefficient. 

1 It refers to Chifeng, Datong, Huaibei, Hegang, Hebi, Huainan, 

Jincheng, Jixi, Jiaozuo, Pingdingshan, Qitaihe, Shuangyashan, 

Shuozhou, Wuhai, Xianyang, Yulin, Yangquan, Zaozhuang. 


K for material input: indicate as fixed assets 

investment amount in current year with reference to 

previous research, the unit is ten thousand Yuan. 

L for labor input: taking into account "the number of 

unit employees" statistical coverage limitations and data 

sources availability, this paper replaces it by district 

population at the end of year, unit is ten thousand people. 

F for the financial investment: taking into account 

data availability, this paper replaces it by the loan balance 

of regional financial institutions; the unit is ten thousand 

Yuan. 

α, β and γ each represent the output elasticity of 

capital, labor output elasticity of output and financial 

flexibility. 

b. Econometric model analysis 

For the study of the overall characteristics and 

differences between cities in coal resources city between 

financial development and economic development, this 

paper takes use of Eviews5.0 to establish following 

econometric model. 

i. Mixed estimated model 

Most of the coal resource cities passed a similar 

development path, therefore they should have the same 

characteristic of economic and financial relations. By 

ordering software outputs from Eviews5.0, draw the 

conclusion: 

LNQ =0.739+0.396LNK + 0.177LNL+ 0.46LNF (3) 

(5.31) (21.20) (6.76) (11.63) 

2 

R2=0.9998 R =0.9998 F=243201 P=0.000 DW=0.58 

See from the related indicators of output equation, the 

estimated model fit good(R 2), the overall equation is 

significant(passed the F test), the T value tests of single 

parameters are satisfactory, the estimate results are 

reliable and reasonable from the economic sense. From 

the elasticity of each variable, during the sample interval 

2000 -2008, financial development played a much more 

important role,and contributed more to the economic 

development of coal resource cities. The overall financial 

development adapted to the urban economic 

development. As the level of the financial scale increased 

1%, the total regional economy will increase 0.46%. 

ii.Phased mix estimated model 

Due to the impact by external policy, economic 

environment changes and changes in the city’s own stage, 

coal resource cities may show different relationship 

between economic and financial development in different 

phases. Use Eviews5.0 to analyze phases and organize 

available: 

2000-2001: 

LNQ = 1.838+0.11LNK + 0.35LNL+ 0.46LNF (4) 

(4.3) (2.05) (3.93) (6.5) 

2 

R2= 0.9998 R =0.9998 F=266630 P=0.000 DW=0.56 

2002-2003: 

LNQ =1.426+0.216LNK + 0.302LNL+ 0.46LNF (5) 

(5.23) (6.60) (6.76) (8.19) 

2 

R2=0.9999 R =0.9999 F=156612 P=0.000 DW=0.54


2004-2005: 

LNQ =1.178+0.357LNK + 0.30LNL+ 0.38LNF (6) 

(3.28) (5.72) (5.38) (3.89) 

R 2= 0.9998 2 

R =0.9998 F=74307 P=0.000 

DW=1.00 

2006-2008: 

LNQ = 1.32+0.42LNK + 0.22LNL+ 0.32LNF (7) 

(8.89) (14.3) (9.36) (7.62) 

R2=0.9999 

2 

R =0.9999 F=192508 P=0.000 DW=0.78 

In various stages, hybrid estimation model fitness, the 

overall equation satisfaction and T value test are both 

fine. Comparing the above-mentioned equation elements 

output elasticity coefficient, we find out: firstly, coal 

resources cities, as investment-driven cities, whose 

overall features are increasingly strengthen. Fixed assets 

investment output elasticity coefficient α increased in the 

sample time interval from 0.11 to 0.42, investment in the 

development of coal resources cities is getting 

increasingly important; secondly, input-output elasticity 

coefficient of labor β negatively developed, the growing 

role of regional economic development reduced,it is 

more in line with the reality that in recent years, cities 

step up modernization of coal mine construction, mine 

reduced capital investment and the actual labor; thirdly, 

financial output elasticity coefficient γ fell to 0.32 from 

0.46, indicating that financial development of coal 

resources based on the contribution of urban economic 

development is declining, the output effect of credit funds 

is gradually decreased. 

iii.Variable coefficient estimation model 

In order to find out the difference between coal 

resource cities of financial development and economic 

development, this paper establishes vary coefficient 

estimation model for financial development variable. 

Relevant output situation ordered by Eviews5.0 is as 

Table 1. 

Through comprehensive analysis of the financial 

output elasticity coefficient of each city in Table 1, it can 

be drawn that during the sample interval, the coal cities 

financial development and economic development are 

positively correlated in case other factors not considered, 

but the contribution varies a certain from level of the 

financial scale of the regional economic development. 

Output elasticity in [0.3239 0.3817], with an average 

output elasticity of 0.3596, the standard deviation of 

output elasticity is 0.014. 

Table1: Coal-based city financial input-output coefficients situation 

Variable Coefficient Std. Error t-Statistic Prob. 

LNA 0.6227 0.8174 0.7618 0.4475 

LNK? 0.4224 0.0179 23.6284 0.0000 

LNL? 0.4312 0.4152 1.0385 0.3008 

Cf 0.3391 0.0672 5.0440 0.0000 

Dt 0.3560 0.0593 6.0032 0.0000 

Hb 0.3560 0.0563 6.3246 0.0000 

Hb 0.3715 0.0524 7.0916 0.0000 

Hg 0.3582 0.0497 7.2111 0.0000 

Hn 0.3503 0.0552 6.3440 0.0000 

Jc 0.3565 0.0548 6.5011 0.0000 

Jx 0.3757 0.0537 6.9994 0.0000 

Jz 0.3659 0.0619 5.9159 0.0000 

Pds 0.3651 0.0683 5.3459 0.0000 

Qth 0.3727 0.0511 7.2983 0.0000 

Sys 0.3568 0.0515 6.9303 0.0000 

Sz 0.3648 0.0523 6.9708 0.0000 

Wh 0.3818 0.0561 6.8048 0.0000 

Xy 0.3394 0.0686 4.9500 0.0000 

Yl 0.3239 0.0631 5.1334 0.0000 

Yq 0.3724 0.0504 7.3911 0.0000 

Zz 0.3676 0.0630 5.8376 0.0000 

R 2 =0.9997 

2 

R =0.9997 F=31790 P=0.000 DW=1.01 

B. Structural support (structural effect) 

The transformation of resource-based cities is 

essentially a process of economic restructuring, namely, 

an industrial restructuring process. The study result of 

quantitative effect between financial development and 

economic development reflects that financial 

development has become an important factor in the 

development of coal resources cities, but in different 

years or different cities there are some differences in this 

role, that is, the same financial development can not 

produce the same total amount of economic success, the 

financial input-output affects differently. This difference 

should be a structural difference, that is, the fit issues 

between financial development and economic 

development structure. 

a. Two phases comparison and analysis 

Compare the financial output elasticity coefficient with 

industrial structure changes in two stages (2000-2003, 

2004-2008) (see Table 2) in coal-resourced cities, known: 

during 2004-2008, among the sample cities, the average 

proportion of secondary industries in other cities except 

Jixi, Wuhai has increased compared with 2000-2003. But 

along with the increase in the proportion of secondary 

industry, the financial outputs are not synchronized grow, 

the overall trend of both is always on reverse, the 

proportion of changes in the opposite direction was 

94.4%. 

Table 2: Financial output coefficient and industrial structure diversification in coal resource cities 

Cf Dt Hb Hb Hg Hn Jc Jx Jz 

coefficient 2000-2003 0.38 0.37 0.35 0.34 0.31 0.36 0.36 0.35 0.39 

2004-2008 0.13 0.18 0.19 0.22 0.25 0.18 0.20 0.23 0.17 

variation symbol - - - - - - - - - 

Sec. 

2000-2004 33.9 53.9 50.1 50.1 39.9 48.4 55.2 40.8 52.7 

industry proportion 

2004-2008 43.1 54.2 56.3 60.5 42.2 55.2 63.9 32.7 63.7 



variation symbol + + + + + + + - + 

Pds. Qth. Sys. Sz. Wh. Xy. Yl. Yq. Zz. 

coefficient 2000-2003 0.41 0.31 0.33 0.33 0.27 0.38 0.33 0.34 0.40 

Sec. 

industry 

proportion 

2004-2008 0.15 0.26 0.22 0.23 0.31 0.13 0.15 0.24 0.17 

variation symbol - - - - + - - - - 

2000-2004 53.4 48.8 38.8 48.9 66.8 42.6 51.4 58.7 51.8 

2004-2008 61.6 52.6 42.2 60.0 66.3 45.6 69.2 59.5 63.0 

variation symbol + + + + - + + + + 

Remarks: Each city's financial output coefficient in the second phase is the result from Eviews5.0. Industrial structure is the 

average which secondary industry added-value in various stages accounted for GDP. 

b. Allocation efficiency of financial industry 

For further analysis of reverse problems between 

financial development and economic growth in coal 

resource cities, and also take the data availability into 

consideration, this paper selects Huainan2 City as the 

typical case to study financial structure and economic 

structure relationship, and establishes panel analysis 

model based on regional industrial structure and credit 

structure in 2000-2008 as a substitute for economic and 

financial structure. 

i. Development of industrial and credit structure 

3000000 

2500000 

2000000 

1500000 

1000000 

500000 

0 

2000 2001 2002 2003 2004 2005 2006 2007 2008 

3500000 

3000000 

2500000 

2000000 

1500000 

1000000 

500000 

增加值第一产业增加值第二产业增加值第三产业 

贷款余额第一产业贷款余额第二产业贷款余额第三产业 

Figure 1: The added value of industries and the loan balance in recent 

years in Huainan City 

Unit: Ten thousand yuan 

In recent years, Huainan City, increasing credit funds 

from financial institutions increasingly focused on 

coal-based secondary industry whose proportion of loans 

increased from 70.7% in 2000 to 79.2% in 2008. 

Industrial and customer concentration of financial 

institution loans is quite high3; meanwhile, the status of 

the second pillar industry is consolidated continuously, 

and the proportion of the added value of the secondary 

2 As a coal resource-based city built by coal mines, Huainan has a very 

rich coal resources of 44.4 billion tons vision, proven reserves of 15.3 

billion tons, accounting for 32% of east China.In 2008, coal and power 

industries achieved added value of 20.9 billion yuan, accounting for the 

city's industrial added value, GDP 84.6%, 46.13% respectively. 

3 By the end of 2008, the year-end value of loans was 24.482 billions of 

9 coal mining and power enterprises, including Huainan Mining (Group) 

Co., Ltd and Huainan-Zhejiang coal and power Co., Ltd, accounting for 

64.38 percent of the aggregate value of loans, 42.3 percentage points 

higher than that in 2001. 


0 

industry accounting for the regional GDP rose to 61.1%4 

in 2008 from 46.6% in 2000. Regional industrial structure 

adjustment pressure increased. 

ii. Panel data estimation 

The results of the Eviews5.0 output of coal cities 

reached the relationship between credit structure and 

industrial structure as the following (Table 3). 

Table 3: The relationship between financial structure and industrial 

structure in coal resource cities 

2000- 

2003 

2004- 

2008 

Variable Coefficient Std.Error t-Statistic Prob. Statistics 

Y--F 0.1296 0.1938 0.6688 0.5285 R 2 =0.999 

E--F 1.1978 0.4170 2.8728 0.0283 F=174815 

S--F 0.6296 0.0746 8.4362 0.0002 P=0.000 

Fixed 

Effects 

Y-C=4.750 E-C=-1.209 S-C=2.265 DW=2.83 

8 

Y--F 0.0920 0.0314 2.9301 0.0168 R 2 =0.999 

E--F 0.6823 0.1616 4.2232 0.0022 F=4753 

S--F 1.1078 0.4682 2.3659 0.0422 P=0.000 

Fixed 

Effects 

Y-C=5.06 E-C=1.947 S-C=-0.330 DW=2.39 

6 

Remarks: Y, E and S represent primary, secondary, tertiary 

industrial added value logarithm respectively, F Y, FE, FS, each 

represents primary, secondary, tertiary industrial credit balance 

logarithm at end of year in same period. Y-C, E-C, and S-C are 

the permanent effects of each industrial development. 

The relevant data in Table 3 show that with the 

financial development, the credit loan output effect in 

each industry changed: firstly, the credit loan elasticity 

coefficient in primary and secondary industries 

respectively declined 0.038 and 0.515, in which the 

decline degree of credit loan elasticity coefficient of 

tertiary industry obviously exceeds the growth rate of 

tertiary industry. In three industrial credit loan 

configuration efficiency, secondary funds configuration 

efficiency has been the highest in 2000-2003 dropped to 

the second highest in 2004-2008. The decline in the fund 

allocation efficiency and the increasing concentration of 

credit loans to coal-based secondary industry, co-led the 

4 In 2008, the added value of coal and power industries was up to 20.9 

billions in Huainan, accounting for 84.6 percent of the aggregate 

industrial added value and 46.13 percent of the regional GDP.


overall decreasing efficiency of the allocation of credit 

funds. 

C. Overall conclusion 

The research which analyzes the coal city’s financial 

support for economic development from two 

perspective--"quantity" and "structure" shows: Under the 

circumstance that investment boosts growth is still the 

growth mode of coal resource cities, financial 

development has clear effect and high contribution to 

economic development, generally adapted to the 

economic development need, but the contribution to 

economic presents falling trend. Coal resource city's 

economic structure and credit structure show a trend of 

mutual strengthening. Credit funds are increasingly 

concentrating to coal-based secondary industry; financial 

support for primary and tertiary industries is weakening. 

The high industry-concentration and high 

customer-concentration features of credit loans finally 

lead to the financial assets allocation efficiency decline in 

coal resources cities. 

III. STRATEGIES FOR PROMOTING FINANCIAL 

SUPPOTT FOR COAL-BASED CITTIES 

SUSTAINABLE DEVELOPMENT 

A. Promoting financial development 

Establish a sound financing system for city 

transformation, vigorously develop local financial 

institutions, and actively introduce joint-stock 

commercial banks, strengthen the financial viability of 

coal resources city; develop direct financing and give full 

play to support resource-based capital market in the role 

of urban transformation, to encourage support of large 

coal companies using short-term financing bills, bonds 

and notes, reducing large-scale enterprises, enterprises 

with financial dependence on bank credit and small and 

medium private enterprises of the credit squeeze effect; 

strengthen the economic and financial information 

exchange, and promote political bank-enterprise 

communication and collaboration; strengthen the credit 

system building, create a favorable sustainable 

development of coal resources city a good financial 

environment. 

B. Optimize credit loan structure 

In the meantime of actively supporting local 

characteristic industries and pillar industries, financial 

institutions should also increase effective credit loan 

release focusing on funds need of resource-based 

economy transformation follow-up industries. Financial 

institutions should reasonably allocate credit resources, 

improve the fund input intensity for tertiary industry and 

other high credit fund allocation efficiency industries, 

improve credit fund application efficiency; increase the 

support intensity for recycling economy, bio-medical and 

other high-tech industries, avoid further solidity of 

coal-resource cities economy and financial structure 

result from excessive credit concentration, and more 

difficulties in resource cities transformation; enhance risk 

assessment and monitoring, strengthen and improve the 


awareness of energy financial risks, to effectively protect 

credit fund security in urban transformation. 

C. Adjust investment structure 

Single industrial structure is a major issue which 

resource cities will face in its development, for which 

resource cities are inevitably face "resource-curse". 

"Today's investment structure is tomorrow's industrial 

structure." An effective program for the curse is to reduce 

dependence on resource sectors, that is, the 

implementation of industrial diversification. Relevant 

government departments should encourage coal 

companies "based on coal," as well as make "extended 

coal" article, stimulate large-scale coal enterprises 

achieve diversification business and development 

methods transfer; actively guide the private capital to 

increase investment to non-coal tertiary industries, do a 

good "surpass coal" article, particularly increase 

investment to high-tech industries which have major 

breakthrough and stimulates effect to economic growth, 

provide new credit carrier for financial sectors, eventually 

through "incremental" tune "deposit quantity "approach 

to achieve industrial structure optimization and upgrading 

of coal resource cities . 


This work is supported by the National Natural 

Science Foundation of China(71071003), the MOE 

Project of Youth Foundation of Humanities and Social 

Science(09YJC630004), and the Anhui Philosophy 

Society and Science Projects (AHSK05-06D55). 

REFERENCES 

[1] Sheng Kerong, Sun Wei. “Discussing the factors of 

economical development of the coal-based cities in China,” 

Mining Research and Development, vol.24, mar.2004,pp. 

[2] Goldsmith. 1969, Financial Structure and Development, 

New.haven, CT: Yale. University press. 

[3] King and Levine. 1993, Finance and Growth: Schunpeter 

might be Right, Quarterly Journal of Economics, vol.108, 

pp.717-738. 

[4] Han Tingchun. “Financial development and economical 

growing: a positive analysis of China,” Economical 

technology, Vol.3, 2001, pp 

[5] Tan Ruyong. “A positive analysis of the relationship 

between finacial development and economical growing in 

china,” Economical Research, vol. 10, 1999, pp 

[6] Cao Tingqiu, Wang Xihang. “Financial development and 

economical growing: a positive analysis on the cities in 

Shandong province,” Shandong society technology, vol. 

1, 2006, pp. 

[7] Tang Jianying, Zhou Dequn etc. “A positive analysis on 

the variation of the total factor productivity of the coal 

cities in China,” Journal of China University of Mining & 

Techonology, vol. 11, 2007, pp. 

Zuhuai Yuan is currently a doctor candidate in the School of 

Management at the University of Mining and Technology of 

China, Xuzhou, Jiangsu, China. As well, he is the director of the 

municipal research institution, Huainan, Anhui, China. (E-mail: 

yzhhn@tom.com).


Solving the Sparsity Problem in Recommender 

Systems Using Association Retrieval 

YiBo Chen 

Computer school of Wuhan University, Wuhan, Hubei, China 

chenyibo8224@yahoo.com.cn 

ChanLe Wu, Ming Xie and Xiaojun Guo 

Computer school of Wuhan University, Wuhan, Hubei, China 

National Engineering Research Center for Multimedia Software, Wuhan, China 

Email:{chanle.wu, Guoxiaojun}@gmail.com 

Abstract—Recommender systems are being widely applied 

in many fields, such as e-commerce etc, to provide products, 

services and information to potential customers. 

Collaborative filtering as the most successful approach, 

which recommends contents to the current customers 

mainly is based on the past transactions and feedback of the 

similar customer. However, it is difficult to distinguish the 

similar interests between customers because the sparsity 

problem is caused by the insufficient number of the 

transactions and feedback data, which confined the usability 

of the collaborative filtering. This paper proposed the direct 

similarity and the indirect similarity between users, and 

computed the similarity matrix through the relative distance 

between the user’s rating; using association retrieval 

technology to explore the transitive associations based on 

the user’s feedback data, realized a new collaborative 

filtering approach to alleviate the sparsity problem and 

improved the quality of the recommendation. In the end, we 

implemented experiment based on Movielens data set, the 

experiment results indicated that the proposed approach 

can effectively alleviate the sparsity problem, have good 

coverage rate and recommendation quality. 

Index Terms—collaborative filtering; association retrieval; 

sparsity problem; recommendation quality 


Along with the rapidly development of the Internet, the 

number of the servers connected to Internet and the Webs 

on WWW show a trend of exponential growth. The 

rapidly development of the Internet present a mass of 

information to us at the same time, for example, there are 

tens of thousands movies in Netflix, millions of books in 

Amazon, more than 10 billion page collection in 

Del.icio.us, so much information, not to mention find 

some interesting contents, it is impossible that to gave all 

of information the once-over. The traditional search 

Manuscript received January 1, 2010; revised June 1, 2010; accepted 

July 1, 2010. 

Copyright credit, project number, corresponding author, etc. 


doi:10.4304/jcp.6.9.1896-1902 

algorithm only presents the same ordered results to all of 

users; can not to provide different service to different 

users according to their different interests The 

information explosion reduced the use ratio of the 

information, this phenomenon is called information 

overload. Personalized recommendation, included 

personalized search, has been thought as one of the most 

effective tools to resolve the problem of information 

overload. Radically, the recommendation problem is to 

substitute user to evaluate the strange products, which 

include books, movies, CD, web and so on, it is a process 

from know to unknown [1]。. 

Recommendation as a social process plays an 

important role in many applications for consumers, 

because it is overly expensive for every consumer to learn 

about all possible alternatives independently. Depending 

on the specific application setting, a consumer might be a 

buyer, an information seeker, or an organization 

searching for certain expertise [2]. 

Until 1990s, personalized recommendation research as 

an independent concept be advanced. It rapidly 

development origin from the web2.0‘s maturity, which 

make the user become a participant from browser. In an 

actually recommender system, there are tens of 

thousands, or even more than one millions products need 

to be recommended, for instance, Amazon, eBay, 

Youtube, etc, also there are huge users. Accurate and 

high-performance recommender system can mine the 

potential propensity to consume of the user and provide 

personalized services for users. In the increasingly fierce 

competitive environment, personalized recommendation 

system is not just business marketing means, more 

importantly, it can improve the user‘s loyalty and prevent 

the loss of users. 

A recommender system is compose of three parts: 

action recorder module collect the user‘s information, 

model analysis module analyze the user‘s preference and 

recommendation algorithm module, thereinto, the 

recommendation algorithm module is the most core part 

of the recommendation system [3]. At present, 

recommendation algorithm mainly includes collaborative 

filtering algorithm, content-based algorithm, the bipartite


relationship graph recommendation algorithm based on 

user-product and hybrid recommendation algorithm. This 

paper focus on the sparsity and precision problem, 

compute the similarly matrix through the relatively 

distance between the user‘s rating and use the association 

retrieval technology to realize a new collaborative 

filtering approach. 

The remainder of the paper is organized as follows. 

Section 2 surveys existing work on collaborative filtering 

and discusses the sparsity problem in detail. Section 3 

introduces associative retrieval and summarizes our 

associative retrieval-based approach to dealing with the 

sparsity problem and improve the quality of the 

recommendation. Section 4 presents an experimental 

study and the experimental data analysis. Section 5 

concludes the article by summarizing our research 

contributions and pointing out future directions. 

II. COLLABORATIVE FILTERING AND THE SPARSITY 

PROBLEM 

A. Collaborative filtering 

Collaborative filtering aggregates the experiences of 

similar users in the system to generate personalized 

recommendations. One key aspect of collaborative 

filtering is the identification of users similar to the one 

who needs a recommendation depends on the preference 

patterns of users makes it more general than other tasks 

such as ad hoc information retrieval and content-based 

filtering [4]. 

Collaborative filtering has been the most successful 

recommendation system approach to date and has been 

widely applied in various applications, thereinto, Grundy 

have been considered the first collaborative filtering 

system [5]. Grundy system can build user‘s preference 

model to recommend relevant books to every users. 

Tapestry mail processing system, manpower deal with the 

similarity between users. The more users, the lower 

precision [6]. GroupLens build the user‘s information 

group, within group of users can publish their own 

information, and with other users make collaborative 

recommendation [7]. Ringo make use of the same social 

information filtering method to recommend music to 

users [8]. There are some other typically collaborative 

recommendation system, such as Amazon.Com [9], Jester 

[10], Phoaks [11], and so on. 

Many algorithms have been proposed to deal with the 

collaborative filtering problem. Most collaborative 

filtering algorithms can be categorized into two classes 

[12]: Memory-based algorithms and model-based 

algorithms. 

The memory -based algorithms first find the users from 

the training database that are most similar to the current 

test user in terms of the rating pattern, and then combine 

the ratings given by those similar users to obtain the 

prediction for the test user. The two most commonly 

methods is Pearson correlation and cosine of the angle. 

Many enhanced method have been applied into the 

Pearson correlation and cosine of the angle. For example, 

absentee voting, case extended, weighted advantage 

predication, etc. Otherwise, Chen and Cheng make use of 


the order within product list to compute the similar 

degree between users; the high-order products have 

higher weight when computing the user‘s comparability 

[13]. But Yang and Gu proposed that using user‘s 

behavior information to construct the user‘s interest 

point, make use of the interest point to compute the 

comparability [3][14]. 

Model-based algorithm collects rating data to study, 

infer the user‘s action model, and predicate rating for a 

product. The difference between model-based 

collaborative filtering and memory-based collaborative 

filtering is that model-based approach not based on some 

of heuristic rule to predicate, but based on data 

application statistics and machine learning to get model 

to predicate. Breese et al. proposed two selection 

probability models: Clustering model and Bayes network 

[15]. In first model, suppose the user‘s rating 

independently, the similarly user cluster into a class, give 

the user class a mark number. In Bayes network, the 

number of class and model parameter can obtain from 

existing data through learning. Other model-based 

collaborative filtering system have probability correlation 

model [16], maximum entropy model, linear regression 

model, and so on. 

Despite its success in many application settings, the 

collaborative filtering approach nevertheless has been 

reported to have several major limitations including the 

sparsity, scalability, and synonymy problems. The 

sparsity problem occurs when transactional or feedback 

data is sparse and insufficient for identifying neighbors 

and it is a major issue limiting the quality of 

recommendations and the applicability of collaborative 

filtering in general. Our study focused on developing an 

effective approach to making high-quality 

recommendations even when sufficient data is 

unavailable. 

B. The sparsity problem 

In collaborative filtering systems, users or consumers 

are typically represented by the items they have 

purchased or rated. For instance, in an online cinema 

have 3 million movies; each consumer is represented by a 

Boolean feature vector of 3 million elements. The value 

for each element is determined by whether this consumer 

has viewed the corresponding movie in the past time. 

Typically the value of 1 to 5 indicates that such a view 

had occurred and 0 indicates that no such event has 

occurred. When multiple consumers are concerned, a 

matrix composed of all vectors representing these 

consumers can be used to capture past view events. We 

call this matrix the consumer–product interaction matrix. 

In this article, we use C to denote the set of consumers 

and I to represent the set of items. We represent the 

consumer–product interaction matrix by a |C|×|I| matrix R 

= (rij), such that 

In many large-scale applications, both the number of 

items and the number of consumers are large. In such 

cases, even when many events have been recorded, the 

consumer–product interaction matrix can still be 

(1)


extremely sparse, that is, there are very few elements in R 

whose value is not 0. This problem, commonly referred to 

as the sparsity problem, has a major negative impact on 

the effectiveness of a collaborative filtering approach. 

Because of sparsity, it is highly probable that the 

similarity (or correlation) between two given users is 

zero, rendering collaborative filtering useless [17]. Even 

for pairs of users that are positively correlated, such 

correlation measures may not be reliable. 

The cold-start problem further illustrates the 

importance of addressing the sparsity problem. The coldstart 

problem refers to the situation in which a new user 

or item has just entered the system [18]. Collaborative 

filtering cannot generate useful recommendations for the 

new user because of the lack of sufficient previous ratings 

or purchases. Similarly, when a new item enters the 

system, it is unlikely that collaborative filtering systems 

will recommend it to many users because very few users 

have yet rated or purchased this item. Conceptually, the 

cold-start problem can be viewed as a special instance of 

the sparsity problem, where most elements in certain 

rows or columns of the consumer–product interaction 

matrix A are 0 [2]. 

Many researchers have attempted to alleviate the 

sparsity problem. In [19], the author proposed an itembased 

approach to addressing both the scalability and 

sparsity problems. Another proposed approach, 

dimensionality reduction, aims to reduce the 

dimensionality of the consumer–product interaction 

matrix directly. A simple strategy to reduce the 

dimensionality is to form clusters of items or users and 

then use these clusters as basic units in the prediction. 

More advanced techniques can be applied to achieve 

dimensionality reduction. Examples are statistical 

techniques such as Principle Component Analysis (PCA) 

[10] and information retrieval techniques such as Latent 

Semantic Indexing (LSI). Essentially, dimensionality 

reduction approaches deal with the sparsity problem by 

generating a denser user-item interaction matrix that 

considers only the most relevant users and items. 

Predictions are then made using this reduced matrix. 

Empirical studies indicate that dimensionality reduction 

can improve recommendation quality significantly in 

some applications, but performs poorly in others, the 

potentially useful information might be lost during this 

reduction process [20]. 

Researchers have also attempted to combine 

collaborative filtering with content-based 

recommendation approaches to alleviate the sparsity 

problem [21][22]. In addition to user-item interactions, 

such techniques also consider similarities between items 

derived from their content, which allow them to make 

more accurate predictions. However, the hybrid approach 

requires additional information regarding the products 

and a metric to compute meaningful similarities among 

them. In practice, such product information may be 

difficult or expensive to acquire and a related similarity 

metric may not be readily available. 

Another category of methods consider the data as a 

bipartite graph where nodes represent the users and items, 


and an edge (i, j) exists between a user i and an item j if i 

has rated j. Moreover, edge (i, j) is given a weight 

corresponding to the rating given by i to j. These methods 

then derive global similarities between users or items 

using graph theoretic measures. For instance, one such 

method computes similarities between two users as the 

average commute time between their respective nodes in 

a random-walk of the graph. Other graph theoretic 

measures were also investigated, such as the minimal hop 

distance between nodes of the graph, and the spread 

activation of the nodes in the graph. The main drawback 

of these approaches is that there is often no good 

interpretation of the similarity measures in the context of 

the prediction problem [23]. 

Our research focuses on developing a computational 

approach to exploring transitive between users to address 

the sparsity problem and improving the accurate in the 

context of collaborative filtering. 

III. COLLABORATIVE FILTERING BASED ON ASSOCIATION 

RETRIEVAL 

A. Association retrieval 

Associative retrieval has its origin in statistical studies 

of associations among terms and documents in a text 

collection. The basic idea behind associative retrieval is 

to build a graph or network model of documents and 

index terms and queries, and then to explore the transitive 

associations among terms and documents using this graph 

model to improve the quality of information retrieval. 

This relationship is also reflected in people's daily life, 

for instance, Lisi is Wanwu‘s friend, Zhanshan is Lisi‘s 

friend, Wanwu can recommend movie A to Zhanshan, so 

there is an association relationship between Zhanshan and 

Wanwu. We found that recommender system can make 

use of this relationship between users to address the 

sparsity by studying. 

B. Finding the relationship between users by association 

retrieval 

Firstly, we supposed that represent a 

user set which includes 3 users, 

represents a movie set which includes 4 movies, 

represent a user‘s rating matrix which 

includes elements. 

The rows represent the user, the columns represents the 

movie, for example, the first row represents the user c1 

viewed the movies i2 and i4, the rating is 3 and 4 

respectively. 

From the second line in the matrix B, we can know that 

the user c2 viewed the movie i2, i3 and i4. It is easy to find 

that the user c1 and c2 viewed the movie i2 and i4 from 

matrix R and B. According to similarity theory, we can 

ascertain that the user c1 is similarity with the user c2, so 

the movie i3 can be recommended to the user c1 through


the user c2, but the movie i1 cannot be recommended to c1 

forever. However, the above example only has 4 movies. 

At present, the online movie provider more than millions 

movies, the ―dark information‖ will appear if only 

through the direct similarity users to recommend, some of 

movies will cannot be recommended to some of users, the 

requirements of the user cannot be satisfied. 

According to the association retrieval theory, users as a 

set of nodes, the products as a set of nodes, we use the 

bipartite graph to express the matrix B, as shown in Fig1. 

Customer Nodes 

C1 

C2 

C3 

Product Nodes 

i1 

i2 

i3 

i4 

Figure 1. Transitive associations in collaborative filtering. 

Accordering to Fig 1, the length of the association path 

is assumed to be 3, there are c1-i2-c2-i3 and c1-i4-c2-i3 two 

paths, the movie i3 is recommended to the user c1, but 

there is not a path whose length is 3 between i1 and c1, so 

i1 will not be recommended to the user c1. If the length of 

the path is extended to 5, we can find that the movie i1 

can be recommended to the user c1 through the path c1-i2c2-i3-c3-i1 

and c1-i4-c2-i3-c3-i1. 

Accorder to the above analysis, this paper makes some 

of define are as follows: 

Definition 1: direct recommendation path represent a 

user recommend item to a target user directly. 

Definition 2: indirect recommendation path represent a 

user recommend item to a target user through one or 

more than one user. 

Definition 3: user direct similarity degree represents 

the similarity degree between users in direct 

recommendation path. 

Definition 4: user indirect similarity degree represents 

the similarity degree between recommendation user and 

target user in indirect recommendation path. 

From the above analysis, we know that the association 

retrieval method can explore the transitive between users 

to get a set of paths and the direct or indirect similarity 

degree. Through formula (2) to compute the value of 

in the sparsity matrix to address the sparsity problem. 

(2) 

Note that i represents user, j is item, 

is the set of recommendation path, 

represents an ordered set of a recommendation path the 

user passed, is similarity degree between and . 

C. Computing the direct similarly matrix 

In the computing of the direct similarity matrix, we do 

not use the Pearson-correlation and cosine of the angle. 

Through the research we find that whatever the user 


rating is high or low after the user viewed the movie, to 

some extent, which express some of similarity between 

users both on the personal preferences and the preference 

of ratings. For example, in the matrix R, the user c1 and c2 

rated i2 and i4, the rating value of the c1 is 3 and 4, the 

rating value of the c2 is 2 and 5, we can use formula (3) to 

compute the rating similarity degree between c1 and c2 for 

the same movie and , 

max is the maximum value function; abs is the 

absolute value function; R represents the value set of the 

rating, such as R={0,1,2,3,4,5}; , the value of the user 

i rate product k. Formula (4) was used to compute the 

user similarity between i and j after get the rating 

similarity degree. 

Note that m, the number of the products. We use the 

rating matrix R as an example, the user similarity 

, according to this method, 

we can get the user similarity matrix as follows: 

Next, we combine the association retrieval and direct 

similarity matrix to compute in order to get the 

recommendation matrix after getting the user similarity 

matrix. 

D. Computing the recommender matrix 

We use the data the section 3.2 provided to recommend 

for the user c1. When M=3, we can find that c1 has two 

recommendation path c1-i2-c2-i3 and c1-i4-c2-i3 from the 

data; the similarity between c1 and c2 is 0.4 from the 

similarity matrix in section 3.3, the weight of the path is 

0.4; so we get the correlation degree of the i3 is 

; Because c1 and c2 have the highest 

similarity, the rating value of the c2 for i3 is 3, so the 

recommendation value is . When M=5, 

there are two recommendation path c1-i2-c2-i3-c3-i1 and c1i4-c2-i3-c3-i1, 

the weight is 

, the value of the correlation degree is , 

the rating value of the c3 for i1 is 4, so the 

recommendation value is . 

The recommendation matrix was defined in 

(5) 

(3) 

(4) 

(5) 

Note that R, the rating matrix, is the similarity 

matrix, B is the marked matrix. Using the data in section 

3.2, we get the recommendation matrix and 

through formula (4) when M=3 and M=5.


From the above recommendation matrix, we can know 

that , 

, which consistent with the above computing 

outcome. 

IV. ALGORITHM 

The algorithm is as follows: 

Algorithm1. Collaborative algorithm based on 

association retrieval 

Input:user rating matrix R,the length of path M 

Output:Recommendation matrix 

Step1. Matrix B = Matrix R, If not equal 0 then 

for each . 

Step2. Set the iteration variable N=1. 

Step3. Original recommendation matrix . 

Step4. Compute the direct similarity matrix 

according to formula (3) and (4). 

Step5. Compute the transpose . 

Step6. Compute the matrix according to 

formula (5). 

Step7. If N+2 less than M then N=N+2, goto Step 3 until 

N larger than M. 

A. Experiment data 

V. EXPERIMENT AND ANALYSIS 

The datasets were collected by the GroupLens 

Research Project at the University of Minnesota. 

The data set consists of 100,000 ratings (1-5) from 943 

users on 1682 movies. Each user has rated at least 20 

movies; the sparsity degree is 99.937%. 

R 

C. Experiment results 

In our experiment, we called our approach ARC. We 

respectively compute the precision, recall and F-measure 

based on Movielens data set for the ARC, PC and COS 

algorithms. In the ARC, the value of the M is 3. 

Summarized bar charts are shown in Figs. 2–5. Table1 is 

the comprehensive comparison about the precision, recall 

F-measure and coverage between ARC, PC and COS 

algorithms. 

In the aspect of the precision, the ARC increased by 

18.40% compared with PC and 33.58% compared with 

COS. In the aspect of the recall, the ARC increased by 

17.65% compared with PC and 66.68% compared with 

COS. In the aspect of the F-measure, the ARC increased 

by 18.39% compared with PC and 34.13% compared 

with COS. In the aspect of coverage, the ARC increased 


B. Experiment procedure 

For each target consumer, we retrieved the entire set of 

previously viewed items and sorted them into 

chronological order by view date. The first 90% of these 

items was treated as ―past‖ views to serve as input to be 

fed into different methods to generate recommendations. 

For comparison purposes, the second 10% of these items 

were treated as ―future‖ views of the customer and hidden 

from the recommender system. 

In the experiment, we compared the outcome of the 

Pearson-correlation, Vector similarly, Item-based and our 

approach. We use precision, recall, coverage and Fmeasure 

to measure the effectiveness of a given 

recommendation approach. These measures are widely 

accepted in information retrieval and recommender 

system research [24]. 

The baseline methods are described below. 

Pearson Correlation Coefficient (PCC) 

Pearson Correlation Coefficient method predicts the 

rating of a test user x on item i as: 

(5) 

Where the coefficient is computed as 

Vector Similarity (VS) 

This method is very similar to the previous method 

except that the correlation coefficient is 

computed as: 

The definition of the precision, recall, coverage and Fmeasure 

are as follows. 

(8) 

(9) 

(10) 

(11) 

by 4.66% compared with PC and 24.78% compared with 

COS. From the results, we can see that there are greatly 

improved in the aspect of the precision, recall, F-measure 

and coverage. But from the above data, we find that the 

COS is worst in the situation of the sparsity. Otherwise, 

in the aspect of coverage, the ARC increased by only 

4.66% compared with PC. We also make another 

experiment, the results show that the coverage can 

increase more than 10% when the M equals 5, the 

overhead of the computing have great increased, but the 

increase was very little in the recommendation precision. 

This paper considers that a low coverage rate increase for 

two reasons, on the one hand, it is because the value of 

the M is 3; on the other hand, maybe the sparse degree of 

the experiment data set is not enough. 

(6) 

(7)


ASS 

Value of the precision 

Value of the recall 

0.018 

0.016 

0.014 

0.012 

0.01 

0.008 

0.006 

0.004 

0.002 

0 

1 

0.9 

0.8 

0.7 

0.6 

0.5 

0.4 

0.3 

0.2 

0.1 

0 

TABLE1. 

COMPREHENSIVE COMPARISON TABLE 

PC COS 

Precision Recall F-measure Coverage Precision Recall F-measure Coverage 

D-value 0.00256 0.1429 0.0503 0.0378 0.00414 0.381 0.0824 0.1685 

The percent of the 

improving 

0.01647 

18.40% 17.65% 18.39% 4.66% 33.58% 66.68% 34.13% 24.78% 

Precision 

0.01391 

Figure 2. The comparison of the predictive precision 

Figure 3. The comparison of the recall 

VI. CONCLUSION 

0.01233 

ARC PC COS 

Name of the method 

0.9524 

0.8095 

0.5714 

ARC PC COS 


In this paper, we aimed to alleviate the sparsity 

problem and improve the recommendation precision in 

collaborative filtering systems. We use the association 

retrieval technology to alleviate the sparsity problem and 

proposed a new collaborative filtering algorithm to 

increase the recommendation precision. The effectiveness 

of the approach was evaluated experimentally using data 

from the movielens data set. The experiment indicated 

that our approach alleviated the sparsity problem and 

achieved significantly better recommendation quality 

than the standard collaborative filtering approaches 

Meanwhile, there is a great problem for the proposed 

approach in this paper. The volume of data these s ystems 

utilize will continue increasing over time. In this 

situation, our approach will cause the data overload 

problem. As a result, it will present a significant 


Recall 

0.4 

0.35 

0.3 

0.25 

0.2 

0.15 

0.1 

0.05 

0 

Value of the F-measure 

Value of the coverage 

0.9 

0.8 

0.7 

0.6 

0.5 

0.4 

0.3 

0.2 

0.1 

0 

0.3238 

F-measure 

0.2735 

Figure 4. The comparison of the F-measure 

Figure 5. The comparison of the coverage 

challenge for the scalability of collaborative filtering 

recommenders. So, the next research, we will consider 

the scalability problem of collaborative filtering 

recommenders. 


This work was supported in the National Natural Science 

Foundation of China under Grant No. 60672051. The 

Fundamental Research Funds for the Central Universities 

(3105005). Wuhan science and technology plan projects 

(201010621209) 

REFERENCES 

0.2414 

ARC PC COS 


0.8484 

Coverage 

0.8106 

0.6799 

ARC 

Name of the 

PC 

method 

COS 

[1] Liu Jianguo, Zhou Tao, et al. Progress of the personalized 

recommendation systems. Progress of Nature and Science, 

200919(1):1-15 

[2] Zan Huang, Hsinchun Chen, et al. Applying Associative 

Retrieval Techniques to Alleviate the Sparsity Problem in


Collaborative Filtering. ACM Transactions on Information 

Systems, Vol. 22, No. 1, January 2004, 116–142. 

[3] Liu Jianguo, Zhou Tao, et al. Overview of the Evaluated 

Algorithms for the Personal Recommendation Systems. 

Complex System and Complexity Science. 2009, Vol.6, 

No.3, 1-10. 

[4] Rong Jin, Luo Si, et al. Collaborative Filtering with 

Decoupled Models for Preferences and Ratings. CIKM ’03, 

New Orleans, Louisiana, USA, November 3-8, 2003. 

[5] Rich E. User modeling via stereotypes. Cognitive Science, 

1979, 3(4):329-354. 

[6] Goldberg D, Nichols D, et al. Using collaborative filtering 

to weave an information tapestry. Comm ACM, 1992, 

35(12):61-70. 

[7] Konstan JA, Miller BN, et al. GroupLens: Applying 

collaborative filtering to usenet news. Comm ACM, 1997, 

40(3):77-87 

[8] Shardanand U, Maes P. Social information filtering: 

Algorithms for automating ‗Word of Mouth‘. Proc Conf 

Human Factors in Computing Systems Denver, 1995, 210- 

217. 

[9] Linden G, Smith B, et al. Amazon.com recommendations: 

Item-to-item collaborative filtering. IEEE Internet 

Computing. 2003, 7(1):76-80. 

[10] Goldberg K, Roeder T, et al. Eigentaste: A constant time 

collaborative filtering algorithm. Information Retrieval. 

2001, 4(2):133-151. 

[11] Terveen L, Hill W, et al. PHOAKS: A system for sharing 

recommendations. Comm ACM, 1997, 40(3):59-62. 

[12] J. S. Breese, D. Heckerman, et al. Empirical Analysis of 

Predictive Algorithms for Collaborative Filtering, 

Proceeding of the Fourteenth Conference on Uncertainty 

in Artificial Intelligence (UAI). 1998. 

[13] Chen YL, Cheng LC. A novel collaborative filtering 

approach for recommending ranked items. Expert Systems 

with Applications, 2008, 34(4):2396-2405. 

[14] Yang MH, Gu ZM. Personalized recommendation based 

on partial similarity of interests. Advanced Data Mining 

and Applications Proceedings, 2006, 4093:509-516. 

[15] Breese JS, Heckerman D, et al. Empirical analysis of 

predictive algorithms for collaborative filtering. Proc 14th 

Conf Uncertainty in Artificial Intelligence Madison, 1998, 

43-52 

[16] Getoor L, Sahami M. Using probabilistic relational models 

for collaborative filtering. Proc Workshop Web Usage 

Analysis and User Profiling, San Diego. 1999. 

[17] Billsus, D., Pazzani, M. J. Learning collaborative 

information filters. In Proceedings of the 15th 

International Conference on Machine Learning, 1998, 46– 

54. 

[18] Schein, A. I., Popescul, A., et al. Methods and metrics for 

coldstart recommendations. In Proceedings of the 25th 

Annual International ACM SIGIR Conference on Research 

and Development in Information Retrieval (SIGIR 2002). 

(Tampere, Finland), 2002, 253–260. 

[19] Sarwar, B., Karypis, G., et al. Item-based collaborative 

filtering recommendation algorithms. In Proceedings of the 

10th International World Wide Web Conference. 2001, 

285–295. 

[20] Sarwar, B., Karypis, G., et al. Application of 

dimensionality reduction in recommender systems: A case 

study. In Proceedings of the WebKDD Workshop at the 

ACM SIGKKD. ACM, New York.2000. 

[21] Good, N., Schafer, J., et al. Combining collaborative 

filtering with personal agents for better recommendations. 

In Proceedings of the 16th National Conference on 

Artificial Intelligence, 1999, 439–446. 


[22] Huang, Z., Chung, W., et al. A graph-based recommender 

system for digital library. In Proceedings of the 2nd 

ACM/IEEE-CS Joint Conference on Digital Libraries 

(Portland, Ore.). ACM, New York, 2002, 65–73. 

[23] Chrsistian Desrosiers, George Karypis. Solving the 

Sparsity Problem: Collaborative Filtering via Indirect 

Similarities. Technical Report. Department of Computer 

Science and Engineering University of Minnesota 4-192 

EECS Building 200 Union Street SE Minneapolis, MN 

55455-0159 USA. 2008. 

[24] Sarwar, B., Karypis, G., et al. Analysis of recommendation 

algorithms for e-commerce. In Proceedings of the ACM 

Conference on Electronic Commerce. ACM, New York, 

2000, 158–167. 

Yibo Chen, born in 1982, Ph.D. candidate. The research 

interests include personalization recommendation and semantic 

web. 

Chanle Wu, born in 1945, professor, The interests include 

computer networks, e-Learning, grid computing and semantic 

web. 

Ming Xie, born in 1978, Ph.D. candidate. The research interests 

include data mining and semantic web. 

Xiaojun Guo, born in 1984, Ph.D. candidate. The research 

interests include semantic web and e-Learning.


Integrated Structure and Control Design 

for Servo System Based on Genetic 

Algorithm and Matlab 

Dingzhen Li 

Department of Electronics and Electrical Engineering, Nanyang Institute of Technology, Henan 473004, China 

Email: lidingzhen.student@sina.com 

Ruimin Jin 

Department of Electronics and Electrical Engineering, Nanyang Institute of Technology, Henan 473004, China 

Email: jinruimin2004@163.com 

Abstract—The integrated design was introduced about the 

pitching part of airborne radar servo system. The paper 

analyzed both the servo system model and the dynamic 

character of this system. Then the research about parameter 

optimization and simulation had been done by using the 

Matlab. The electromechanical coupling model and the 

optimization model were built up based on model of 

mechanism transmission system and electricity control 

system. The optimization model includes the integrated 

design of structure parameters and control parameters. The 

dynamics model of mechanism transmission system includes 

the nonlinearity of backlash. It considered the influence of 

parameters for dynamics property in structure of the 

mechanism transmission system. Furthermore, the model of 

electricity control system of the airborne radar servo system 

has three feedback-control loops. The method of integrated 

structure and control design was applied on the 

optimization model using Genetic Algorithm (GA). 

Simulation had been done based on Matlab/Simulik. 

Simulation results showed that the method of integrated 

structure and control design is very good. It is feasible and 

effective for airborne radar servo system. It proved the 

method used in the task is right and the practicability of 

Genetic Algorithm. 

Index Terms—Radar Servo System, Integrated Design, 

Electromechanical Coupling Model, Genetic Algorithm, 

Matlab Simulation 


Modern mechanical and electrical systems have higher 

demanding for the system accuracy and steady-state 

dynamic performance in a variety of extreme conditions. 

The traditional design method is that the mechanical part 

is designed first, then the control part is designed. This 

method ignores the dynamics of the mechanical system, 

the interactions and the mutual coupling of the control 

system. The repeated design of the structural parameters 

and control parameters also causes the design cycle 


doi:10.4304/jcp.6.9.1903-1912 

getting long and the cost increasing. It is also difficult to 

achieve the best performance of the electromechanical 

systems. 

In order to improve the overall performance of the 

system, we should unify the model to the mechanical and 

electrical systems and have an integrated design on the 

base of this. Namely, the integrated design of structure 

parameters and control parameters will be done. 

This paper studied the airborne radar pitch servo 

system. The first unit is a single stage gear meshing. We 

research the dynamics model of the pitch 

servo-mechanical drive system. The influence of the 

system dynamic characteristics is analyzed on the 

structural parameters of the mechanical transmission 

system. Secondly, the design of the loop circurt of servo 

system is introduced. Then we establish a servo-electric 

control system of this model and study the simulation 

model. Finally, the servo control system is coupled with 

mechanical system and electrical system. The mechanical 

structure is established including the structure parameters 

and control parameters of the electromechanical coupling 

model. Then we create an integrated optimization model 

for the structure/control integrated design. 

Airborne radar servo system includes machine system 

and electrical control system [1]. Traditional design 

methods all make the structure and control of radar servo 

system as a separate module, later the structure design is 

optimized, and then is added the optimal controller to the 

structure. It is sequential, serial. All parts are independent 

design, which ignore the strong coupling between the 

structural parameters and control parameters. It is very 

difficult to achieve the global optimum. To overcome this 

shortcoming, the design uses a method that makes the 

structure and control integrated. During the process of the 

design, we synchronize the optimization of structure 

parameters and control parameters. The simulation results 

show that the method can improve the synthetic 

performance of radar servo system effectively.


II. INTRODUCTION OF THE OPTIMIZATION 

ALGORITHM 

The model of radar pitch servo system considers the 

influence of backlash nonlinear factors. Because of the 

multiple constraints of structure and control, the 

optimization design is more difficult, so that the 

optimization problem is finally reduced to the solution of 

a nonlinear function which is constrained. Using 

quasi-Newton method of nonlinear optimization 

algorithm is not only large, but also falling into local 

optimal solution frequently, resulting in the failure of 

optimization. The genetic algorithm is based on the 

fitness function, by manipulating implementation for the 

genetic of all individuals of the population to achieve 

restructuring within the group iterative process of 

individual search method. And the search does not 

depend on gradient information, especially for dealing 

with complex problems and nonlinear problems that 

traditional search methods can not solve [2], and get 

optimal solution of the global system, so it has been 

widely used in integrated design of the structure and 

control. Therefore, the design uses a genetic algorithm 

theory. 

Genetic Algorithm(GA) is random optimization search 

method by simulating the natural selection and heredity 

mechanism in the natural biology evolvement.In the 

engineering applications,there are many problems in 

dealing with the multi-parameter and multi-objective, 

such as optimizing the parameter of the servo system 

adjustor. It can be worked by using the GA which is used 

to design the program by different requirements of the 

system design. With the Matlab/Simulink, the best result 

can be searched in whole space and be given at last. The 

paper used Genetic Algorithm to optimize the parameter 

of P and PI, and received fine impression. It proved the 

method used in the task is right and the practicability of 

genetic algorithm. 

x , endPop , bPop , traceInfo ] = ga( 

bounds , 

selectFN 

x is the optimal solution obtained. endPop is the final 

population we get. bPop is the search trajectory of the 

optimal population. traceInfo is the information 

optimized. bounds is the matrix which represents the 

upper and the lower bounds of the variable input 

parameters. evalFN is the fitness function whose format 

is: function [val, sol] = evalFN (sol, options), of which 

val indicate fitness defined in the fitness function, sol is 

the design bariables in the process of the optimication. 

Namely the genetic algorithm is individual. startPop is 

the initial population function, which is used to initialize 

the genetic algorithm. This is the format: startPop = in 

itializega (PopulationSize, bounds, evalFN, evalOps, 

options ), in which, populationSize is used to specify the 

size of the initial individuals of each generation, bounds, 

evalFN is the same as the previous definition. For other 

parameters, you may see the help. 

With these prepared knowledge above, you can use the 

GAOT toolbox directly to design the system of this paper 

in the following integrated design. 

Genetic algorithm (GA) is a searching for optimization 

algorithm that is based on the principle of natural 

selection and genetic mechanism. The major steps 

include coding, initial generation of population, 

adaptation detection and evaluation, selection, crossover 

and mutation. The flow chart is shown in Fig.1. Optimal 

value is finally output. 

Figure 1. GA flow chart 

There are three toolboxes of genetic algorithm, say 

GAOT, GATBX and GADS. GAOT is a free toolbox 

circulated on the Internet. It is not the software that 

comes from MATLAB. But it can be easily configured to 

use. The default toolbox is the objective function for 

solving the maximum, while the structure/control 

integrated design goal is to solve the objective function of 

the maximum. So it is optimized GAOT toolbox and 

more convenient. 

GAOT toolbox includes many useful functions [3]. 

The main program provides a genetic algorithm toolbox 

and the external interface. Its function format is as 

follows: 

evalFN , evalOps , startPop , opts, 

termFN , , 

, selectOps, 

xOverFNs, 

xOverOps, 

mutFNs, 

mutOps) 

[ termOps 


III. ESTABLISH OF ELECTROMECHANICAL 

COUPLING MODEL 

In order to do the structure/control integration design, 

we must establish a servo system electromechanical 

coupling model which combines the mechanical system 

and control system [4]. In order to show the coupling 

model clearly, we can use Matlab/Simulik technology- 

related to group the function-related modules as 

subsystems and establish multi-level hierarchical model. 

Fig. 2 is the pitch servo system electromechanical 

coupling model established. The internal structure of each 

module is shown in Fig. 2(a) ~ Fig. 2(e) below. We will 

introduce them in detail. 

In the basic design of the parameters of PWM 

controller, time constant and the loop filter constants are 

very small, which has very little effect on the results of 

the simulation system, so we ignore these factors in the 

coupled model [5].


In the three-ring control structure, current loop and 

speed loop are inner ring, while position loop is outer 

ring. This structure can get better dynamic following 

performance and anti-jamming performance. Among 

them, the function of current loop is to change the 

transfer function and improve speed of the system. It 

timely inhibits internal interference of current loop and 

limits maximum current. It makes the system with 

enough accelerate torque and ensures safety operation of 

the system. The role of the speed loop is to enhance the 

ability of system disturbances and inhibit speed 

fluctuation. Position loop is to guarantee system static 

precision and dynamic tracking performance, making the 

whole servo system stabilize, with high-performance 

operation. Three-ring controller fits quality or not 

directly relates to the servo drive system stability, 

accuracy and quickness. For control system has 

multi-ring structures, its controller parameters were set 

as follows: we first design controller parameters of the 

inner loop, then design the outer controller parameters 

regarding inner loop as a link, and ultimately design 

parameters of all the control loops in this manner. 

A. Electrical Machine Model 

Fig. 2 (a) is the module of electrical machine model. 

In the armature current consecutive cases, the armature 

voltage balance equation will be 

di di 

u− E = iR+ L = R( i+ Ti 

) . (1) 

dt dt 

In equation(1), u is the voltage added to the two 

terminals of the motor. E is back emf. i is armature 

current. R is total resistance. L is all the armature loop 


Figure 2. Diagram of the electromechanical coupling model 

Figure 2 (a). Module of the electrical machine model 

inductance. Ti L R = is called electromagnetic time 

constant in the armature loop. Then there is 

E = K θ . (2) 

e m 

In equation(2), Ke is called back emf coefficient. 

θm is the motor shaft corner. 

Concerning motor, its moment balance equation of 

rotation axis are as follows 

⎧ ⎪Tm 

= I 

mθm + Bmθ m + M l 

⎨ 

. (3) 

⎪⎩ Tm = K ti 

In equation (3), T representes the output torque. 

m 

K is the moment coefficient. t 

l M is the motor 

disturbance moment. I is inertia moment of motor 

m 

armature. B is damper of motor rotor. 

m 

Take laplace transformation to the above equation (1) 

~ equation (3) at the same time. Then we get equation 

(4). 

⎧U 

( s) − E( s) = R( I( s) + TiI( s) s) 

⎪ 

⎨ E( s) = K eθ 

( s) 

⎪ 

I( s) K t = M l + Imθ ⎩ 

( s) s 

. (4) 

Simulation square diagram of motor mathematical 

model by equation (4) can be shown in Fig. 2(a). 

Each symbol in the diagram have been definited. 

Related parameters with pitch servo system drive motor 

in this paper can be seen in table .


TABLE Ⅰ.BASIC PARAMETERS OF PITCH SERVO MACHINE 

Rated current Ic 5.42 amps 

Peak current Ip 25.3 amps 

Resistance R 1.81 ohms 

Induction L 5.1 mH 

Torque constant Kt 0.462 N . m/amp 

Constant of emf Ke 48.4 V/krpm 

Inertia moment Im 1.96E-04 kg . m 2 

Damper Bm 4.653E-04 N . m . s/rad 

B. Design of the Current Loop 

In practical engineering applications, we often add the 

current loop circuit in the speed loop circuit to ensure 

rapid start-up performance of the system. PI controller 

makes the current loop have the steady-state tracking 

performance of the step signal without static error. It can 

also effectively reduce the time constant of the motor 

circuit. This provides the design basis of the speed loop 

controller with rapid response. 

Fig. 2 (b) is the module of electrical current loop. In 

the current loop diagram, transfer function of the 

Pulse-Width Modulator is following 

K pwm 

Gpwm 

= 

T s+ 

1 

. (5) 

pwm 

Figure 2 (b). Module of the electrical current loop 

In equation (5), Kpwm is the voltage amplification 

factor of PWM controller. TPWM is the time constant of 

the PWM controller. 

Transfer function of PI regulator in current loop is 

following 

Tao _i× s + 1 

GIT = K _ i× 

. (6) 

Tao _ i × s 

In equation (6), Tao-i is integral time constant of 

current loop. K_i is the scale factor of current loop. K_A 

is the current amplification factor of feedback loop. 

C. Design of the Speed Loop 

Fig. 2(c) is the module of speed loop. In practical 

system debugging, the speed loop has good or bad effect 

on system performance considerably. On the one hand, 

increasing the speed loop gain can increase the stiffness 

of the speed loop, reduce the sensitivity of the system for 

dynamic and static friction, overcome the dead zone, and 

reduce the torque fluctuations. On the other hand, it can 

effectively expand the system bandwidth and prevent the 

mechanical resonance of turntable. During the debugging, 

we found that we wanted to realize the system load and 


other disturbances on the robustness, the speed must be 

well designed to ensure three-ring steady. This paper 

used the speed loop with PI controller still. In the system 

block diagram of the speed loop, transfer function of the 

speed loop with PI controller is 

T _ v× s+ 

1 

GST = K _ v× 

T _ v× s 

. (7) 

In equation (7), T_v is integral time constant of the 

speed loop. K_v is proportional coefficient of the speed 

loop. Transfer function of the speed loop filter is 

GSL 

1 

= 

. (8) 

K _ b× s+ 

1 

In equation (8), K_b is the filter constant of the speed 

loop. In addition, K_B is the current amplification factor 

of the feedback loop. 

Figure 2 (c). Module of the speed loop 

D. Design of the Position Loop 

Fig. 2(d) is the module of the position loop. Upon 

completion of the current loop and speed loop, the 

system block diagram is established. The position loop 

can be designed as closed-loop. Finally, the performance 

indicators of the system need be achieved in the last 

position loop. Therefore, the position loop is a very 

important part of servo control system in the whole 

design. This position loop controller uses PI controller. 

Transfer function of the position loop with PI controller 

is 

Ti × s + 1 

GPT = Kp× 

. (9) 

Ti × s 

In equation (9), Ti is integral time constant and Kp is a 

proportion coefficient in the position loop. Transfer 

function of the position loop filter is 

GPL 

1 

= 

. (10) 

K _c× s+ 

1 

In equation (10), K_c is filter constant of the position 

loop. 

Figure 2 (d). Module of the position loop 

Fig. 2(e) is the module of the load model. Load 

module is very complex. We did not introduce in detail.


In the end, we established the three closed-loop 

simulation model of control system and got its 

simulation. The corresponding initial parameters were 

Jeq = 0.018, Beq 

= 0.0004653, 

K pwm = 4.6, Tpwm= 5e−5, K _ A= 0.3333, K _ a = 5e−5, K _ B = 0.5, K _ b = 0.0002, 

K _ c=0.002. 

Using the original parameters, we debuged the control 

system with PI controller repeatly. Then we adjusted the 

controller parameters in the ring. As follows: 


Figure 2 (e). Module of the load model 

K _ i= 5, Tao_i= 0.0028, K _ v= 47.876, T _ v= 

0.32265, 

Kp = 59.178, Ti = 789.62. 

For example, the position loop simulation curve was 

shown in Fig. 3. 

The system established under ideal conditions is stable 

in Fig. 3(a) and Fig. 3(b). At the same time, the system 

transient response of step signal is essential to meet the 

performance requirements of the radar servo system. We 

will adjust the controller parameters and system tricyclic 

initial design parameters of the other. These parameters 

are the choice of parameter ranges of the structure and 

control integration design.


Position / rad 

1.4 

1.2 

1 

0.8 

0.6 

0.4 

0.2 

Reference Signal and Response Signal 

Reference Signal 

Response Signal 

0 

0 1 2 3 4 5 6 7 8 9 10 

10 

Time(s) 

-2 

10 0 

10 2 

-720 

Frequency (rad/sec) 

(a) Response curve of step signal (b) bode diagram 

IV. ESTABLISHMENT OF OPTIMIZATION MODEL 

Structure/control integrated design methods can 

overcome the internal contradictions and the waste of 

source brought about by the independence of structure 

and control. It can also promote each other by 

coordinating the relationship between them, so that we 

can achieve a optimum control efficiency. So, it is of 

great importance. Here we take an example of the stage 

signal to do the integrated design of the radar pitch servo 

system. 

A. Determination of the Objective Function 

Since we mainly study the influence on the effect of 

the servo system brought about by nonlinear factors of 

the gear gap, we take gear gap as the objective function 

in the optimization model. Under the premise of meeting 

the demand of the system functions, structure/control 

integrated design optimized the maximum gear gap value 

that the system allowed [6]. 

There are two levels gear gap in the designed servo 

system model: high speed gear gap and low speed gear 

gap [7]. To take gear gap as the objective function, we 

should consider the weight coefficient of the two gear 

gaps. So the objective function is taken as in 

F = λ b + λ b . (11) 

1 1 2 2 

The impact of the low-level backlash is larger than 

that of the high level in the second gear transmission 

mechanism. So the weight factor is taken as in 

λ = 0.2, λ = 0.8 . (12) 

1 2 

B. Design Variables and Constraint Conditions 

1) Selection of design variables 

The selection of design variables on structure:the teeth 

number of each gear is Z1, Z2, Z3 and Z4 respectively. 

Semi-backlash are b1 and b2. We take tricyclic PI 


Magnitude (dB) 

Phase (deg) 

100 

0 

-100 

-200 

-300 

-400 

0 

-180 

-360 

-540 

Figure 3. Location-loop simulation curve 

Bode Diagram 

controller parameters K_i, Tao_i, K_v, T_v, Kp and the 

corresponding drive parameter K_A, PW, K_B as the 

design variables in controlling. 

2) Specific constraints of the needed design variables 

After determining the variables of the integrated 

design, we introduce the specific constraints for the 

variables: 

a) Specific constraints of variables 

The mechanical drive mode of the airborne radar pitch 

servo system is two-level gear transmission mode. This 

is shown in Fig. 4. 

10 4 

Figure 4. Two-level gear transmission mode 

The system has very strict limitations of space, so 

space constraints of the model is an very important 

constraint. We assumed that the two-level gear 

transmission system is limited in a space of an outside 

diameter W. The constraints are 

⎧2( 

r1 + r2) ≤ W 

⎨ 

. (13) 

⎩r2 

+ r3 + 2r4 

≤ W 

b) Constraints of reduction ratio 

The constraints of the servo system reduction ratio 

take the given initial value as reference to define its 

boundary of reduction ratio. The constraint conditions of 

the servo system reduction ratio take the given initial 

value as reference to define its boundary, namely 

⎧ a 

⎪ 

⎨b 

⎪ 

⎩c 

− 

− 

− 

≤ i 

≤ i 

≤ i 

12 

34 

总 

= r 

= r 

= i 

2 

4 

12 

/ r ≤ a 

/ r 

i 

34 

1 

3 

≤ b 

≤ c 

+ 

+ 

+ 

10 6 

. (14)


The boundary of the high-level gear transmission ratio 

a , . 

and c , c refer to the boundary of two-gear 

and the low-level gear transmission ratio are − a+ 

b− , b+ 

− + 

transmission ratio respectively. 

c) Performance constraints 

The steady-state error of e in the system should satisfy 

the system precision. That is e


We can see the results from the table and figures. 

Mechanical structure parameters and control parameters 

of the servo system are matched with each other after the 

system is integrated design [9]. The gear of the meshing 

has changed at the premise of meeting the performance 

index of the system. And the transmission ratio is 

re-assigned. The space the gear box used was reduced by 

2.46% after the integrated design, meeting requirements 

Fittness 

E rror/rad 

-100 

0 10 20 30 40 50 

Generation 

60 70 80 90 100 

for the constraints on space of the airborne radar servo 

system. In addition, the high-speed level and the 

low-speed level backlash in this paper had an increase by 

22.8% and 76.7%. The system can contain a bigger 

backlash after the integrated design, so that the life of the 

gear is prolonged, the waste of resource and the cost of 

production are reduced. As a result, the combination 

property of the radar servo system is finally promoted. 

Position / rad 

1.4 

1.2 

1 

0.8 

0.6 

0.4 

0.2 

Reference Signal and Response Signal 

Reference Signal 

Response Signal 

0 

0 0.5 1 1.5 2 2.5 

Time(s) 

3 3.5 4 4.5 5 

Figure 6.The best solution and the average solution of each generation Figure 7 (a). Optimization response curve 

1.2 

1 

0.8 

0.6 

0.4 

0.2 

20 

-20 

-40 

-60 

-80 

0 

0 

Fittness and Generation 

Tracking Error 

the best solution 

the average solution 

-0.2 

0 0.5 1 1.5 2 2.5 

time(s) 

3 3.5 4 4.5 5 

Figure 7(b). Optimization tracking error curve Figure 7(c). Optimization bode diagram 

Through the analysis and simulation above, we can 

make the conclusions: 

a) In order to build the model, this model can reflect 

the actual system. When establishing the dynamic model 

of pitch servo-mechanical drive system in this paper, the 

nonlinear factors is considered into the backlash and 

backlash model uses a non-linear dead zone model. It is 

in line with this system characteristics. 

b) For multi-gear transmission characteristics of the 

servo-mechanical systems, we built the basic unit of gear 

meshing. Thereby, a mechanical drive system dynamic 

model was established. The basic unit used this method 


Magnitude (dB) 

Phase (deg) 

200 

0 

-200 

-400 

-600 

-800 

0 

-360 

-720 

-1080 

-1440 

10 -2 

10 0 

Bode Diagram 

10 2 

10 4 

Frequency (rad/sec) 

could easily create multi-stage gear transmission system 

model. 

c) The former research results in most of the studies 

only considered the motor control. It nearly regarded the 

matching mechanical and electrical parameters and 

institutional dynamics parameters on the dynamic 

performance of servo system. Pitching in the closed loop 

servo system of radar, the mechanical transmission 

system was included in the position loop. The electrical 

servo control system was not completely separated with 

the subsystems. However a new integrated mechanical 

and electrical coupling systems was formed by a 

10 6 

10 8


feedback loop. Aimed at the feature of this system, the 

mechanical structure and control parameters was 

considered in the structure/control integrated design 

Design 

variable 

Performance parameter 


We built the electromechanical coupling model of the 

airborne radar pitch servo system in this paper. The 

coupled model was the optimization model of the 

structure/control integration design. We used the Genetic 

Algorithm to optimize the parameter of the integrated 

optimization model and received fine impression. It 

proved the method used in the task is right and the 

practicability of Genetic Algorithm. This method can be 

used as a new method in servo system design and can be 

developed in the long ran. 

The model used classical three close loop control 

method. The simulation model is built up by dynamic 

simulation tool Matlab/Simulink and the simulation 

curve lines of reflecting the system performances are 

acquired. According to the simulation model,the effect 

of nonlinear factors on system performances is analyzed, 

and the measures of improving the system performances 

are given. The simulation results show the feasibility and 

effectiveness of the structure/control integration design 

in the servo system, and it is the stage for further 

research. 

Structure and control integration design is not the 

design of simple superposition between mechanical 

module and control module[10]. It analyzes deeply the 

coupling of structure and control and sets up the coupled 

model. Integrated concurrent design is done for the 

structure and control parameters. Modern structure with 

many complexities in itself has a strong dynamic 

coupling, coupled with the control role in the regulation. 

The overall performance of the system can be achieved 

under the combined effects of structure and control. 

Therefore, the establishment of the modern electrical and 


Structure 

Design 

variable 

TABLE Ⅱ. OPTIMIZATION RESULTS 

during the system being designed. The simulation proved 

that the integrated design method was feasible to design 

the radar servo system. 

Parameters Original design value Integrated Design value 

Control 

design variable 

Z1 18 19 

Z2 54 68 

Z3 18 18 

Z4 150 129 

b1 62.5 76.778 

b2 75 132.59 

K_i 5 14.682 

Tao_i 0.0028 0.36888 

K_v 47.876 21.399 

T_v 0.32265 0.08367 

Kp 59.178 58.585 

Ti 789.62 366.47 

K_A 0.3333 0.35943 

PW 4.6 2.3192 

K_B 0.5 0.23749 

W 0.2558 0.2495 

Mp 5% 0% 

tr 0.4 0.21 

ts 1 0.8 

tp 0.8 0.7 

e 1.5e-3 2.8e-4 

mechanical system coupling model can better reflect the 

actual situation. It will be the focus of the future 

research.In the future, research and analysis should be 

done further deeply in the structure/control integration 

design for radar servo system from the following: 

a) During the process of establishing this model, we 

ignore the support bearings and box, and other gear 

stiffness and damping and surface friction and other 

factors. On establishing the future model, we should take 

full account of these factors. 

b) This radar servo system is a three-axis system. It 

contains orientation axis, pitch axis and the 

horizontal-roller axis. It only studies the coupling of the 

mechanical part and the control part in the pitch servo 

system. It doesn’t consider the coupling of the three-axis. 

The future research should be taken into account the 

coupling factor of the three-axis movement. 

REFERENCES 

[1] Job van Amerongen,Peter Breedveld. Modelling of Physical 

Systems for the Design and Control of Mechatronic 

Systems. Annual Reviews in Control.2003, (27):87-117. 

[2] Liu D K,Yang Y L,Li Q S.Optimum positioning of 

actuators in tall buildings using genetic algorithm[J]. 

Computers and Structures,2003,81: 2823-2827. 

[3] YU Ling, JIA Chun-qiang. Functions and Examples in 

Matlab GA Toolbox. Mechanical Engineer.2004,(11): 

27-28. 

[4] Lu Jianwei,Zhang Xianmin, Shen Yunwen. Integrated 

Structral and Noise Control Design For EIC Linkage 

Mechanism. Journal of Mechanical Engineering. 

2003,39(3):40-43. 

[5] Giacomini D,Bianconi E,Martino L,Palma M.A now fully 

integrated power module for three phase servo motor 

driver applications[J].IEEE Industry Applications Society,


2001,2:981-987. 

[6] IJA RM. FONSECA, PETERM. BAINUM. Integrated 

Structure and Control Optimization. Journal of Vibration 

and Control.2004,(10):1377-1391. 

[7] Wang G J,Fong C T,Chang K J.Neural-network-based 

self-tuning PI controller for precise motion control of 

PMAC motor[J].IEEE Transactions on Industrial 

Electronics,2001,48(2):408-415 

[8] ZHAO Guo-feng,FAN Wei-hua,CHEN Qing-wei. A 

Survey on Backlash Nonlinearity. Acta Armamentarii. 

2006,27(6):1072-1080. 

[9] Long Kai, Cheng Ying. The Research of Parameters by 

The Simulation of Exciting Force in Gears. Computer 

Simulation. 2002,19(6):87-89. 

[10] Ahmad Al-shyyab. Non-Linear Dynamic Analysis of a 

Multi-Mesh Gear Train Using Multi-Term Harmonic 

Balance Method.The University of Toledo.2003. 

[11]Xin M,Balakrishnan S N,Ohlmeyer E J.Integrated guidance 

and control of missiles with method[J].IEEE Tmns on 

Control Systerns Technology.2006,14(6):981-992. 

[12] Fawzi Belblidia, Ernest Hinton, fully integrated design 

optimization of plate structures, Finite Elements in 

Analysis and Design 38 (2002) 227-244. 

[13] Li Q S,Liu D K,Tang J,Zhang N,Tam C M.Cornbinatorial 

optimal design of number and positions of actuators in 

actively controlled structures using genetie algorithms[J]. 

Journal of Sound and Vibration,2004, 270:611-624. 

[14] Jahng-Hyon Park,Haruhiko Asada.Integrated Structure/C- 

ontrol Design of a Two-Link Non-rigid Robot Arm for 

High Speed Positioning. Proceedings of the 1992 IEEE 

International Conference on Robotics and Automation, 

1992,(5):735-741. 

[15] Joseph C.Chen, Jacob Chen. Testing a New Approach for 

Learning Teamwork Knowledge and Skills in Technical 

Education. Industrial Technology.2004, 20(2): 1-10. 

[16] Tan Ping,Dyke S J,Richardson A,et a1.Integrated device 

placement and control design in civil structures using 

genetic algorithms[J]. Journal of Structural Engineering, 

ASCE,2005,131(10):1489-1496. 

[17] David A S,Paul N R,Lin Peiyang.GA-optimized fuzzy 

logic control of a large-scale building for seismic 


loads[J].Engineering Structures,2007,30(2):436-449. 

[18] CUI Ling-li, GAO Li-xin, ZHANG Jian-yu, XIAO 

Zhi-quan. Integrated Structure and Control Design for 

Flexible Manipulator System. Journal of Beijing 

University of Technology.2007,33(8). 

[19] George O’ Neal, An Analytical Approach to Integrated 

Structural and Control Design, PHD,University of 

Michigan,2001. 

[20]Anton C.Pil, haruhiko H. Asada. Integrated 

structure/control Design of Mechatronic Systems Using 

a Recursive Experimental Optimization Method. 

IEEE/ASME Transactions on Mechatronics.1996,9, 

1(3):191-203. 

Dingzhen Li (1972~), female, Nanyang, 

Henan province, China. She received the 

B.S. degree in automatic control of 

electrical engineering from Northeast 

Heavy Machinery Institute, Qiqihaer, 

China and the M.S. degree in Test and 

Measurement Technology and Instrument 

from Nanjing University of Science and 

Technology, Nanjing, China, respectively. 

She is currently an associate professor with 

Department of Electronics and Electrical Engineering, Nanyang 

Institute of Technology, Nanyang, China. Two books and thirty 

papers are published. Her main research interests include 

design of mechatronic systems and automation control of 

intelligent equipments. 

Ruimin Jin (1967~), male, Nanyang, Henan province, China. 

Ph.D. Three books and thirty papers are published. He is now a 

professor in Nanyang Institute of Technology. His main 

research interests are in the fields of in solar cells preparation 

and applications.


A Model to Select System Core and Its 

Application 

Chongming LI 

College of Management, Huazhong Normal University, Wuhan, 430079, China. 

Email: lichongming@eyou.com. 

Yue DING 

College of Management, Huazhong Normal University, Wuhan, 430079, China. 

Email: dingyue_2003@163.com. 

Abstract—The theory of system core is given a method to 

determine key elements of the system based on graph 

theory, but it is difficult to apply in practice for the question 

of how to change a system to a graph and multi-core in a 

system. This paper gives a method to change a system to a 

graph based on correlation analysis, and gives a modle to 

select system core based on cluster analysis. More, in the 

case of the real estate system of Wuhan, a diffusion index 

curve is given based on the elements of system core selected 

by the model, the result illustrates that Wuhan real estate 

arises in 1991 and to its peak in 1993, and then declines to 

the bottom in 1996; next, it fluctuates in a small range and 

becomes smooth since 1999, but a slight upward trend 

during 2000, this conclusion is consistent with the actual 

development status of the real estate in Wuhan and prove 

the validity of the model. 

Index Terms—System Core; Cluster Analysis; Correlative 

Analysis; Real Estate System; Wuhan 


When making research on system problem, it usually 

draws support from some indexes to analyze the whole 

system, that is to say, to establish an index system 

describing the system. But for many systems, especially 

those complex systems, they involve a lot of elements 

which have extremely complex relations within 

themselves. In order to study the system, it has to analyze 

the elements and relations of the system, for the structure 

and function of system determined by them. Under a 

complete system information situation, it can choose 

those indexes which related to the system as many as 

possible, but along with indexes increase, the redundant 

information which has nothing to do with the system also 

will increase. Besides, redundant information will 

submerge which we needed; meanwhile it will increase 

the analysis and computation difficulties. Because the 

elements of system play different roles in the system (XU 

Jin, WANG Yingluo, 1993), some are very important and 

some are unimportance.There are a large number of 

objective facts in nature and human society show that any 

system have some key elements, key elements of system 

play a dominant role to the system, so it hopes to discover 

the essential elements which have the key role through 


doi:10.4304/jcp.6.9.1913-1919 

some methods, and then uses these key elements to study 

the system. The theory of system core using the 

knowledge of graph theory to present a method that can 

determine the essential elements of system, which is also 

called the core of the system (XU Jin,1993), by the theory 

it can discover the essence, the main body and the key 

elements of the system. 

The theory was applied in complex system (WANG 

Jingguang, 2001), fault diagnosis (CAI Bing, ZHOU 

Liuding, 1994) and reliability of communication (CAO 

Qiguo, SUN Yugeng, 1997).But it's few used in 

social-economic systems, for the relations of elements in 

social-economic systems is too complex to explain by 

vertex cut sets and components, it's hard to be turned a 

social-economic system into a graph. Another problem is 

the lack of uniqueness to the system core, there is no 

explain to how a system core important than another in 

system core theory, many research describe the destroy of 

connected graph by vertex cut sets and components based 

on toughness of a graph (Chvata1 ,1973), then denote the 

importance of vertexes through the damage of the 

connected graph, just as relative rupture of 

graph(OUYANG Kezhi, OUYANG Keyi,1993) and the 

relation between degree of rupture and rupture 

number(ZHANG Shenggui, WANG Ziguo,1995), which 

give academic base to settle the question of multi-core. 

This paper gives a method to turn a system into a graph 

and formulates a model to select system core within the 

cluster analysis and grey correlative analysis, then the 

model was used in the real estate system of Wuhan, the 

conclusion according with the actual development status 

of the real estate of Wuhan, thereby demonstrating the 

validity of the model. 

II. SYSTEM CORE AND ITS MULTIPLE VALUED 

The system core theory given a method to study 

complex systems, it is describe the system center through 

the system core by qualitative and quantitative method. It 

is regard system as graph in system core theory and the 

system core is composed of vertex cut set which is 

important or has dominant role to the system function, it


is used the number of sets and connected components to 

calculate system core. 

The basic idea of system core theory is that such 

component elements couldn’t have the same effect on a 

given system, while some elements are minor and some 

are very important to the system (XU Jin, 1993). 

Eliminating or destroying these key elements will make 

the system break down. These several essential elements 

are called the core of this system. 

Suppose that X is a system, its elements 

are n x x x ,..., , 1 2 , if x i has relation with x j in X , 

then denote by x i x j , and connect this two with an 

edge e ij . In that way, we can use a graph to express the 

system X. The vertex set can represent element(index) of 

X, edges in the graph express the relations of them, then 

structuring out graph G of system X , its vertex set 

is ( ) { 1, 2,..., 

n} 

x x x G V = , edge set 

is E ( G) 

= { xix 

j | xiandxj 

have relation } , in practice the 

relations between xi and x j are broad, they depend on 

the nature of researched system and the problem. 

Definition1. Suppose that G is a connected 

graph, V ( G) 

≥ 4 , then 

h( G) 

= max{ ω ( G − S) 

− S ; S ⊂ C( 

G)} 

( 1 

) 

h (G) 

is called core degree which is the value of 

system core of system graph G , C (G) 

are all vertex 

cut sets in graph G , ω ( G − S) 

reflect the branches 

when graph G is cut off by vertex cut sets S ; S is 

the number of vertex in vertex cut sets S . 

∗ 

If there are vertexes cut sets S satisfied 

∗ ∗ 

h( G) 

= ω ( G − S ) − S ; 

(2) 

then ∗ 

S is the core of systemG . 

∗ 

S are the vertex-cut sets that satisfy the value of core 

degree, core degree meaning in graph theory is the most 

destructive measurement for graphG , so the definition 

of system core is based on that of vertex-cut sets and the 

connected graph, thus it first requests that graph G is 

connected, at the same time, there is not core in 

full-connected graph K n for the lack of vertex-cut sets 

in K n . 

From the definition of system core, any system which 

contains a binary relation can regard as a graph, vertexes 

are elements of the system, and edges are relations of 

elements. Since the born of system core theory, it has 

developed a lot in theory as well as practice, however, for 

the lack of uniqueness for most of the systems, there 

exists a problem concerning the option of core in 


application. As following two simple graphs, 1 G 

andG 2 . 

Figure1. Simple system graph 

from definition 1, it's know easily that the core degree of 

G 1 is 2, namely ( 1 ) 2 = G h ,the core of G 1 is unique, 

∗ 

S = { x1} 

; for G 2 , ( 2 ) 1 = G h ,but there are 3 

∗ 

∗ 

∗ 

cores, S 1 = { x2} 

, S 2 = { x3} 

, S 3 = { x2 

, x3} 

. 

Because the core are the vertex-cut sets of graph, so for 

some graphs of complex system, the number of the core 

is more than that of the vertex is also possible, under such 

circumstances, the most commonly used method is based 

on the nature of studied system and the studied problem 

to choose which core to analyze the system, if we want to 

know which core can mostly reflect the nature of system 

and solve the studied system problem, it need to analyze 

every core, if the number of core is more than that of the 

vertex, then taking advantage of system core to analyze 

the system is not only failed to simplify the problem, but 

make it more complex, there it need another kind of 

method to solve the problem, namely find out the core 

that can solve the studied system problem. 

III. GRAPH OF SYSTEM AND CLUSTER OF INDEXES 

Although there has a specific algorithm to solve the 

question of system core Identification, but the algorithm 

is only used to small connected graph and it is difficult to 

be used in complex system, In addition, system core 

theory and its specific applications main used in the 

communication network, the function communication 

network is transfer Information and connectivity is the 

basis of transfer Information, So the dominant vertex is 

core. However, much relations of social network is 

difficult to describe by vertex cut sets and connected 

graph, such as friendship, trust, acquaintance, like or 

other relations. 

In the course of resolving some problem about system, 

especially, socioeconomic system, it can transform the 

system into the form of graph. In the process of 

constructing graphs, the key of a system graph is how to 

make sure the edge between vertices (system index), as to 

a system, the relation between indexes should reflect 

integrity of system at first that is the correlation between 

indexes. So the graph of a system can be determined by 

the method of associative analysis.


As to indexes xi ( i = 1, 

2,..., 

n) 

taken as 

characterization of system X , there is relation 

between xi and x j which can be seen from the integrity 

of system (MIAO Dongsheng, 1998), the relation 

between xi and x j can be expressed by gray correlative 

analysis, gray correlative analysis is the measurement of 

relationship between two systems or two elements, and it 

describes the status including the magnitude of change, 

comparative change in direction and velocity, and so on of 

comparative change between elements in the process of 

system development. If the trend of change between two 

elements is consistent in the process of development, the 

two elements change in a high degree of synchronization, 

the degree of relationship is comparatively high, vice 

versa. Gray correlative analysis is a kind of method with 

which can be used to analyze and determine the degree of 

mutual effects between system elements and the degree in 

which elements contribute to system. 

The basic idea of gray correlative analysis is to 

determine the behavior of a system, find out numeric 

array of behavior, search for the elements which affect 

the behavior of system, collect the element data arrays 

affecting behavior of system, calculate the correlative 

degree between data array of each element and data array 

of behavior, so the relation between xi and x j can be 

expressed by correlation coefficient r ij 

r 

ij 

= 

∑ 

k= 

1 

m 

∑ 

k= 

1 

m 

( x ( k) 

− x )( x ( k) 

− x ) 

( x ( k) 

− x ) 

i 

i 

i 

i 

2 

m 

∑ 

k= 

1 

j 

( x ( k) 

− x ) 

j 

j 

j 

2 

(3) 

here, i , j = 1, 

2,..., 

n , is number of indexes of system; 

k = 1, 

2,..., 

m is number of indexes data; x i , x j is 

average of indexes data. 

Using formula (3), correlation coefficient r ij between 

xi and x j of system X can be determined, the 

correlation matrix of all indexes is R . 

⎡r11 

⎢ 

⎢ 

r21 

R = 

⎢ ... 

⎢ 

⎣r1n 

r12 

r22 

... 

r2n 

... 

... 

... 

... 

1 , r = r , i, 

j = 1, 

2,..., 

rii ij ji 

here, = n . 

r1n 

⎤ 

r 

⎥ 

2n 

⎥ 

... ⎥ 

⎥ 

rnn 

⎦ 

In the process of constructing graph of a system, the 

relation between xi and x j is generalized, in this paper, 

according to the character of system and the problem of 

system, it chooses a critical value r0 of a correlation 

coefficient, when rij ≥ r0 

, there is correlation between 


xi and x j , using an edge to join 

index i, j correspondingly. So it can gain graph G of 

system X. In order to make sure the graph of system is 

connected, choose the smallest correlation coefficient r c 

that can make all indexes of system to be a connected 

graph as critical value. 

As it can be seen from the correlation matrix R of 

indexes, some correlation coefficient are very big, the 

other are opposite smaller. So indexes with big 

correlation coefficient have strong correlation in system, 

then indexes can be grouped by clustering of index 

correlation: according to correlation matrix R , it's can 

search from the first row to find out r1 j ( j = 1, 

2,... 

n ) 

bigger than r c ( r1 j ≥ rc 

), and take the indexes xi , x1 

as 

one group, for example, if r12 , r15, 

r19 

are all bigger than 

x , x , x , x can be as one group, then 

r c ,the indexes 1 2 5 9 

searching from the second row to find out r2 j ≥ rc 

, then 

indexes xi , x2 

as one group, and so on, till last row of 

R ; then take the index with the smallest tab as 

representative stands for each group, if two group has the 

same index, merge the primary classification, for 

example, in the primary classification, x1 , x2 

, x5, 

x9 

is 

as one group, x7 , x9 

, x12 

is as another group, here 

, x , x 

x , finally take 

merge x7 9 12 into the group in which 1 

the indexes unlisted as one group respectively, by this 

way, n indexes can be divided to M groups, each index 

aggregate J i ( J i ∩ J j = Φ, 

i, 

j = 1, 

2, 

3,..., 

M )reflect 

system from one side, so index system which used to 

describe system should include one index per group at 

least. 

IV. MODELTO SELECT SYSTEM CORE 

The aim of turning a actual system into a graph is 

taking advantage of the system core theory to identify the 

key elements and simplify the indexes system of complex 

system. Therefore, after getting the system graph G by 

correlation analysis, we need to calculate the core of 

G .According to the formula (1) and (2), it can get the 

core of graph G , but the core is uniqueness. The 

definition of the core shows that the core is the main body 

of the system and it can completely describe the system, 

at the same time, because of the clustering process of 

system index, needing at least one in every kind of index 

to describe the system, if we want to completely describe 

a system, so, in the case that the core is not the only one, 

∗ 

0 

S should satisfy 

∗ 

S ≥ 

∗ 

0 M And S0 ∩ J i ≠ Φ (4)


∗ 

S 0 is the core value (the number of core element)of 

the system, M represents the kind number of indexes 

cluster when the critical value of correlation coefficient 

is r c , J i is the indexes set of class i , i = 1, 

2, 

3,..., 

M . 

It is can combine the correlation clustering with the 

core theory to solve the problem of multi-core. but in 

application, there is another problem that should be 

solved ,namely whether it's can find out the core that 

satisfy formula (4), if system X do not contain the core 

that satisfy formula (4), indexes will be farther clustered, 

if 

xk ∈ A , xl ∉ A , A ∉ M 

there is min( r kl ) , xl ∈ B ,take A and B as one 

cluster, then cluster indexes of system into M ′ 

categories( M ′ < M ),so it's can get a new clustering, 

seeing if there is a core 

then need further clustering till finding out the core 

∗ 

∗ 

S 0 , constitute elements of S 0 are the key ones of 

studied system. 

∗ 

S 0 satisfy formula (4), if not, 

V. A CASE STUDY IN REAL ESTATE SYSTEM OF 

WUHAN 

In this paper, the model of select system core is applied 

to the real estate system of Wuhan, establishing the index 

system of Wuhan real estate system, numbering these 

indexes(YE Yanbing, DING Lieyun, 2001) ,and getting 

index data from the year 1990 to 2001 as TABLE I. 

In order to make every index comparable, A 

dimensionless treatment is needed, it is use the method of 

initial value, and it’s very simple and easily 

understandable. 

Suppose that there is an original numeric array listed as 

follows 

( 0) 

( 0) 

( 0) 

( 0) 

x ( i) 

= { x ( 1), 

x ( 2),..., 

x ( n)} 

( 1) 

( 0) 

x ( i) 

is produced by x ( i) 

which processed with 

initial value method by formula (5). 

( o ) ( o ) 

( o ) 

( 1) 

x ( 1) 

x ( 2) 

x ( n) 

x ( i) 

={ ( o ) , ( o ) ,..., ( o ) } 

x 

( 1) 

x 

( 1) 

x 

( 1) 

x x x n 

(5) 

( 1) 

( 1) 

( 1) 

= { ( 1), 

( 2),..., 

( )} 

using formula ( 3 ) to get the related coefficient 

between indexes, when r 0. 

85 , take indexes as 

c 

vertexes v i , the relations of indexes as edges e ij , it's can 

get a connected graph of Wuhan real estate system, as 

Figure2. 

Take advantage of formula (1) and (2), calculating out 

that the value of system core is 6, obtaining 43 cores as 

TABLE II. 

According to the method of clustering, when the real 

estate coefficient critical Value r 0 = 0. 

85 , obtaining 

one clustering from the index system of the real estate 


= 

in Wuhan,{1, 7, 15}, {2, 6, 9, 13}, {3, 4, 8}, {5, 

10,11,12,14,16}. 

Compared to the clustering of indexes with the core in 

TABLE II, it can get the core which satisfies formula(4 

)is C31, indexes included in C31 can be substituted for 

the ones in table 1 to analyze the real estate system of 

Wuhan. It takes advantage of the Diffusion Index(DI) to 

describe the situation about the real estate system in 

Wuhan city from the year 1990 to 2001. It also verifies if 

the indexes included in C31 can describe the actual 

situation of real estate system in Wuhan, according to the 

indexes from C31, it gets the Diffusion Index curve as 

Figure 3. 

TABLE I. INDEXES DATA OF WUHAN REAL ESTATE SYSTEM 

Number 1 2 3 4 

Index Investment Residence Quantity of Construction 

Year 

investment Employment area 

1990 33452 84525 6793 641 

1991 34190 92853 7256 624 

1992 55821 112277 7576 671 

1993 246322 245672 12827 924 

1994 570618 470834 10652 1310 

1995 916194 630835 11173 1560 

1996 972609 678365 12743 1597 

1997 1065729 693015 9981 1585 

1998 1067997 814564 9894 1655 

1999 967130 814564 16049 1741.3 

2000 1013105 975576 11211 2086.5 

2001 1153400 861100 12421 2836.9 

CONTNUED TABLE I 

Number 5 6 7 8 

Index Housing Trade area Trade value House 

completed 

price index 

Year building area 

1990 534 27 10315 99.4 

1991 507.1 39 16025 98.9 

1992 488.5 82 39940 174.3 

1993 549.2 93 63601 132.8 

1994 692.6 179 113956 99.9 

1995 844.6 285 158145 109.9 

1996 1006.2 336 691315 208.6 

1997 1180.3 286 255678 98.3 

1998 1260.5 676 675331 98.1 

1999 1294.5 735 537350 104.4 

2000 1274.3 774 622500 110.8 

2001 1255.78 892 741400 100.1



Number 9 10 11 12 

Index Total real Total Per capita Consume 

estate GDP GDP income price 

Year 

index 

1990 2.12 176.83 1555.8 103 

1991 2.85 207.95 1771.68 107.3 

1992 3.47 255.42 2116.9 111.4 

1993 5.17 357.23 2872.9 119.8 

1994 8.96 485.76 3769.8 126.3 

1995 10.24 606.91 4453.9 118.4 

1996 12.7 782.13 4915.86 112.2 

1997 14.52 912.33 5573.04 103.1 

1998 24.14 1015.89 5912.52 97.4 

1999 26.32 1085.68 6198.24 96.1 

2000 30.28 1206.84 6953.94 100.6 

2001 37.08 1347.8 7305.05 99.5 


Number 13 14 15 16 

Index Capital Land sale Construction Per capita 

cost area industry living space 

Year 

output value 

1990 142665 89.42 207581 6.1 

1991 177894 98.8 237301 6.2 

1992 324181 130.8 303877 6.3 

1993 471368 293.05 502134 6.5 

1994 751348 275.72 795827 6.9 

1995 1114959 142.64 1071878 7.2 

1996 1353224 161.6 1113665 7.5 

1997 1220022 282.17 1088014 7.8 

1998 1297707 190 1165540 8.1 

1999 1322586 196.54 1352129 8.6 

2000 1411232 212.6 1553941 8.8 

2001 1653600 223.1 1098300 9.65 

From Figure 3, it is show that the development of 

real estate of Wuhan coincides with the development of 

the real estate in China, from1990 to 2002, the real estate 

market of Wuhan experienced a complete cycle which 

has 4 period, underway development of real estate in 

years 1978~1991. Since 1991, there is a quick 

development of real estate in China, for the development 

of real estate in the city near the sea and the increase of 

investment from abroad. Then there is a quick 

development of real estate in Wuhan, due to quick arising 

of real estate price. Real estate is excessive development 

and bubbles in real estate of China which leads the 

overheated economy, so The Chinese Government began 

to deflate money and restrict credit of real estate. The 

Chinese real estate has entered the adjustment period 

since the last of 1993, then the real estate of Wuhan 


decline from the peak in 1993 to the bottom in1996. 

Since the early of 1998, the government of China 

published series policies to arouse the development of 

real estate, the real estate of China has entered the golden 

age with a steady development, so the real estate of 

Wuhan has a steady development, since 1999 and a slight 

upward trend since2000. 

The result shows that Wuhan city cycle in real estate is 

general agreement with the national real estate cycle, this 

shows that elements which were chosen can be used to 

describe the real estate system of Wuhan, so the model to 

select system core is effective. 

TABLE II. SYSTEM CORES OF WUHAN REAL ESTATE SYSTEM 

Core Index 

C1 3,4,10,11,13,14 

C2 3 , 4,5,10,11,13,14 

C3 3, 4, 6 ,10 ,11, 13,14 

C4 3, 4,8 ,10, 11, 13,14 

C5 3 , 4 ,9,10,11,13,14 

C6 3, 4,10,11,12,13,14 

C7 3, 4,10,11,13,14,16 

C8 3,4,5,6,10,11,13,14 

C9 3,4,5,8, 10,11,13,14 

C10 3,4,5,8,11,13,14,16 

C11 3,4,5,9,10,11,13,14 

C12 3,4,5,10,11,12,13,14 

C13 3,4,5,10,11,13,14,16 

C14 3,4,6,8,10,11,13,14 

C15 3,4,6,10,11,12,13,14 

C16 3,4,6,10,11,13,14,16 

C17 3,4, 8,9,10,11,13,14 

C18 3,4,8,10,11,12,13,14 

C19 3,4, 8,10,11,13,14,16 

C20 3,4, 9,10,11,12,13,14 

C21 3,4,9,10,11,13,14,16 

C22 3,4,10,11,12,13,14,16 

C23 3,4,5,6,8,10,11,13,14 

C24 3,4,5,6,8,11,13,14,16 

C25 3,4,5,6,10,11,12,13,14 

C26 3,4,5,6,10,11,13,14,16 

C27 3,4,5,8,9, 10,11,13,14 

C28 3,4,5,8,9,11,13,14,16 

C29 3,4,5,8,10,11,12,13,14 

C30 3,4,5,,8,10,11,13,14,16 

C31 3,4,5,8,11,13,14,15,16 

C32 3,4,5,9,10,11,12,13,14 

C33 3,4,5,9,10,11,13,14,16 

C34 3,4,5,10,11,12,13,14,16 

C35 3,4,6,8,10,11,12,13,14 

C36 3,4,6,8,10,11,13,14, 16 

C37 3,4,6,10,11,12,13,14,16 

C38 3,4,8,9,10,11,12,13,14 

C39 3,4,8,9,10,11,13,14, 16 

C40 3,4,8,10,11,12,13,14,16 

C41 3,4,9,10,11,12,13,14,16 

C42 3,4,5,6,8,10,11,12,13,14 

C43 3,4,5,8,9,10,11,12,13,14


Figure 2. Connected graph of Wuhan real estate system 

Figure 3. The diffusion index curve of real estate of Wuhan in 

1990-2001 

VI. CONCLUSIONS 

Commonly, any one complex system has multiple 

cores and each core describes the system in different way, 

but according to system problems to be studied, it is can 

find a perfect core to study the system based on some 

method and model. From the case study in real estate 

system of Wuhan, the core of real estate system of 

Wuhan can be used to study the real estate system, for it 

is accords with the actual development status of the real 

estate of Wuhan and China, this means that the system 

can be replaced by the elements which are called system 

core, thereby demonstrating the validity of the model to 

select system core. 

With the development of system core theory, it will be 

widely used in the social and economic system. As an 

effective way to simplify the complex system, many 

economic problems can be studied and solved by the 

system core theory, just as the theory can be used in 

researching the stability of the society and the structure of 

economic system(WANG Jingguang,2001), core 


competence of enterprise(ZHAO Binxin, ZHAO 

Jinghua,2000), etc. 

REFERENCES 

[1] Xu jin. The theory of system core and its application. 

XiDian University Press,1994. 

[2] W. Duckworth, B. Mans. Connected domination of regular 

graphs. Discrete Mathematics, Volume 309, Issue 8, 28 

April 2009, Pages 2305-2322. 

[3] Bruno Escoffier, Laurent Gourvès, Jérôme Monnot. 

Complexity and approximation results for the connected 

vertex cover problem in graphs and hypergraphs .Journal 

of Discrete Algorithms, Volume 8, Issue 1, March 2010, 

Pages 36-49. 

[4] CAO Qiguo,SUN Yugeng. The Hypergraph Design 

Method of Multibus Structures of Reliable 

Communication Networks. Acta Electronica Sinica, 

1997,(10):55-62. 

[5] Konstantin Avrachenkov, Vivek Borkar, Danil 

Nemirovsky. Quasi-stationary distributions as centrality 

measures for the giant strongly connected component of a 

reducible graph .Journal of Computational and Applied 

Mathematics, Volume 234, Issue 11, 1 October 2010, 

Pages 3075-3090. 

[6] OUYANG Kezhi, OUYANG Keyi. Relative Breaktivity of 

Graphs. Journal of Lanzhou University(Natural Science 

Edition), 1993(3):78-82. 

[7] ZHANG Shenggui, WANG Ziguo. On Using Concept of 

Degree of Rupture for Designing Reliable Network. 

Journal of northwestern polytechnical universty, 

1995(2):310-313. 

[8] Firdovsi Sharifov, Hakan Kutucu. Minimum Cost ≤k 

Edges Connected Subgraph Problems.Electronic Notes in 

Discrete Mathematics, Volume 36, 1 August 2010, Pages 

25-32. 

[9] Hu Xuefeng. On the establishment and improvement of the 

statistical indexes system of Real estate, The Journal of 

ShanXi Einance and Economics University, 2002.2,86-89. 

[10] Ye Yanbing. Ding Lieyun, Design and study of real estate 

early warning indexes system. Optimization of Capital 

Construction, 2001.3,1-3. 

[11] Wang Xiaobo. Study on economy cycle and early warning. 

Press of Metallurgy Industry,1993. 

[12] Xu jin, XI Youmin, WANGYingluo. system core and core 

degree(I). Journal of Systems Science and Mathematical 

Sciences 1993,(02):20-28. 

[13] W. Ananchuen, N. Ananchuen, R.E.L. Aldred. The 

structure of 4-γ-critical graphs with a cut vertex .Discrete 

Mathematics, Volume 310, Issues 17-18, 28 September 

2010, Pages 2404-2414. 

[14] SHOU Jilin, LI Fei, Point Weighted Core and Coritivity of 

Network System and Its Applications, Systems 

engineering--theory and practice, 1996(6):58-63. 

[15] WANG Jingguang. Study of the Relation Between 

Reliability & Complexity of Information Systems 

Structure. Measurement & Control Technology, 

2001,(02):26-34. 

[16] ZHAO Bingxin; ZHAO Jinghua. Researching on the Core 

Competence with Networks. Chinese Journal of 

Management Science, 2000,(S1):45-51. 

[17] Michael R. Fellows, Guillaume Fertin, Danny Hermelin, 

Stéphane Vialette. Upper and lower bounds for finding 

connected motifs in vertex-colored graphs .Journal of 

Computer and System Sciences, In Press, Corrected 

Proof, Available online 3 August 2010.


[18] Stephen P. Borgatti.Identifying sets of key players in a 

social network.Computational & Mathematical Organization 

Theory, 2006, Volume 12, Number 1, Pages 21-34. 

[19] Yong Yeon Shin and Jai Sang Koh.An algorithm for 

generating minimal cutsets of undirected graphs. Journal of 

Applied Mathematics and Computing, 1998, Volume 5, Number 

3, Pages 681-693. 

[20] Chang C. Y. Dorea and Ary V. Medino.Anomalous 

Diffusion Index for Lévy Motions.Journal of Statistical 

Physics, 2006, Volume 123, Number 3, Pages 685-698. 

Chongming LI earned a B.S. in Electrical Technology from 

Shandong University of Technology in 1995, M.S. in 

Philosophy of Science and Technology from Wuhan University 

of Technology in 2001, and Ph.D. in Systems Engineering from 


Huazhong University of Science and Technology in 2005. He is 

Associate Professor of the College of Management, Huazhong 

Normal University. 

His current research interests include Information 

Management, Land Resource Management. 

Yue DING earned a B.S. in Accounting and Auditing from 

Wuhan University in 1992, M.S. in Industrial Economy from 

Wuhan University in 1998, She is currently working toward 

her Ph.D. degree in Economic Management at Zhongshan 

University. 

Her current research interests include Information 

Management, Land Resource Management.


De-noise Comprehensive Research 

On Airplane Cockpit Signals Recorded by CVR 

Dao-Lai Cheng 

College of Urban Construction and Safety Engineering, Shanghai Institute of Technology, Shanghai, 201418, China 

daolaicheng@163.com 

Chui-JieYi 

Qingdao R&D Center of Energy Equipment, Qingdao Technological University, Qingdao, China 

chuijieyi@vip.163.com 

Hong-Yu Yao 

Aviation Safety Technical Center, General Civil Aviation Administration of China (CAAC), Beijing, China 

yaohy@mail.castc.org.cn 

Abstract—The characteristic of cockpit sound recorded by 

CVR is the key evidence in investigating accident causes for 

wrecked aircraft. However, cockpit signals ( or CVR sound 

information) are complex, they include crew's voices (or 

pilot's voices), environmental noise and different kinds of 

backgrounds sound signals , and many factors from inside 

and outside cockpit affects the analysis results, especially 

noise. To obtain the pure cockpit signal (no noise) from 

mixed cockpit signals, after CVR Signals’ classification, the 

comprehensive de-noise research for cockpit signals are 

made, including the DWT threshold de-noise, the cockpit 

sound’ ICA de-noising based on BSS and wavelet de-noise 

are put forward. Through different de-noise methods 

comparative research for cockpit signals, some valuable 

conclusion can be drawn in the end of the paper. These 

conclusions are very useful for judging and diagnosing the 

wreckage aircraft by pure cockpit signals (background 

sound signals). 

Index Terms—Information process; airplane; Blind Source 

Separation (BSS); Independent Components Analysis 

(ICA); Cockpit signals 


In order to record flight information and to 

reconstruct or diagnosis aircraft accident, most all of 

large commercial aircraft (airplane) and other aircraft are 

necessary equipped with “black boxes, CVR & FDR”. 

Both CVR (Cockpit Voice Recorder) and FDR (Flight 

Date Recorder) play an indispensable role in aircraft 

accident investigation [1-3] . Compared with FDR, CVR 

(Fig.1) is one of the key evidence in the aircraft accident 

investigation. It is not only able to judge the unit's 

control, consciousness, determination, physical and 

mental state, but also can analyze the aircraft status and 

their environment by CVR. Cockpit signals (or CVR 

This paper is supported by the project of National Natural Science 

Foundation of China (Grant No. 61071203, 60772149). 

Copyright belongs to the papers all author and the units. 

Corresponding author: daolaicheng (daolaicheng@163.com) 


doi:10.4304/jcp.6.9.1920-1925 

sound information) includes crew's voices (or pilot's 

voices), environmental noise and different kinds of 

backgrounds sound signals (such as switch sounds, an 

overspend warning signal). To effectively identify the 

background sound signals, de-noise must be done firstly 

for airplane cockpit signals. To solve de-noise problems, 

comprehensive de-noise researches are made 

systematically in the paper. 

Figure 1. Cockpit voice recorder hull (CVR) 

The paper structures are arranged as follows: Firstly, 

CVR Signals are classified into speech & non-speech 

signals; then, the discrete wavelet transform (DWT) 

threshold denoise for cockpit signals are described; 

thirdly, principles of blind source separation (BSS) and 

the principle of de-noising analysis based on ICA of BSS 

are made in details, including blind source separation 

and analysis of OGWE, maximum ratio of signal to noise 

of blind source separation algorithm; Fourthly; process of 

cabin sound de-noising analysis based on ICA is done; 

finally , some research conclusions are obtained in the 

end of the paper. 

II. CVR SIGNALS’CLASSIFICAT [4-6] 

As we known that the CVR signals are recorded on 

4 channels connected by four wires: channel 1 from the 

cockpit area microphone of the CVR records non -speech 

information; channel 2 and channel 3 of the CVR record 

speech audio information from the captain and first 

officer’s audio selector panels; channel 4 records the


audio information from the jump seat/observer’s radio 

panel. 

To analyze and de-noise CVR’ signals, 

conveniently, here the signals from different channels are 

divided into speech and no-speech signals. Speech 

information means the voice of captain and first officer or 

other crew; non-speech information mean noises and 

background sounds. On-speech signals can be divided to 

noises and background sound signals. For noises, they 

include engine noise, exterior air flow noise, sliding 

noise, selector noise, motor noise, loud frequency 

noises induced by tape itself or the recording circuit, etc.; 

The background sound signals include overspend 

warning, fire alarm, flight altitude advice, wing flutter, 

turbulence, landing gear extension and retraction or 

flaps/slaps up, switch sound, wheel landing, thumps, 

clicks, squeaks, rattles, airframe vibration or whirl 

flutter, etc. . 

III. DISCRECTE WAVELET TRANSFORM(DWT) 

THRESHOLDING DENOISE [5-9] 

Threshold is a technique used for signal and image 

denoising. The discrete wavelet transform uses two types 

of filters: (1) averaging filters, and (2) detail filters. 

When we decompose a signal using the wavelet 

transform, we are left with a set of wavelet coefficients 

that correlates to the high frequency subbands. These 

high frequency subbands consist of the details in the data 

set. If these details are small enough, they might be 

omitted without substantially affecting the main features 

of the data set. Additionally, these small details are often 

those associated with noise; therefore, by setting these 

coefficients to zero, we are essentially killing the noise. 

This becomes the basic concept behind threshold-set all 

frequency subband coefficients that are less than a 

particular threshold to zero and use these coefficients in 

an inverse wavelet transformation to reconstruct the data 

set. 

There are the double-density DWT and doubledensity 

complex DWT for 1-D signals. Here, the doubledensity 

DWT method is only discussed in the following 

papers simply. 

The method can be implemented by program. This 

program takes as input two parameters, one of which is 

the noisy input signal (to be threshold) and the other of 

which is the threshold point. A sample CVR noisy signal 

is shown below (Fig.2), whose length is 512. To denoise 

the signal, we first take the forward double-density DWT 

over four scales. Then a denoising method, knows as soft 

threshold, is applied to the wavelet coefficients though 

all scales and sub bands. The soft threshold method sets 

coefficients with values less than the threshold T to 0, 

and then subtracts T from the non-zero coefficients. The 

double-density DWT method results in the following the 

CVR denoised signal (Fig.3). 


Figure 2. a sample CVR noisy signal 

Figure 3. the CVR signal by DWT threshold denoise 

IV. PRINCIPLES OF BLIND SOURCE 

SEPARATION((BSS) [10-13] 

Blind Source Separation (Blind Source Separation, 

BSS) is a process that the source signal is extracted from 

the mixed-signal. Normally, cockpit signals (CVR sound 

signals) are mixed signals, so, the pure cockpit signal 

(no noise) can be acquired by BSS. Some researches are 

made in the following paragraph. 

A Blind source separation 

Blind signal need to be in the certain conditions, 

and the "Blind" has double meanings: the source signal is 

unknown; how the source signals mixed is also unknown. 

(1) Definition of blind source separation 

Blind Source Separation (Blind Source Separation, 

BSS) is a process that the source signal is extracted from 

the mixed-signal. Blind signal need to be in the certain 

conditions, and the "Blind" has double meanings: the 

source signal is unknown; how the source signals mixed 

is also unknown. 

(2)The mathematical model of blind source 

separation 

The observation signal is M signals which are 

mixed, that is from the N statistical mixture of unknown 

source. Purpose of the BBS study is to separate the source 

signal from the signal mixture, which mathematical 

model described as follows: 

There are N independent sound sources, sound 

source signal is s i (t) ( i=1, 2, …… , N ), in order to


separate the source signals, M(M ≥ N) measurement 

points are measured, the measured signal is y j (t) 

( j=1 , 2 , ……, M ), the mathematical expression as 

follows(1): 

N 

∑ aij i 

i=1 

y j () t = s () t + Q 

Here: yj() t – number j masured signal; 

si() t - number i sound source; 

ij 

(j=1, 2, ……, M) (1) 

a - The corresponding coefficient; 

Q - Noise vector. 

This is the simplest model of blind source separation 

and is the separation of instantaneous linear mixing 

model. If the blind signal is collected in a silent room, 

the noise vector is negligible, That is to say: Q = 0, 

Equation 1 can be written in matrix form (2): 

Y=AS (2) 

Here: A - coefficient matrix constituted by a ij , unknown; 

S - Source signal, unknown; 

Y - Observed signal, known. 

(3) The uncertainty of source signal 

Blind source separation theory based on the 

observed signals and source signals which are 

independent only. Because the weak conditions for blind 

source separation, amplitude and the location of source 

signal separated are uncertain. Then, equation (1) is 

written as follows (3): 

N aij 

y j () t = d i si 

() t (j=1, 2, ……, M) (3) 

d 

∑ 

i= 

1 

j 

We can see that di is a constant, disi(t) as the source 

signal, the equation still holds, which reflects the 

uncertainty of the source signal amplitude, and secondly, 

when exchange location of coefficient and corresponding 

signal at the same time, the equation still holds. If the 

signal information is mainly contained in the signal 

waveform, the uncertainty does not affect the separation 

of the signals. 

(4) Simplify the problem of blind source separation 

BSS is to identify A and S in unknown 

circumstances of A and S, according to the independence 

of Y and the source signal, the separation process is to 

find a separation matrix W(4): 

Y=WX=WAS=CS, C=WA (4) 

For simplicity, the time t did not be written (the 

same below). If you can indeed find such a matrix, 

makes the C is the unit diagonal matrix (M = N), there 

are yi=is(i=1 , 2 , 3 , …, M) , then solution of the blind 

source separation problem is transformed to find the 

matrix W. 

B Independent components analysis (ICA) 

(1) Definition of independent components analysis 

(ICA) 


The current BSS description of the problem are 

mostly based on ICA model, BSS and ICA is equivalent 

normally, the difference is that ICA is a mathematical 

model that can solve various different problems, but BSS 

model is a real problem model that can be applied to 

other solutions, not only in ICA method. 

(2)The understanding of the independent 

components analysis (ICA) 

What is the independent components analysis (ICA) 

Look at a simple example: three people xi(t) are talking 

in a room at the same time , three microphones are placed 

at different locations, three speech signals si(t) are get 

through the three microphones, expression as follows(5): 

3 

() () 

x t = ∑ a s t 

(5) 

i ij i 

j= 

1 

Here: a ij (i, j = 1, 2, 3 )is the mixing coefficient. 

According to the distance between the room 

microphone and the people (ignore the delay and other 

additional factors such as the sound wave diffraction and 

refraction, etc.), the problem of a similar cocktail party 

is solved, only under the condition that three of their 

words can get on the basis of xi (t), . If more than three 

people and different locations, equation (5) can be 

expressed as the generalized mathematical model(6): 

xi ( t) 

= ai1 

s1( 

t) 

+ ai 

2s 

2 ( t) 

+ … + ains 

n ( t) 

(i, j = 1, 2, 3……, n) (6) 

Here, the independent component si () t and the 

mixed matrix a ij are unknown, the observation signal 

xi ( t) 

is known. ( t) 

xi is to be used to estimated ( t) 

and a ij in practical applications. That is to say, in order 

to make Y = S, the separation matrix W must be sought. 

In the blind source separation algorithm, such as the 

kurtosis maximization algorithm (based on higher-order 

cumulates), the minimum mutual information method, 

maximum likelihood estimation method, the joint 

digitalization method, the maximum SNR method, are 

based on ICA. 

V. PRINCIPLE OF DE-NOISING ANALYSES BASED 

ON ICA 

According to the theoretical research, ICA process 

is to separate the independent component mostly 

approaching each source signal, that is, set the objective 

function to achieve the approximation. To achieve this 

approximation, we establish an objective function J (W) 

with separation matrix W according to information theory, 

statistical theory and other methods, which is based on 

the definition of the objective function in this paper. W is 

found to make J (W) into a maximum (or minimum) 

value. 

In order to find both the best objective function J (W) 

and an effective algorithm for solution, sound de-noising 

analysis module is made based on ICA process in this 

paper. And the speech de-noising analysis is made in the 

pilot cabin based on ICA, including two aspects: one is 

OGWE blind source separation algorithm with the usage 

s i


of higher-order cumulates; the other is based on the 

largest ratio in signal to noise. 

A Blind source separation and analysis of OGWE 

Mixed signal need to be reduced into the mean and 

the whitened treatment in OGWE independent 

component analysis algorithm; because the mean is to 

simplify the calculation, while the whitening treatment 

can simplify the matrix, decreasing the problem 

complexity. Minimal contrast function (objective 

function) is used in OGWE algorithm; the rotation vector 

and higher-order cumulates matrix is calculated; and 

given transformations and angle calculations are 

combined for the realization of mixed-signal separation. 

B Maximum ratio of signal to noise of blind source 

separation algorithm 

Signal to noise ratio function (7): 

T 

s•s SNR = 10log T 

e•e (7) 

T 

s•s = 10log T 

( s− y) •( s− y) 

Here: S is the source signal, y is the estimated 

signal, noise signal is expressed as e = s − y 。 As the 

source signal is unknown, y (n) concludes noise, it is 

estimated that moving-average y ~ of y (n) is used to 

replace the source signal s, y = Wx ; ~ y = W~ 

x ; W is 

the separation matrix; x ~ is processed by the moving 

p 

average, ~ 1 

xi 

( n) 

= ∑ xi 

( n − j) 

, i=0, 1, 2, …, p-1 

p j= 

0 

Therefore, the objective functions of maximized 

signal to noise ratio can expressed as follows (8): 

T T 

Wxx W 

SNR = (8) 

T T 

W ( ~ x − x)( 

~ x − x) 

W 

The purpose of the algorithm is to find theW ~ . 

VI. PROCESS OF CABIN SOUND DE-NOISING 

ANALYSES BASED ON ICA [17-18] 

To take advantage of the theory of ICA, at least the 

same sampling signal points must be re-structured, one is 

the translated source signal, and the other is the pure 

noise signal in cabin. These two signals could be denoised 

by above-mentioned method. 

A Sound de-noising analysis by OGWE 

Here, a pilot's voice signal (the cockpit signals 

recorded by CVR inside the cabin) and a pure noise 

signal in cabin are given. They are separately shown in 

Fig.4 and Fig.5.The two kinds of signals are firstly 

analyzed through using OGWE based on higher-order 

cumulates ICA blind source separation. Fig.6 and Fig.7 

are separated signals respectively. 


Figure 4. Waveform of a typical cabin sound (source signal) 

Figure.5 Noise waveforms in cabin 

Figure. 6 Waveform of pilot voice signal 

Figure. 7 Noise Waveform 

From the research and comparisons, we found that 

the uncertainty of blind source separation is order and 

amplitude. Fig.5 and Fig.7 show that the clear distinction 

of the amplitude of the noise signals. In fact, obvious 

difference also is included in amplitude of the pilot voice 

signal. 

B Sound de-noising module by the maximum signal to 

noise ratio 

Maximum SNR can be used to analyze the blind 

source separation. Matrix W is got as follows: W= [-


.048306233, -0.965696385-0.914864537, 0.050990232]. 

Absolute value of the separation matrix W in the second 

diagonal approximates to 1, which being effectively 

separate the pilot voice signal. 

C Comparative analysis of de-noising module sound 

based on wavelet 

To illustrate the accuracy of de-noising by the ICAbased 

blind source separation method, the effects of denoising 

was compared based on wavelet in MATLAB. 

Three different methods are made for different 

signals, which are separately based on OGWE after denoising; 

based on maximum signal to noise ratio after denoising; 

based on Mat lab wavelet toolbox after denoising. 

Their wave charts are shown from Fig.8~Fig.10. 

FFigure.8 Pilot Speech Signal Waveform 

Based on OGWE after de-noising 

Figure.9 The Pilot Speech Signal Waveform 

Based on SNR after de-noising 

Figure. 10 The pilots voice signal aveform 

Based on SNR after de-noising 

From these charts, we can conclude that blind 

source separation method based on ICA has the 

approximate effects with the successful wavelet for denoising 

except small difference in amplitude; and, that 

blind source separation method based on ICA has the 

approximate effects with the successful wavelet for denoising 

except small difference in amplitude. 


D Comparative analyses of denoising cabin sound on 

Short-term zero-crossing rate and short-term 

energy 

Short-term zero-crossing rate and short-term energy 

are further used to accurate the analysis of the three 

methods of de-noising effect. 

In order to analyze CVR source signal by typical 

short-time zero-crossing module, the 2984 points is 

sampled in source signal. A unit has 600 points in shortterm 

analysis; two units with 300 points overlapping 

regions, and the signal are divided into 9 units. For the 

short-term zero-crossing number; OGWE; the maximum 

SNR method and Wavelet De-noising, their comparative 

analysis results can be seen in Table.1, Table 2. 

From Table 1, OGWE blind source separation 

method is much closer to wavelet in the Zero-crossing 

numbers and de-noising effect. 

From Table 2, OGWE blind source separation 

method is much closer to wavelet in the short-term 

energy and de-noising effect. De-noising effect based on 

maximum signal to noise ratio is somehow less. 

TABLE Ⅰ RESULTS OF THE DIFFRERENT 

ANALYSES MODULE 

ShortZero OGWE Maximum SNR Method Wavelet 

1 26 18 26 

2 29 28 29 

3 29 34 29 

4 28 23 26 

5 33 35 34 

6 34 39 33 

7 29 25 27 

8 31 29 29 

9 30 27 27 

TABLE II SHOR ENERGY, OGWE, MAXIMUM SNR METHOD, 

WAVELET DENOISE RESULTS OF SOUND 

Short 

Energy 

OGWE Maximum SNR Method Wavelet 

1 19.027 34.324 14.221 

2 39.04 17.329 29.87 

3 34.362 17.852 22.56 

4 34.156 20.756 24.455 

5 41.723 27.137 33.812 

6 37.916 23.129 33.026 

7 31.498 19.81 24.235 

8 14.418 23.966 12.089 

9 5.2407 13.585 4.651 

VII. SPECTRAL ANALYSES OF CANIN SOUND 

SIGNAL BY ADOBE AUDITION SOFTWARE 

Pilots’ voice signal and pure noise signal are both 

directly separated by ICA-based blind source separation 

method and spectra analysis by Adobe Audition software. 

Their results can be got in Fig.11. The three different 

spectrum (the source signal spectrum, pilot voice signal 

spectrum and the noise signal spectrum can be seen in 

figure11 from left to right. 

- 

1 

0


From the spectra of three signals in Fig.11, we can 

conclude that 1) higher energy intensity are occurs for 

separating the pilot speech signal than the source noisy 

signal in voice by ICA-based blind source separation 

methods;2)the noise is at a relatively high frequency band. 

Figure.11 Cabin sound spectrums 

VIII. CONCLUSIONGS 

Through above typical cockpit signals de-noise 

analysis and research; some conclusions can be drawn as 

followings: 

(1)The cockpit signals are complex, can be 

classified into speech information and non -speech 

information; There are noise signal and background 

sound signals; 

(2)Discrete wavelet transform (DWT) shareholding 

denoise play an vital role for cockpit signals; 

(3)The pilots’ source signals of aircraft cockpit voice 

are separated based on the principle of ICA blind source 

separation methods. 

(4)Three different blind source separation methods 

(based on OGWE, maximum signal to noise ratio after 

de-noising and Mat lab wavelet toolbox after de-noising) 

have the approximate effects except small difference in 

amplitude. 

(5)Through the spectra of three signals (the source 

signal spectrum, pilot voice signal spectrum and the 

noise signal spectrum), we can conclude that 1) higher 

energy intensity are occurs for separating the pilot speech 

signal than the source noisy signal in voice by ICA-based 

blind source separation methods;2)the noise is at a 

relatively high frequency band. 

REFERNCES 

[1] Shu Ping, Zhong Minzhu, Yang Lin, Amelioration of 

Cockpit Voice Recorder Decoding System [M].Beijing: 

Aviation Industry Press, 2004.97-100. 

[2] Kendall W.Neville (USA).“Research on Flight Techniques 

and Aviation Safety (ISBN 7-5364-6055-4), ” Published by 

Si Chuan Science and Technology Press, Chengdu, 

China, pp.40-46, August 2006. 

[3] Sound Spectrum Study Cockpit Voice Recorder – 12. 

Statistical Summary of Commercial Jet Airplane Accidents 

Worldwide Operations 1959 – 2008.Boeing, 2009 

Statistical Summary, JULY 2009. 

[4] McKinney Martin F, Jeroen Breebear.Feature for Audio 

and Music Classification [EB/OL]. http://mckinney. 


philps.com, 2008 

[5] Dao Laicheng, Chui Jieyi, Hongyu Yao.Nonstationary 

quiver spindle background sound analysis of airplane vie 

Wigner-Ville and Wavelet time-scale distribution[J] . 

Published by China Machine Press, Beijing, China, 

vol.43, no.5 pp.150-154, May 2007. 

[6] Hairong Guo, Daolai Cheng, Zhoufeng Liang.Sound 

spectral analyses of black boxes based on Matlab and VC 

program [J].Micro computer information, Beijing, China, 

2008, 24(3):299 -301\ 

[7] Zhibing Gao, Chuijie Yi, Yangming Zhou The information 

management system’ design and implement for the 

characteristics of cockpit background sound [J].Micro 

computer information, Beijing, China, 2009, 25(34):18 - 

19. 

[8] Daolai Cheng, Chuijie Yi, Hongyu Yao, .Sound Signal 

Analysis of CVR Based on CVDS, WT & CZT 

Algorithm[J].DCDIS Series B, Vol.14(S5), 2007: 129- 

134. 

[9] CHENG Daolai, YI Chuijie, Zhang Zhiqiang. 

Comparative Analyses and Experiment Verification on 

Cockpit Background Sound’ Characteristic Frequency [J]. 

Fourth International Conference on Innovative Computing, 

Information and Control December 7-9, 2009, Kaohsiung, 

Taiwan. 

[10] DaoLai Cheng, QingCheng Wang ChuiJie Yi, HongYu 

Yao.Analysis and Research for Airplane Cockpit 

Sound’ICA Denoise Based on Blind Source Separation 

Principle.2011 International conference on Industry, 

information System and Material Engineering 

(IISME2011)(C).April 16-17, 2011, Guangzhou, China 

[11] Malgorzata Zygarlicka, Janusz Mroczka. Reduction of the 

cross-terms of the Wigner Ville distribution by image 

processing. X international PhD workshop OWD’2008, 

215-220. 

[12] ZhiQiang Zhang Characteristics acquirement and analyses 

aircraft black boxes cockpit voice [P]. a dissertation for the 

master degree in engineering.Qingdao, Shangdong 

province, China. Qingdao Technological University, 

2010.12 

[13] Zhibing Gao, Chuijie Yi, Yangming Zhou The information 

management system’ design and implement for the 

characteristics of cockpit background sound [J].Micro 

computer information, Beijing, China, 2009, 25(34):18 - 

19. 

Daolai Cheng, born in 1965, PhD, 

Professor, the vice director of College of 

Urban Construction & Safety 

Engineering, Shanghai Institute of 

Technology, Shanghai (201418), China. 

His research interests include thermal 

energy and power engineering, signal 

process. 

Tel:+86-21-60873631, 13311998959 

E-mail:daolaicheng @163 . com 

Chuijie Yi, born in 1958, PhD, professor, president of Qingdao 

Technological University, China. His research interests include 

noise and vibration control. 

Hongyu Yao, born in 1963, PhD, researcher, chief engineer of 

General Civil Aviation Administration of China (CAAC), 

China. His research interests include cockpit signal analyses.


Fuzzy Support Vector Machines Control 

for 6-DOF Parallel Robot 

Dequan Zhu 

Institute of Intelligent Machines, Chinese Academy of Sciences, Hefei, China 

Department of Automation, University of Science and Technology of China, Hefei, China 

College of Engineering, Anhui Agricultural University, Hefei, China, 

Email: adqzhu@sina.com 

Tao Mei 

Institute of Intelligent Machines, Chinese Academy of Sciences, Hefei, China 

Email: tmei@iim.ac.cn 

Lei Sun 

College of Engineering, Anhui Agricultural University, Hefei, China 

Email: SL961102@163.com 

Abstract—In order to realize the trajectory tracking control 

of six degrees of freedom parallel robot, the dynamics 

equation of six degrees of freedom parallel robot was 

established. The parallel robot has obvious nonlinear, 

uncertainty characteristics and external disturbance, so the 

sliding mode variable structure theory was introduced into 

the system control. A fuzzy support vector machines control 

strategy based on sliding mode control was designed to 

reduce the oscillation of the sliding mode control. 

Parameters of fuzzy support vector machines controller 

were optimized by hybrid learning algorithm, which 

combines least square algorithm with improved genetic 

algorithm, to get the optimal control performance for the 

controlled object. The controller designed consists of a fuzzy 

sliding mode controller and a fuzzy support vector machines 

controller, and the compensation controller is selected by 

comparing switching function with the thickness of 

boundary layer. Simulation results show that under the 

condition of model error and external disturbance, the 

control strategy designed gets tracking effect with high 

precision and speed. 

Index Terms—parallel robot, fuzzy control, support vector 

machines, sliding mode control, dynamics equation 


The six degrees of freedom (6-DOF) Stewart platform 

parallel robot is a closed-loop mechanism in which the 

end-effector (mobile platform) is connected to the base 

by six extensible legs [1]. Compared with serial robot, 

the parallel robot has potential advantages in terms of 

compliance, accuracy, high speed and payload. Therefore, 

it has been used in precision lathe, assembly robotic 

manipulator and electronics manufacture [2,3]. Because 

of the complex condition and the uncertain object, the 

parallel robot is not only a complicated nonlinear 

multivariable and strong coupling system, but also a time- 


doi:10.4304/jcp.6.9.1926-1934 

varying system. The parallel robot is not controlled 

accurately and its track is not kept better by the general 

model-based control method [4,5]. 

In the domain of artificial intelligence techniques for 

the system control, different control algorithms have been 

used to realize the trajectory tracking control for the 

parallel robot. The sliding mode control algorithm has 

complete adaptability for system disturbance and stirring, 

which is extensively applied in the control of the parallel 

robot [6,7]. In the fuzzy control, the mathematical model 

for the system does not be set up precisely and the joints 

of robots can be decoupled, but fuzzy control system is 

easily influenced by nonlinear, time-varying and random 

disturbance [8]. Neural network control algorithm has 

many advantages, such as self-learning, self-organizing, 

self-adaptive capacity, nonlinear and parallel distributed 

processing, and so on. However, it also has the congenital 

defects, such as it falls into local minimum easily, and it 

is weakly normalized for few samples [9-11]. These 

defects make it difficult to meet control precision for 

parallel robot. 

Thus, some control methods are combined to realize 

the trajectory tracking control for the parallel robot. A 

cascade-control algorithm based on a sliding mode in the 

legspace was proposed by Hongbo Guo, Yongguan Luo, 

Guirong Liu and Hongren Li to realize the trajectory 

tracking control of hydraulically driven six degrees of 

freedom parallel robotic manipulator [6]. A control 

approach which is based on the coupling of sliding mode 

and multi-layers perceptron neural networks was 

proposed by Achili B, Daachi B, Amirat Y and Ali-cherif 

A to deal with the robust adaptive control tracking of a 6 

degree of freedom parallel robot [9]. A sliding mode 

control with discontinuous projection-based adaptation 

laws was proposed by Yangjun Pi and XuanyinWang to 

improve the tracking performance of the parallel robot


manipulator [10]. A robust neural-fuzzy-network control 

system was presented by Rongjong Wai and Pochen 

Chen to realize the joint position control of an n-rod 

robotic manipulator for periodic motion in order to deal 

with the uncertainties in application, such as friction 

forces, external disturbances, and parameter variations 

[12]. A new discrete sliding mode control approach for 

parallel robot was presented by Shaocheng Qu and 

Yongji Wang to achieve accurate servo tracking in the 

presence of load variations, parameter variations and 

nonlinear dynamic interactions [13]. A fuzzy-PI 

compound control system for three-cylinder hydraulic 

parallel robot was designed by Qidan Zhu, Xunyu Zhong 

and Bo Xu [14]. 

Support vector machines, which is based on the 

structural risk minimization rule, overcomes the 

shortcoming that neural network structure relies on the 

experience of designer. Its topology structure is decided 

by support vectors. It solves these problems well, such as 

high dimension, local minimum and small samples, and 

has advantages of both neural network and traditional 

model [15,16]. So the support vector machines is 

combined with fuzzy control to design a fuzzy support 

vector machines controller for parallel robot to reduce the 

chattering in sliding mode control. It is important to 

select the proper SVM parameters for improving the 

learning and generalizing capacity of the control system 

[16]. Thus, the fuzzy proportional coefficients were 

adjusted with the controlled object, and the parameters of 

the controller were optimized by least square (LS) 

learning algorithm and improved genetic algorithm (IGA) 

in order to improve control precision and working 

stability of the parallel robot. 

II. DYNAMICS MODEL OF 6-DOF STEWART PLATFORM 

PARALLEL ROBOT 

l6 l l l 

5 

4 3 

l1 

l2 

Figure 1. Schematic diagram of 6-DOF Stewart platform parallel robot 

A. Reference frame 

The 6-DOF Stewart platform parallel robot is shown in 

Fig.1. It consists of mobile platform, base platform and 

six extensible leg, each of which is connected with the 

two platforms by spherical joints [2,3]. The legs are 

driven by six servo-electromotors. To describe the motion 

of the mobile platform, two reference frames are chosen: 

a fixed reference frame {B, XB, YB, ZB} attached to the 

base platform and a mobile reference frame {P, XP, YP, ZP} 

attached to the mobile platform, as shown in Fig.1. Six 

coordinates are used to further describe the position and 


the orientation of the mobile platform in detail. Three 

coordinates are the positional displacements [Xp, Yp, Zp] T , 

which describe the position of a fixed point in the mobile 

platform with respect to the fixed reference frame. The 

other three coordinates are the angular displacements, 

represented by Euler angles [γ, β, α] T , which describe the 

orientation of the mobile platform with respect to the 

fixed reference frame. Therefore, the generalized 

coordinate vector, whose elements are the six variables 

chosen to describe the position and orientation of the 

mobile platform, can be defined as [XP, YP, ZP, γ, β, α] T . 

The rotation matrix from mobile reference frame to fixed 

reference frame can be described as follow: 

⎡cαcβ 

cαsβsγ 

− sαcγ 

sαsβcγ 

⎤ 

R = 

⎢ 

⎥ 

⎢ 

sαcβ 

cαcγ 

+ sαsβsγ 

sαsβsγ 

− cαsγ 

⎥ 

⎢⎣ 

− sβ 

cβsγ 

cβcγ 

⎥⎦ 

B 

(1) 

P 

where c(⋅) denotes cos(⋅); s(⋅) denotes sin(⋅). 

The ith leg vector li & with respect to the fixed reference 

frame can be described as 

B 

l& 

i = c& 

+ RP 

B& 

i − P& 

i ,i=1,2,,6 (2) 

where c is the translation vector of the origin of the 

mobile reference frame with respect to the fixed reference 

frame; Bi is the position vector of the ith joint point of 

base platform with respect to the fixed reference frame; 

Pi is the position vector of the ith joint point of mobile 

platform with respect to the mobile reference frame. 

The extended length of the ith leg is described as 

Δ l = l& 

− l 

(3) 

i 

i 

where li0 is the original length of the ith leg. 

Well-controlled lengths of six legs make the mobile 

platform follow the desired trajectory. 

B. Dynamics equation 

In order to solve the kinetic energy and the potential 

energy of parallel robot, the whole system is separated 

into the mobile platform and the six legs with the base 

platform. 

Suppose the angle velocity of mobile platform is ωk , 

and then the kinetic energy of mobile platform KEh, 

which includes translational and rotating kinetic energy, 

can be described as 

1 2 2 2 

KEh = ( mu 

( x& 

P + y& 

P + z& 

P ) + ωk 

Ichωh 

) (4) 

2 

where mu is the mass of mobile platform; xP, yP, zP is the 

displacement about the axis XP, YP, ZP respectively; o I ch is 

the inertia matrix of the mobile platform with respect to 

mobile reference frame. It can be computed by 

o h T 

I ch = RIc 

R 

(5) 

where R is the corresponding rotating matrix, defined by 

the angle rotating rule of Roll-Pitch-Yaw; h 

I c is the 

rotating inertia with respect to the mobile reference frame. 

It is described as 

⎡I 

⎤ 

X 0 0 

P 

h ⎢ 

⎥ 

Ic 

= ⎢ 0 IY 

0 

P ⎥ 

(6) 

⎢ 

⎥ 

⎣ 

0 0 I ZP 

⎦ 

i 0 

T 

o


where I , I , I is the rotating inertia with respect to 

X P 

YP 

Z P 

the axis XP, YP, ZP respectively. 

The angle velocity of mobile platform is described as 

v r r 

ω = & γR 

( α ) R ( β ) X + & βR 

( α) 

Y + & αZ 

k 

Z B 

YB 

⎡CαCβ 

− Sα 

0⎤ 

⎡ & γ ⎤ 

= 

⎢ 

⎥ ⎢ ⎥ 

⎢ 

SαCβ 

Cα 

0 & 

⎥ ⎢ 

β 

(7) 

⎥ 

⎢⎣ 

− Sβ 

0 1⎥⎦ 

⎢⎣ 

& α ⎥⎦ 

where R , Z R is the corresponding rotating matrix 

B YB 

respectively, defined by the angle rotating rule of Roll- 

Pitch-Yaw; c(⋅) denotes cos(⋅); s(⋅) denotes sin(⋅). 

The kinetic energy of mobile platform is rewritten as 

1 T 

KEh = q& 

M h ( q) 

q& 

(8) 

2 

where Mh(q) is defined as 

⎡mu 

0 0 0 0 0 ⎤ 

⎢ 

⎥ 

⎢ 

0 mu 

0 0 0 0 

⎥ 

⎢ 0 0 mu 

0 0 0 ⎥ 

M h ( q) 

= ⎢ 

⎥ 

⎢ 0 0 0 M h44 

M h45 

− I xsβ 

⎥ 

⎢ 0 0 0 M 

⎥ 

h54 

M h55 

0 

⎢ 

⎥ 

⎢⎣ 

0 0 0 − I xsβ 

0 I X ⎥⎦ 

(9) 

2 

2 2 

2 2 

where M h44 

= I X sin β + IY 

sin γ cos β + I Z cos γ cos β , 

M h45 

= ( I x − IZ 

) cosγ 

sinγ 

cos β , M h454 

= ( IY 

− I Z ) cosγ 

sin γ cos β , 

2 

2 

M h55 

= IY 

cos γ + I Z sin γ . 

The potential energy of mobile platform is described as 

T 

P = m gz = [ 0 0 m g 0 0 0] 

q (10) 

h 

u 

p 

u 

where g is the gravity acceleration. 

The extensible legs, which are driven by the servoelectromotors, 

are separated into the cylinders and the 

rods. They are regarded as the rigid parts with rotating 

inertia with respect to themselves. Each leg is represented 

by the centroid point of it. 

The position of the ith leg centroid point Gi is 

described as 

1 

r m i 

BiGi 

m ilBi 

m i Li 

l i ui 

lˆ 

2 r 

= [ 1 + 2 ( − 2 )] = [ i + Li 

] ui 

m1 

i + m2i 

m1i 

+ m2i 

(11) 

where l im 

i l im 

i 

lˆ 

1 1 − 2 2 

i = 

, l1i is the distance between the 

m1i 

+ m2i 

center point of the ith base joint and the centroid point of 

cylinder of the ith leg; l2i is the distance between the ith 

upper joint and the rod of the ith leg; m1i is the mass of 

cylinder of the ith leg, m2i is the mass of rod of the ith leg; 

Li is the length of the ith leg; ui is the orientation of the ith 

leg. ui can be defined as 

r BiPi 

u = 

(12) 

i 

L 

i 

The velocity of centroid point of the ith leg VGi r 

described as 

is 

r 

VGi 

lˆ 

r r 

i 

r r m2 

= [ VP 

− ( V 

i P ⋅ u 

i i ) ui 

] + 

L 

m + m 

r 

V (13) 

Pi 

i 

where VPi r is the velocity vector of the centroid point of 

upper joint of the ith leg Pi. 


Z B 

1 

2 

The kinetic energy of the ith rod is described as 

1 r r 

T 

K L = ( m 

i 1 i + m2i 

) VG 

V 

i G 

(14) 

i 

2 

Then, substituting equation (13) for equation (14), the 

kinetic energy of rod can be rewritten as 

1 

r r r 

T 

T r r r 

T 

K L = ( m 

i 1i 

+ m2i 

)[ hiVb 

V 

i b − k 

i iVb 

u 

i i ( ui 

) Vb 

(15) 

i 

2 

lˆ 

2 

where i m2 

i 

m2 

i 2 

hi 

= [ + ] ; ki 

= hi 

−[ 

] . 

Li 

m1i 

+ m2i 

m1i 

+ m2i 

The total mass of six legs can be written as[3] 

T 

2 

T 

1 

6 

2] 

∑ 

i= 

1 

M = [ J ( H − J K J ) J ( m + m ) (16) 

legs 

legs 

where Klegs denotes the total kinetic energy of six legs, 

rT 

rT 

rT 

rT 

rT 

rT 

−1 

J1 

= diag{ 

u1 

, u2 

, u3 

, u4 

, u5 

, u6 

} , J 2 = Vbq& 

r 

. 

The Jacobian matrix J for parallel robot may be 

described as 

J = J1 J2 (17) 

Suppose the structure of six legs is same, and then the 

total kinetic energy of six legs may be expressed as 

6 

1 T 

Klegs = ∑ K L = q& 

M legs ( q) 

q& 

(18) 

i 2 

i= 

1 

The potential energy of six legs may be written as 

6 

lˆ 

m2i 

Plegs 

=∑ [( m1i 

+ m2i 

)( + )( z p − xb′ 

Sβ 

+ y CβSγ 

z CβCγ 

)] 

i b′ 

+ ′ 

i 

L m 

1 

1 m 

i 

i i + 

= 

2i 

(19) 

Lagrange equation for parallel robot can be expressed 

as 

d ⎡∂ 

L( 

q, 

q& 

) ⎤ ∂L( 

q, 

q& 

) 

⎢ ⎥ − = τ i , i=1,2,• • •,.n (20) 

dt ⎣ ∂q& 

i ⎦ ∂qi 

where q∈R n is nominal coordinate; L is the Lagrange 

function of mechanism system; τi is the force on the ith 

nominal coordinate. 

The dynamic Lagrange equation of 6-DOF rigid 

parallel robot is described as 

M ( q) 

q& 

& ( t) 

+ V ( q, 

q& 

) q& 

( t) 

+ G( 

q) 

+ τ = J τ ( t) 

(21) 

6 

m 

where q, q& 

, q& 

& ∈ R is the position, the velocity and the 

acceleration of centroid point of mobile platform 

respectively; M(q) is the mass of mobile platform and six 

legs; ( q, 

q& 

) is the velocity vector of mobile platform 

V m 

and six legs; G(q)∈R 6 is the gravity vector of mobile 

platform and six legs; τ(t)∈R 6 is the control force vector 

of mobile platform and six legs; τd∈R n is the model error 

and external disturbance; M(q), Vm ( q, 

q& 

) , G(q) can be 

computed by the equations of the kinetic energy and the 

potential energy of mobile platform and six legs. 

If the model designed is precise, control law of robot is 

expressed as 

T 

( t) = j [ M ( q)( 

qd 

− kve 

− k pe) 

+ V ( q, 

q) 

q + G( 

q)] 

− τ & 

& 

& & (22) 

where qd is expect angle, e = q − qd, e& = q& 

− q& 

. d 


control equation of stable close-loop system is expressed 

as 

1 

d 

1i 

T 

2i


+ k e + k e = 0 

& (23) 

e v p & 

Because it is difficult to build the real model of the 

object precisely, desired model is only built. Its control 

law is expressed as 

T 

( t) = J [ M o ( q)( 

qd 

− kve 

− k pe) 

+ Vo 

( q, 

q) 

q + Go 

( q)] 

− τ 

& 

& 

& & 

(24) 


below equation is established. 

M ( q) 

q& 

& + C( 

q, 

q& 

) q& 

+ G( 

q) 

= M o ( q)( 

q& 

& d − kve& 

− k pe) 

+ ( q, 

q& 

) q& 

+ G ( q) 

+ F( 

t) 

(25) 

Vo o 

If ⊿M=MO − M, ⊿V=VO − V, ⊿G=GO − G, the 

below equation is obtained. 

−1 e& & + kve& 

+ k pe 

= M O ( ΔMq& 

& + ΔVq& 

+ ΔG) 

(26) 

From equation (26), the decline of the control 

performance is brought up partly by the parameter and 

non-parameter uncertainty. Thus, the compensation for 

the uncertainty is needed for improving the control 

precision of the robot. 

Ⅲ. STRUCTURE OF CONTROL SYSTEM OF PARALLEL 

ROBOT 

According to dynamics equation of the robot and its 

control law, sliding mode control method and fuzzy 

support vector machines are used to compensate for the 

uncertainty. Structure of control system for robots is 

shown in Fig.2. In the figure, q denotes the real track of 

robot; qd denotes the expect track; e denotes the error 

vector; FSVMC denotes the fuzzy support machines 

controller; FSMC denotes the fuzzy sliding mode 

controller; R(e) denotes the switching function, whose 

inputs are e and e& . 

qd 

q& & d 

q& d 

kv 

k 

kv 

p 

M O (q) 

u 

u 

1 

2 

GO (q) 

τ 

CO ( q, 

q& 

) 

Figure 2. Structure of control system for robots 

Compensation control law for fuzzy sliding mode is 

defined as 

τ i ( t ) = M ( q)( 

q& 

& d − kve& 

− k pe) 

+ V ( q, 

q& 

) q& 

+ G( 

q) 

+ ui 

, i = 1. 

2 

(27) 

where u1 is the control compensation of fuzzy sliding 

mode controller; u2 is the control compensation of fuzzy 

support vector machines controller . 

Function R(e) is used to decided which controller as 

the compensation controller. Suppose the thickness of 

boundary layer is Q ; if R ( e) 

> Q , FSMC is used for 

control compensation; if R ( e) 

< Q , FSVMC is used for 


q& 

q 

control compensation; if R ( e) 

= Q , the below sliding 

algorithm is used for control compensation. 

u ( t) 

= ( 1 − d ( e)) 

u2 

( t) 

+ d ( e) 

u1( 

t) 

+ τ (28) 

where τ is the control torque; u1 and u2 are the outputs of 

FSMC and FSVMC respectively; d(e) is the sliding 

function, whose function is to make FSMC and FSVM 

switch smoothly. 

⎧ 0, 

E( 

t) 

∈AFSVMC 

⎪ 

d( 

e) 

= ⎨0 

< d( 

e) 

< 1, 

E( 

t) 

∈ A − AFSVMC 

(29) 

⎪ 

⎩ 1, 

AFSMC 

where AFSVMC is the control range of fuzzy support 

machines controller; AFSMC is the control range of sliding 

mode controller. They are defined as 

A = E | e ≤ Q 

(30) 

{ } p 

{ E e ≤ + ζ } 

FSVMC 

A = | Q 

(31) 

p 

where ζ is the thickness of switching layer, 0 < ζ < Q . P 

norm is defined as 

1/ 

p 

p ⎪⎫ 

i ⎬ 

⎪⎧ 

e = e 

p ⎨∑ 

(32) 

⎪⎩ i ⎪⎭ 

According to P norm definition, the sliding function 

used is described as 

d( e) 

= max( 0, 

sat( 

R − Q) 

/ ζ ) (33) 

⎧ y, 

where sat ( y) 

= ⎨ 

⎩sgn( 

y), 

y < 1, 

y = e . 

p 

If R ( e) 

< Q , e∈AFSVMC, and d(e)=0; if R ( e) 

> Q + ζ , 

e∈AFSVMC, and d(e)=1. If A′ = A − AFSVMC 

, 0


Suppose sliding mode plane s=0, the below equation is 

obtained. 

ce + e& 

= 0 

(35) 

The control compensation of sliding mode may be 

defined as 

u = τ = −Ksign( 

s) 

(36) 

1 

1 

where K is the coefficient matrix, K≥M ( ΔMq 

+ ΔCq 

+ ΔG+ 

τd) 

− && & . 

Lyapunov function is defined as 

1 

V = (37) 

2 

2s 

The below equation is obtained by equation (1), (7) 

and (8). 

V = ss& 

≤ −K( 

t) 

| s | < 0 

(38) 

Under the new control law, sliding mode exists and 

may be attained. Sliding mode switches at the sliding 

mode plane s=0, which brings the strong oscillation. 

In the sliding mode control law, the switching gain 

K(t), which is used to compensate the uncertainty, easily 

arises the chattering. For reducing the chattering, K(t) 

should be varied with the time. 

Fuzzy rules may be expressed as: 

If s>0, K(t) should be increased; If s0. 


Substituting K ˆ ( t) 

for K(t) in equation (36), control law 

can be rewritten as 

u = −Kˆ 

× sign( 

s) 

(40) 

1 

Ⅴ. FUZZY SUPPORT VECTOR MACHINES CONTROLLER 

A. Structure of FSVM controller 

Structure of FSVM controller is shown in Fig.5. Inputs 

of FSVM system are { qd, q& d }; the output compensation 

of FSVM system for uncertainty is u2; qd is the desired 

positions of two joints; q is the real positions of two 

joints; e is the position error of two joints; e& is the varying 

rate of position error of two joints. 

qd 

+ 

_ 

e 

LS & IGA hybrid 

optimization algorithm 

d/dt 

e& 

Ke 

E 

e& Ke& E& FSVM controller 

K( 

x, 

x1) 

K 

K( 

x, 

x1) 

e 

t 

. 

. 

. 

K x, 

x ) 

( 1 

U 2 

Ku 

u2 

Figure 5. Structure of FSVM control system 

Parallel 

robot 

Inputs and outputs of control system are fuzzified; {e, 

e& , u2} are fuzzified respectively as {E, E& , U2}; their 

fuzzy subsets are {NB, NM, NS, Z, PS, PM, PB}, which 

respectively denotes {negative big, negative middle, 

negative small, zero, positive small, positive middle, 

positive big}; quantified grades of them are {-6, -5, -4, -3, 

-2, -1, 0, 1, 2, 3, 4, 5, 6}. Triangle distribution function is 

selected as their membership function. According to 

varying range of inputs and outputs of control system, the 

proportional coefficients of fuzzy control are e K , e K & and 

Ku respectively. Decision process of fuzzy control is 

2 

described by the three-layer SVM model. 

1) Input layer 

The function of input layer is that input variables are 

fuzzified as the input of control system x. 

⎧E 

= Kee 

⎪ 

⎨E& 

= Kece& 

(41) 

⎪ 

⎩x 

= ( E, 

E& 

) 

2) Hidden layer 

The function of inner layer is to realize the kernel 

computation of four-dimension input x and SVM. The 

radial basis function kernel function is expressed as 

2 2 

K( x, 

x ) = exp( − x − x / 2σ 

) (42) 

i 

where σ is the kernel width, which reflects the radius of 

close boundary. 

3) Output layer 

The function of output layer is to obtain the real input 

control value of the controlled object by computing SVM 

regression value. 

i 

q


N ⎛ 

⎜ y ( xi 

) = ∑ akK 

( xi, 

xk 

) + b i = 1, 

2 (43) 

⎜ 

k = 1 

⎜ 

⎝ui 

= KUi 

y( 

xi 

) 

The controller parameters are optimized by the hybrid 

learning algorithm. First, least square algorithm is used 

for off-line optimize the parameters of support vector 

machines. Then, improved genetic algorithm is used for 

on-line optimizing the parameters of support vector 

machines and fuzzy proportional coefficients. 

B. Parameter optimization of control system 

The parameters of affecting SVM performance are 

number of training sample set D, penalty coefficient γ, 

kernel width σ and insensitive coefficient ε , and so on. 

The system performance is also affected by fuzzy 

proportional relations between real values and fuzzy 

values in decision-making process of control system. 

Only when these parameters are in finite range, the 

system has the better control performance. Optimal 

parameter combination varies with the object. 

1) Off-line optimization of γ and σ 

Because ε may be preset by the noise, which reflects 

the prediction of data noise level by support vector 

machines, least square algorithm was only used to offline 

optimize γ and σ. 

First, the method of rising exponent was used to search 

the proper γ set and σ set. For example, γ = 2 -4 , 2 -2 , …, 

2 10 ;σ = 2 -10 , 2 -8 ,… , 2 -2 . 

Second, using the method of net search, the parameter 

combination(γ , σ) was selected to verify it crossly. The 

training sample set was divided into S groups {G1, 

G2,⋅⋅⋅GS}. Selecting randomly S-1 groups as training set, 

and another as verifying set; generalization capability was 

evaluated with the following equation. 

S 

N N 2 

P = ∑∑ y − y x i 

i= 1 y∈Gi 

)) x | ( ( (44) 

where Gi is ith group verifying set; y N is the sample of 

verifying set; xi is the parameter vector [a,b] where D - Gi 

was set as train sample; y(x|θi) is the output of SVM 

system. 

Final, the parameter combination was selected 

circularly to verify it crossly and the performance index P 

was computed until the optimal parameter combination 

(γ , σ) is obtained. 

2) On-line optimization of fuzzy SVM parameters 

(a) Encoding 

Because of the complexity and continuity of 

optimizing process of SVM parameters, the coding 

method of float number is used to avoid the effect of 

binary coding on the evaluation of algorithm performance 

and computing precision. 

(b) Selection of fitness function 

In improved genetic algorithm, the individual 

evolution is decided by individual fitness value. Thus the 

individual fitness value need be computed. The individual 

is sequenced by fitness value and sequenced population is 


lined out by the upper limit and lower limit. Fitness 

function is used to evaluate SVM individual and fitness 

function designed influences directly the performance of 

genetic algorithm. According to the feature of robot 

system, fitness function was described by the sum of 

error among given system input and real output. It was 

expressed as 

M 

Fi = Ei 

max 

k = 1 

i i 

− Σ S ( k) 

−T 

( k) 

(45) 

where i = 1,2,⋅⋅⋅, N is the number of individual in 

population; k is the number of individual variable. 

Mean error EMSE of system track was expressed as 

N1 

1 

2 2 ∑ [( Ti 

− f ( xi 

)) ] 

i= 

1 EMSE 

= 

(46) 

N 

(c) Genetic operation 

Genetic operations include selection, crossover and 

mutation. Its objective is to substitute the new generation 

population into next generation population. The 

procedure of operation is given as follows: 

Step1: Generation of initial population; 

Step2: Re-evaluation and adding age; 

Step3: Selection of parents: prior selection of elder 

individuals; 

Step4: Crossover and mutation: generation of new 

individuals; 

Step5: Evaluation: evaluation of new individuals; 

Step6: Natural selection: selection considering the 

diverseness of individuals; 

Step7: Steps 2 to 6 are repeated until the convergence 

is achieved. 

In genetic operation, set population is 200; set 

crossover probability is 0.75; set mutation probability is 

0.02. Each parameter was set by following: 

1 255 

1 31 

D ∈[ 

1, 

512] 

, γ ∈[ 

, 255 ] σ ∈[ 

, 127 ] 

256 256 , 32 32 , 

1 63 1 15 

ε ∈[ 

, 255 ] e [ , 255 ] 

64 64 , 16 16 

∈ K 

1 63 

Ke& ∈[ 

, 255 ] 

, 64 64 , 

1 127 

Ku 

∈[ 

, 15 ] 

128 128 . 

where D, γ, σ and ε are coded respectively by 8 bit (8 bit 

integer), 9 bit (9 bit integer),14 bit(8 bit integer, 6 bit 

decimal),16bit (8 bit integer, 8 bit decimal) binary strings; 

fuzzy proportional coefficients Ke, e K & , Ku are coded 

respectively by 12 bit (8 bit integer, 4 bit decimal), 10 bit 

(4 bit integer, 6 bit decimal), 11bit (4 bit integer, 7 bit 

decimal), binary strings. Thus, they are coded by 72 bit 

binary strings and their values are discrete. Their units are 

1, 1/256, 1/64, 1/32, 1/16, 1/64, 1/128 respectively. After, 

the individual fitness function are computed using these 

parameters, the individuals in the new population are 

selected by the desired value. 

Ⅵ. SIMULATION AND APPLICATION 

To verify the effectiveness of the presented control 

strategy for the parallel robot, the comparative simulation 

experimental researches were carried out between the 

1


designed control strategy and the fuzzy sliding mode 

control strategies using experimental simulation. In 

simulation experiment of control performance, the mobile 

platform is driven by six asymmetric cylinders with a 

cylinder diameter of 85mm and a rod diameter of 64mm, 

and a full stroke of 840mm, which are controlled by six 

servo-electromotors. The installed sensors measure the 

leg lengths and forces between the centroid point of rods 

and the heads of the cylinders. The radius of the base 

platform and the mobile platform are 1250 and 540mm 

respectively. The simulation experiments of parallel robot 

were conducted by the simulation software. In the 

simulation experiments, the experimental values (100, 15, 

1.0, 0.01, 100,1.0,0.1) are set as the initial values of 

control parameter combination (D, γ, σ, ε, Ke, e K & , Ku) , 

trace error EMSE is 2.134; after hybrid optimization, 

optimal parameter combination (D, γ, σ, ε, Ke, e K & , Ku) is 

(1.6, 3.2, 0.2, 1.5, 65, 0.4, 0.07), trace error EMSE is 0.014. 

The experiments concerned position tracking of centroid 

point of mobile platform for the following reference 

trajectories qd(t)=1.0+0.20sin(2πt) mm and 

qd(t)=1.0+0.40sin(2πt) mm by the designed control 

system. The experimental results are shown in Fig.6 and 

Fig.7. 

Figure 6. Position tracking (qd(t)=1.0+0.20sin(2πt)) 

Figure 7. Position tracking (qd(t)=1.0+0.40sin(2πt)) 

The experiments concern the position tracking error of 

centroid point of mobile platform for the following 

reference trajectories qd(t)=1.0+0.20sin(2πt) mm and 



Fig.9. 


It can be seen that the designed control system 

performs much better than the fuzzy sliding mode control 

methods from Fig.6, Fig.7, Fig.8 and Fig 9. It can be 

obtained that position tracking error of centroid point of 

mobile platform for the following reference trajectory 

qd(t)=1.0+0.40sin(2πt) mm is smaller than that for the 

following reference trajectory qd(t)=1.0+0.20sin(2πt) mm 

by the designed control system. From above figures, 

using designed controller, track error is low. Considering 

uncertainty and complexity of the system, the track error 

may be permitted. 

Figure 8. Position tracking error (qd(t)=1.0+0.20sin(2πt)) 

Figure 9. Position tracking error (qd(t)=1.0+0.40sin(2πt)) 

The experiments concern the velocity tracking of 





Fig.11. 

Figure 10. Velocity tracking (qd(t)=1.0+0.20sin(2πt))


Figure 11. Velocity tracking (qd(t)=1.0+0.40sin(2πt)) 

The experiments concern the velocity tracking error of 





Fig.13. 

Figure 12. Velocity tracking error (qd(t)=1.0+0.20sin(2πt)) 

Figure 13. Velocity tracking error (qd(t)=1.0+0.40sin(2πt)) 

It can be seen that the designed control system 

performs much better than the fuzzy sliding mode control 

methods in the velocity tracking from Fig.10, Fig.11, 

Fig.12 and Fig.13. It can be obtained that velocity 

tracking error of centroid point of mobile platform for the 

following reference trajectory qd(t)=1.0+0.40sin(2πt) mm 

is smaller than that for the following reference trajectory 


system. 


The experiments concern control input of legs for the 

following reference trajectories qd(t)=1.0+0.20sin(2πt) 

mm and qd(t)=1.0+0.40sin(2πt) mm by the designed 

control system. The experimental results are shown in 

Fig.14 and Fig.15. It can be seen that the control input of 

legs are different and the control input for the following 

reference trajectory qd(t)=1.0+0.40sin(2πt) mm is 

different from that for the following reference trajectory 


system. 

Figure 14. Control input of leg (qd(t)=1.0+0.20sin(2πt)) 

Figure 15. Control input of leg (qd(t)=1.0+0.40sin(2πt)) 

Ⅶ. CONCLUSION 

In this paper, the dynamics equation of 6-DOF parallel 

robot was established. According to the dynamics 

equation, a fuzzy support vector machines control 

strategy based on the sliding mode control was proposed. 

The proposed controller consists of a fuzzy sliding mode 

controller and a fuzzy support vector machines controller. 

The compensation controller is decided by comparing the 

switching function with the thickness of boundary layer. 

Using improved GA and FL algorithm to optimize the 

performance parameters of support vector machines and


the fuzzy proportional parameters, a better control system 

was obtained. The system uncertainty and external 

disturbance was compensated. Experimental simulation 

was carried out with 6-DOF parallel robot to investigate 

the effectiveness of the proposed control method. The 

simulation results show that the control method designed 

gets tracking effect with high precision and speed, as well 

as reduces the chattering under the condition of existing 

model error and external disturbance. 

REFERENCES 

[1] I. A. Bonev and J. A. Ryu, “New method for solving the 

direct kinematics of general 6-6 Stewart platforms using 

three linear extra sensors.” Mechanism and Machine 

Theory, vol. 35, pp. 425- 436, March 2000. 

[2] I. Davliakos and E. Papadopoulos, “Model-based control 

of a 6-dof electro-hydraulic Stewart–Gough platform.” 

Mechanism and Machine Theory, vol.43, pp.1385-1400, 

November 2008. 

[3] M. R. Sirouspour and S. E. Salcudean, “Nonlinear control 

of hydraulic robots.” IEEE Transactions on Robotics and 

Automation, vol.17, pp.173-182, April 2001. 

[4] Xiufeng Zhang and Lining Sun, “Research of precise 

parallel robot control method and system.” Chinese 

Journal of Mechanical Engineering, vol. 40, pp.177-180, 

April 2004. 

[5] B. Dasgupta and T. S. Mruthyunjaya, “The Stewart 

platform manipulator: A review.” Mechanism and Machine 

Theory, vol.35, pp.15-40, January 2000. 

[6] Hongbo Guo, Yongguang Liu, Guirong Liu and HongRen 

Li, “Cascade control of a hydraulically driven 6-DOF 

parallel robot manipulator based on a sliding mode.” 

Control Engineering Practice, vol.16, pp.1055-1068, 

September 2008. 

[7] S. Islam , P X Liu, “Output feedback sliding mode control 

for robot manipulators.” Robotica, vol.28, pp.975-987, 

December 2010. 

[8] J. S. Oh, J. B. Park and Y. H. Choi, “Stable path tracking 

control of a mobile robot using a wavelet based fuzzy 

neural network.” International Journal of Control 

Automation and Systems, vol.3, pp.552-563, December 

2005. 

[9] B. Achili, B. Daachi, Y. Amirat and A. Ali-cherif, “A 

robust adaptive control of a parallel robot.” International 

Journal of control, vol.83, pp.2107-2119, October 2010. 

[10] Yangjun Pi and Xuanyin Wang, “Trajectory tracking 

control of a 6-DOF hydraulic parallel robot manipulator 

with uncertain load disturbances.” Control Engineering 

Practice, vol.11, pp.185-193, January 2011. 


[11] T. Dierks and S. Jagannathan, “Neural Network Output 

Feedback Control of Robot Formations.” IEEE 

Transactions on Systems Man and Cybernetics part B- 

Cybernetic, vol.40, pp. 383-399, April 2010. 

[12] Rongjong Wai and Pochen Chen, “Robust Neural-Fuzzy- 

Network Control for Robot Manipulator Including 

Actuator Dynamics.” IEEE transactions on industrial 

electronics, vol.53, pp.1328-1349, August 2006. 

[13] Shaocheng Qu and Yongji Wang, “Discrete sliding mode 

control for a parallel robot.” In Proceedings of the fifth 

international conference on machine learning and 

cybernetics, pp.871-874, Dalian, China, August 2006. 

[14] Qidan Zhu, Xunyu Zhong and Bo Xu: “Design of Fuzzy-PI 

Compound Control system for three-cylinder hydraulic 

parallel robot.” In Proceedings of the 2007 IEEE 

international conference on mechatronics and automation, 

pp.989-993, Harbin, China, August 2007. 

[15] C.-F. Lin, Sh.-D. Wan., “Fuzzy support vector machines.” 

IEEE Transactions on Neural Networks, vol.13, pp.464- 

471, March 2002. 

[16] F. Orabona, C. Castellini, B. Caputo, L. Jie, and G. 

Sandini , “On-line independent support vector machines.” 

Pattern Recognition, vol.43, pp.1402-1412, April 2010. 

Dequan Zhu received B.S. degree in agricultural mechanism 

from Anhui Agricultural University and received M.S. degree in 

mechanical and electronic engineering from Hefei University of 

Technology in 1997 and 2005 respectively. Currently, he is an 

associate professor at Anhui Agricultural University, and a PhD 

candidate in automation at University of Science and 

Technology of China. His major research experiences and 

interests include modern agricultural equipment and intelligent 

control. 

Tao Mei received B.S. degree in precision mechanism from 

Zhejiang University and received Ph.D. degree in mechanics 

from University of Science and Technology of China in 1982 

and 2001 respectively. Currently, he is a researcher at Institute 

of Intelligent Machines, Chinese Academy of Science. 

His major research experiences and interests include robotics 

and intelligent control. 

Lei Sun received M.S. degree in control theory and engineering 

from Hefei University of Technology and received Ph.D. degree 

in detect technology and automatic mechanism from University 

of Science and Technology of China in 2004 and 2008 

respectively. Currently, he is an lecturer at Anhui Agricultural 

University. His major research experiences and interests include 

modern agricultural equipment and intelligent control.


Parameters Optimization of Least Squares 

Support Vector Machines and Its Application 

Chunli Xie 1,2 

1. Dalian University of Technology/ School of Electronic and Information Engineering, Dalian, 116024, China 

2. Dalian Nationalities University/College of Electromechanical and Information Engineering, Dalian, 116024, China 

Email: chunlix@dlnu.edu.cn 

Cheng Shao 1 and Dandan Zhao 3 

3. Dalian Nationalities University/School of Computer Science and Engineering, Dalian, 116024, China 

Email: cshao@dlut.edu.cn, zhaodd@dlnu.edu.cn 

Abstract—Parameters optimization plays an important role 

for the performance of least squares support vector 

machines (LS-SVM). In this paper, a novel parameters 

optimization method for LS-SVM is presented based on 

chaotic ant swarm (CAS) algorithm. Using this method, the 

optimization model is established, within which the fitness 

function is the mean square error (MSE) index, and the 

constraints are the ranges of the designing parameters. 

After having been validated its effectiveness by an artificial 

data experiment, the proposed method is then used in the 

identification for inverse model of the nonlinear underactuated 

systems. Finally real data simulation results are 

given to show the efficiency. 

Index Terms—Least Squares Support Vector Machines, 

Parameters Optimization, Chaotic Ant Swarm Algorithm 


A novel statistical learning method called Support 

Vector Machines (SVM) was presented by Vapnik in 

1995. Due to the advantages such as the complete 

statistical learning theory foundation and perfect study 

ability, SVM has become quite an active research field in 

machine learning and broadly used in many fields such as 

pattern recognition and regression estimation problems 

[1, 2]. The classical training algorithm of SVM is 

equivalent to solving a quadratic programming with 

linear and inequality constraints. Least squares support 

vector machines (LS-SVM) has been recently introduced 

by Suykens et al. as reformulations to standard SVM [3, 

4], which simplifies the training process of standard SVM 

in a great extent by replacing the inequality constraints 

with equality ones. The simplicity of LS-SVM promotes 

the applications of SVM, and many pattern recognition 

and function approximation problems have been tackled 

with LS-SVM in the last decade [5-9]. 

The parameters in regularization item and kernel 

function are called parameters in LS-SVM, which play an 

important role for the algorithm performance. The 

Corresponding author: Chunli Xie. 


doi:10.4304/jcp.6.9.1935-1941 

existing techniques for tuning the parameters in LS-SVM 

can be summarized into two kinds: one is based on 

analytical techniques, the other is based on heuristic 

searches. The first kind of techniques determines the 

parameters with gradients of some generalized error 

measures [10]. And the second kind of techniques 

determines the parameters with modern heuristic 

algorithms including genetic algorithms (GA), simulated 

annealing algorithms (SA), particle swarm optimization 

algorithms (PSO) and other evolutionary algorithms [11- 

15]. Iterative gradient-based algorithms rely on smoothed 

approximation of a function. So, it does not ensure that 

the search direction points exactly to an optimum of the 

generalization performance measure which is often 

discontinuous. Grid search [16] is one of the conventional 

approaches to deal with discontinuous problems. 

However, it needs an exhaustive search over the space of 

parameters, which must be time consuming. This 

procedure also needs to locate the interval of feasible 

solution and a suitable sampling step. 

In this paper, a novel algorithm of parameters 

optimization is presented based on the principles of the 

chaotic ant swarm (CAS) algorithm. Inspired by the 

chaotic and self-organizing behavior of the ants in nature, 

the novel CAS [17] algorithm is developed in 2006, 

which combines the chaotic behavior of individual ant 

with the intelligent foraging actions of ant colony via the 

organization variable for solving optimization problems. 

Similar to GA, EA and PSO, the CAS algorithm is a 

population-based optimization tool, which searches for 

optima by updating generations. However, unlike GA and 

EA, the CAS algorithm does not need evolutionary 

operators such as crossover and mutation. Compared to 

GA and EA, the advantages of CAS algorithm are that it 

possesses the capability to escape from local optima, is 

easy to be implemented, and has fewer parameters to be 

tuned. Compared to PSO, the advantages of CAS 

algorithm are that it has higher convergent precision. The 

CAS algorithm has been successfully applied to 

parameters estimation, artificial network training and 

fuzzy system control, etc [18-26]. The CAS algorithm is 

used to the parameters optimization of LS-SVM, and the


feasibility of this approach is examined on the testing 

function and nonlinear under-actuated systems. 

This paper is organized as follows. The LS-SVM 

regression algorithm is briefly reviewed in Section 2. 

Parameters optimization algorithm based on the CAS 

algorithm is addressed in Section 3. The results of testing 

and simulation are presented to demonstrate the 

effectiveness of the proposed method in Section 4. The 

application of LS-SVM based on the CAS Algorithm is 

given in Section 5. Finally, the paper is concluded in 

Section 6. 

II. LS_SVM REGRESSION 

The LS-SVM, evolved from the SVM, changes the 

inequality constraint of a SVM into an equality constraint 

and forces the sum of squared error (SSE) loss function to 

become an experience loss function of the training set. 

Then the problem has become one of solved linear 

programming problems. This can be specifically 

described as follows [4]: 

Given the following training sample set (D): 

{ ( , y ) k = 1, 

2, 

L, 

N} 

D = x k k 

where N is the total number of training data pairs, 

k 

n 

R ∈ x is the regression vector and ∈ R is the 

n 

output. According to SVM theory, the input space R is 

mapped into a feature space, and then the linear equation 

in the feature space can be defined as: 

T 

f ( x) = w ϕ ( x) 

+ b 

(1) 

h 

where the nonlinear mapping ϕ : R → R maps the 

input data into a so-called high dimensional feature space 

(which can be infinite dimension). The regularized cost 

function of the LS-SVM is given as: 

where, 

1 T 1 

min J ( w , e) 

= w w + γ 

2 2 

T 

n 

y k 

N 

2 

∑ ek 

k = 1 

s. t. 

yk 

= w ϕ ( xk 

) + b + ek 

, k = 1, 

2, 

L, 

N (2) 

h n 

w ∈ R is the weight vector, ek ∈ R is slack 

variable, b ∈ R is a bias term and γ ∈ R is regularization 

item. The Lagrangian corresponding to Eq. (2) can be 

defined as follows: 

L ( w,b,e; α) 

= 

N 

−∑ 

k = 1 

k 

T { w ( x ) + b + e y } 

J ( w,e) α ϕ 

− (3) 

where α k ∈ R( 

k = 1, 

2, 

L, 

N ) are the Lagrange multipliers. 

The KKT conditions can be expressed by 

N 

∑ 

k = 1 


k 

w = α ϕ(x 

) 

(4) 

k 

k 

k 

n 

k 

T 

α = γe 

(5) 

N 

∑ 

k = 1 

k 

k 

α = 0 

(6) 

k 

w ϕ ( x ) b + e − y = 0 

(7) 

k 

+ k k 

After elimination of w and e k , the solution of the 

optimization problem can be obtained by solving the 

following set of linear equations 

⎡b⎤ 

⎡0 

⎢ ⎥ = ⎢v 

⎣α⎦ 

⎢⎣ 

T 

1 N 

T N 

, N ] ∈ R 

with y = [ y , L, 

y ] ∈ R , 

v −1 

T ⎤ ⎡0⎤ 

−1 ⎥ ⎢ ⎥ 

Ω + γ I ⎥⎦ 

⎣ y⎦ 

r¡ 

N 

= [ 1, 

L, 

1 

T 

] 

N 

∈ R 

α = [ α1 , L α and Ω is an N × N kernel matrix. 

By using the kernel trick [2], one obtains 

T 

Ω = ϕ ( x ) ϕ( 

x ) = ( x , x ) , ∀k, 

l = 1, 

2, 

L, 

N. 

kl 

k 

l 

K k l 

And the resulting LS-SVM regression model becomes 

N 

∑ 

k = 1 

(8) 

f ( x) = α K( 

x, 

x ) + b 

(9) 

where α k , b are the solution to Eq. (8). 

k 

Note that the dot product ϕ ⋅) ϕ( 

⋅) 

in the feature space 

( T 

is replaced by a prechosen kernel function K( ⋅, 

⋅) 

due to 

the employment of the kernel trick. Thus, there is no need 

to construct the feature vector w or to know the nonlinear 

mapping ϕ(⋅) explicitly. Given a training set, the training 

of an LS-SVM is equal to solving a set of linear equations 

as Eq. (8). This greatly simplifies the regression problem. 

The chosen kernel function must satisfy the Mercer’s 

condition [2]. Possible kernel functions are, e.g.: 

Linear kernel 

K( x , x ) = x ⋅ x . 

k 

Polynomial kernel 

k 

l 

K ( x , x ) = ( x ⋅ x + 1) 

. 

l 

Gaussian RBF kernel 

k 

k 

l 

K( x , x ) = exp( − x − x / 2 ) . 

k 

l 

l 

k 

m 

k 

2 2 

l σ 

where d denotes the polynomial degree, σ is the kernel 

(bandwidth) parameter. 

It is well known that LS-SVM generalization 

performance depends on a good setting of regularization 

parameter and the kernel parameter. In order to achieve 

the better generalization performance, it is necessary to 

select and optimize these parameters. 

III. PARAMETERS OPTIMIZATION OF LS_SVM BASED ON 

CAS ALGORITHM


A. Overview of CAS Algorithm. 

Ants have attracted many scientists’ significant 

interests because their colonies can achieve the selforganizing 

behavior and the high level of structure. Most 

of the existing ant-inspired optimization algorithms are 

based on the random meta-heuristic of nondeterministic 

probability theory. However, Cole suggested that ant 

colony exhibits a periodic behavior while single ant show 

low-dimensional deterministic chaotic activity patterns 

[27]. From the view of dynamics, the chaotic behavior of 

single ant has some relation to the self-organizing and 

foraging behaviors of ant colony. The chaotic behavior of 

individual ant and the intelligent organization actions of 

ant colony are adaptations to the environment. These 

behaviors help the ants to find food and survive. 

According to the theory, a novel optimization algorithm, 

called CAS algorithm, was presented. 

In the CAS algorithm, the chaotic system 

θ + = θ exp( µ ( 1−θ 

)) [28] was introduced into the 

n 1 n 

n 

heuristic equation of the CAS algorithm for obtaining the 

chaotic search initially. The adjustment of the chaotic 

behaviour of individual ant is achieved by the 

introduction of a successively decrement of organization 

variable µ i and leads the individual to move to the new 

site acquired with the best fitness value eventually. 

( pid − θid 

) is introduced to achieve the information 

exchange of individuals and the movements to new site 

taken on the best fitness value. pid is selected based on 

the fitness theory which is very widely developed in 

optimization theory such as genetic algorithm and tabu 

search, and so on. θid is the state of the d th dimension 

of ant i . 

The CAS algorithm is a kind of iterative optimization 

algorithm, which is firstly employed in the optimization 

of sequential space. In the sequential space coordinates, 

the mathematic description [17] of the CAS algorithm as 

follows: 

⎧ 

( 1+ 

ri 

) 

µ i ( n) 

= µ i ( n −1) 

⎪ 

⎪ 

7. 

5 

⎪ 

θid 

( n) 

= ( θid 

( n −1) 

+ × Vi 

) × 

⎪ 

ψ id 

⎨ 

7. 

5 

aµ 

i ( n)( 

3−ψ 

id ( θid 

( n−1) 

+ × Vi 

)) 

ψ 

⎪ ( 1−e 

id 7. 

5 

⎪ 

e 

− × Vi 

+ 

ψ 

⎪ 

id 

⎪ ( −2aµ 

i ( n) 

+ δ ) 

⎩e 

( pid 

( n −1) 

−θid 

( n −1)) 

(10) 

where i = 1, 2, 

L, 

N , N is the size of the ant swarm; 

d = 1, 2, 

L, 

L , L is the dimension of the optimization 

space; n means the current iteration, and n −1 

is the 

previous iteration; µ i is the current state of the ith ant’s 

organization variable, µ i( 

0) 

= 0. 

999 ; ri is termed by us 

as the organization factor of ant i ; ψ id determines the 

selection of the search range of the dth element of 

variable in the search space; V i determines the search 

region of ant i and offers the advantage that ants could 


search diverse regions of the problem space. The value of 

Vi should be suitably selected according to concrete 

optimization problems; a is a sufficiently large positive 

constant and can be selected as a = 200 ; 

δ ( 0 ≤ δ ≤ 2 / 3) 

is a constant; pid ( n −1) 

is the best 

position found by the ith ant and its neighbors within 

n −1 

steps; θid is the current state of the dth dimension 

of ant i , θ id ( 0) 

= ( 7. 

5/ 

ψ id )( 1−Vi 

) R , where R is a 

uniformly distributed random number in R ∈[ 

0, 

1] 

. 

r i and ψ id are two important parameters. r i is the 

organization factor of ant i , which affects the 

convergence speed of the CAS algorithm directly. If r i is 

very large, the iteration step of ‘‘chaotic” search is small 

then the system converges quickly and the desired optima 

or near-optima cannot be achieved. If r i is very small, the 

iteration step of ‘‘chaotic” search is large then the system 

converges slowly and the runtime will be longer. Since 

small changes are desired as iteration step evolves, the 

value of r i is chosen typically as 5 . 0 0 ≤ < r i . The format 

of r i can be designed according to concrete problems and 

runtime. Each ant could have different r i , such 

as ri = 0. 3 + 0. 

02⋅ 

rand( 

1) 

. ψ id affects the search ranges 

of the CAS algorithm. If the interval of the search is 

ωid 

ωid 

[ − , ] , then we can obtain an approximate formula 

2 2 

7. 

5 

ω = . 

id 

ψ id 

In principle, a neighborhood can be any ordered finite 

set. These neighbors are not necessarily individuals who 

are near them in the parameter space, but rather ones that 

are near them in a topological space. In fact the CAS 

algorithm does not impose any limitation on the 

definition of the distance between two ants. In order to 

simulate the behaviors of ants, we use the Euclidian 

distance. Supposing there are two ants whose positions 

are θ , L, θ ) and θ , L , θ ) , respectively, where 

( i1 

iL 

( j1 

jL 

− L ¢ 

i , j = 1, 

2, 

L, 

N (where, N is the size of ant swarm) 

and i ≠ j , the distance between the two ants is 

2 ( θi1 2 

θ j1) 

+ 

2 

+ ( θiL 

−θ 

jL) 

In the CAS algorithm, the neighbor selection can be 

defined in two ways. The first is the nearest fixed number 

of neighbors. The nearest m ants are selected as the 

neighbors of single ant. The second way is to consider the 

situation in which the number of neighbors increasing 

with iterative steps. This is due to the influence of selforganization 

behavior of ant i . The impact of 

organization will become stronger than before and the 

neighbor of the ant will increase. That is to say, the 

number of nearest neighbor is dynamically changed as 

time evolves or iterative steps increase. The number q of


single ant is defined to increase for every T iterative 

steps. 

B. Parameters Optimization of LS-SVM Based on CAS 

Algorithm 

As stated before, the CAS algorithm has powerful 

global search ability to find exact or approximate 

solutions for optimization and search problems. Thus, a 

parameters selection approach using the CAS algorithm 

for LS-SVM is presented in this paper. There are two key 

factors to determine the optimized parameters using the 

CAS algorithm: one is how to represent the parameters as 

the ant’s position, namely how to encode. Another is how 

to define the fitness function which evaluates the 

goodness of an ant. These two key factors are given as 

follows: 

Encoding parameters: the optimized parameters for 

LS-SVM include kernel parameter and regularization 

parameter. In solving parameters selection by the CAS 

algorithm, each ant is requested to represent a potential 

solution, namely parameters combination. So let us 

denote an m -parameters combination as a vector of 

dimension m . For example, if Gauss radial basis function 

(RBF) is chosen as a kernel function, we denote the 

vector as v = ( γ , σ ) . 

Fitness function: the fitness function is generalization 

performance measure. There are some different 

descriptions for the generalization performance measure. 

Therefore, the corresponding fitness can be determined. 

The fitness of an ant is evaluated by the mean square 

error (MSE) index, which is defined as the error between 

the function estimation of LS-SVM and the reference 

model. It can be expressed by 

N 

1 

N ∑ 

i= 

1 

( y − f ( x)) 

where N denotes the number of training data, y is the 

reference model, and f (x) 

is the function estimation of 

LS-SVM. 

In the CAS algorithm one aims at minimizing the MSE 

through choosing the optimal parameters combination, 

that is 

2 

f (z , L, 

z ) = minMSE (11) 

min 1 i 

subject to the equality constraints 

gi i i 

≤ z ≤ h , i = 1, 

2 

where the optimization variables are γ and 

σ respectively, [ i, i ] h g denotes the value range for each 

variable, which is different with different reference model 

and training data. 

The flowchart of the CAS-based parameters selection 

algorithm for the LS-SVM is shown in Fig. 1. 


IV. SIMULATION RESEARCH 

Experiment of a typical test function estimation is 

performed to evaluate the performance of the proposed 

parameters selection method. All experiments are 

performed on a PC with Pentium IV 2.93GHz 

processor, 512MB of main memory and the Matlab 6.5 

simulation software. 

Given one-dimensional Sinc function 

f ( x) 

= sinc( x) 

+ v, 

x ∈[-3,3] 

(12) 

where v is the Gaussian noise with zero mean and 

standard deviation 0.1. We select 100 pairs of data as the 

train set from the input variable range. One aims at 

minimizing the MSE via the CAS algorithm to select the 

optimal kernel parameter σ of Gauss RBF kernel 

function and regularization itemγ .The searching ranges 

are set as follows: γ ∈[ 

0, 

30] 

, σ ∈[ 

0, 

5] 

. The CAS 

algorithm parameters are chosen as follows: N = 20 , the 

maximum number of iterations is 200, 

δ = 2/ 

3 , a = 200 , = 0. 05 + 0. 

02× 

rand() 

, ψ 0. 

25 

, 

r i 

1 = i 

ψ 1. 

5 . In simulation, the first way is used to select 

2 = i 

neighbours of single ant. The researching results of 

parameters are γ = 7. 

7379 and σ = 0.8851 

, respectively. 

The training result for LS-SVM via the above parameters 

is shown in Fig. 2. It can be seen from Fig.2 that LS- 

SVM realizes very good function approximation, so the 

CAS algorithm successfully realizes the parameters 

optimization selection for the test function.


Figure 2. Simulation result of sin c function 

In order to explain the effectiveness of this method, we 

adopt the genetic algorithm (GA, crossover rate is 0.8, 

mutation rate is 0.2%, population size is 30, the 

maximum number of iterations is 200) and particles 

swarm optimization algorithm (PSO, the population size 

and maximum number of iterations is the same as GA) to 

carry out many times’ experiments. The model of LS- 

SVM is tested with the testing set about 50 data produced 

by randomly initialized, the average results is recorded in 

Table 1. Table 1 shows the model testing MSE of this 

paper method is the minimum. 

Selection method Testing error 

GA 

PSO 

CAS 

TABLE I. 

…... Simulation 

— Real 

····Training data 

AVERAGE RESULTS OF PARAMETERS OPTIMIZATION OF LS-SVM 

BY DIFFERENT METHODS 

7.9057×10 -4 

5.5782×10 -4 

3.1550×10 -4 

V. APPLICATION OF LS_SVM BASED ON THE CAS 

ALGORITHM 

The inverted pendulum artificially created is a complex 

nonlinear system in order to deeply research the control 

for the nonlinear, high order and under-actuated system. 

Characterized as a typical nonlinear, high order, unstable 

and under-actuated system, it is very difficult to give a 

precise mathematical model. Therefore, the model 

identification research for the inverted pendulum system 

is very important. 

The GPIP2003 single planar inverted pendulum is 

considered as a plant in the paper, whose inverse model is 

identified by LS-SVM. We adopt the example provided 

by the inverted pendulum toolbox, where the pendulum is 

displaced from lower position to the upper. After the 

pendulum reaches the upper position, one applies the 

disturbance by plucking the pendulum. The experiment 

data from the process overcoming the disturbance to the 


stabilization is sampled to the workspace of Matlab 

environment by the communication interface, which is 

stored as the text document by the command “save~”. 

The data includes seven items such as the sampling period, 

the control variable, angle of the pendulum, position of 

the cart, angular rate of the pendulum, velocity of the cart 

and displacement of the objective. Angle of the pendulum, 

angular rate of the pendulum, position of the cart, velocity 

of the cart and the control variable are selected as multiinput 

and single-output model for LS-SVM. 100 pairs 

data from the input variable are chosen as the training 

sample set, in which 40 pairs data are selected as the 

testing sample set. The minimum MSE error as the fitness 

function, we utilize the CAS algorithm to carry the 

optimization selection for the regularization item γ and 

kernel parameter σ . The researching results of 

parameters are γ = 7. 

7379 and σ = 0.8851 

, and the 

testing error is 0.0022. The estimation for the inverse 

model is achieved using the above result. The simulation 

result is shown in Fig. 3. Fig. 3 shows that the estimation 

value approaches to the real sampling value. simulation 

results show the LS-SVM model has good generalization 

performance and stronger robust performance after 

optimized by the CAS algorithm. 


…... Simulation 

— Real 

Figure 3. Simulation result of inverse modeling 

for the inverted pendulum 

Appropriate parameters are very crucial to leastsquares 

support vector machines (LS-SVM) learning 

results and generalization ability. This paper presents a 

novel parameter selection method for LS-SVM is 

presented based on chaotic ant swarm (CAS) algorithm. 

The selection problem of LS-SVM parameters is 

considered as a swarm intelligence optimization problem, 

and a CAS optimization algorithm is employed to search 

the optimal objective function. CAS algorithm is global 

search method and it need not to consider LS-SVM 

dimensionality and complexity. Simulation and 

experiment results show that the proposed method is an 

effective approach for parameter optimization.



The authors are grateful to the anonymous referees for 

their valuable remarks and helpful suggestions, which 

have significantly improved the paper. This work was 

supported in part by a grant from the support of Key 

Project of Chinese National Programs for Fundamental 

Research and Development (973 Program) (2007CB7140 

06), National Nature Science Foundation of China 

(61074020) and the Fundamental Research Funds for the 

Central Universities (DC10040101). 

REFERENCES 

[1] V. N. Vapnik, The Nature of Statistical Learning Theory, 

chapter 6, New York: Springer-Verlag, 1995. 

[2] V. N. Vapnik, “An Overview of Statistical Learning 

Theory,” IEEE Trans on Neural Networks, vol. 10, p. 988- 

999, 1999. 

[3] J. A. K.Suykens, “Support vector machines: a nonlinear 

modeling and control perspective,” European Journal of 

Control, vol. 7, p. 311-327, 2001. 

[4] J. A. K.Suykens, “Nonlinear Modeling and Support Vector 

Machines (Published Conference Proceedings style),” 

Proc. of 18th Annu. IEEE Conference on Instrumentation 

and Measurement Technology, Budapest, p. 287-294, 

2001. 

[5] K. S. Chua, “Efficient computations for large least square 

support vector machine classifiers,” Pattern Recognition 

Letters, vol. 24, p. 75-80, 2003. 

[6] I. Goethals, K. Pelckmans, J. A. K. Suykens and B. De 

Moor, “Identification of MIMO Hammerstein models 

using least squares support vector machines,” Automatica, 

vol. 41, p. 1263-1272, 2005. 

[7] L. Bako, G. Mercere, S. Lecoeuche and M. Lovera, 

“Recursive subspace identification of Hammerstein models 

based on least squares support vector machines,” IET 

Control Theory and Application, vol. 3, p. 1209-1216, 

2009. 

[8] L. K. Hou, Q. X. Yang and J. L. An, “Modeling of SRM 

Based on XS-LSSVR Optimized by GDS,” IEEE 

Transactions on Applied Superconductivity, vol. 20, p. 

1102-1105, 2010. 

[9] Z. J. Li, Y. N. Zhang and Y. P. Yang, “Support vector 

machine optimal control for mobile wheeled inverted 

pendulums with unmodelled dynamics,” Neurocomputing, 

vol. 73, p. 2773-2782, 2010. 

[10] N. E. Ayat, M. Cheriet and C. Y. Suen, “Automatic model 

selection for the optimization of SVM kernels,” Pattern 

Recognition, vol. 38, p. 1733-1745, 2005. 

[11] Y. W. Kang, J. Li, G. Y. Cao, H. Y. Tu, J. Li and J. Yang, 

“Dynamic temperature modeling of an SOFC using least 

squares support vector machines,” Journal of Power 

sources, vol. 179, p. 683-692, 2008. 

[12] P. F. Pai and W. C. Hong, “Support vector machines with 

simulated annealing algorithms in electricity load 

forecasting,” Energy Conversion andManagement, vol. 46, 

p. 2669-2688, 2005. 

[13] X. L. Tang, L. Zhang, J. Cai and C. B. Li, “Multi-fault 

classification based on support vector machine trained by 

chaos particle swarm optimization,” Knowledge-Based 

Systems, vol. 23, p. 486-490, 2010. 

[14] S. J. An, W. Q. Liu and S. Venkatesh, “Fast crossvalidation 

algorithms for least squares support vector 

machines and kernel ridge regression,” Pattern 

Recognition, vol. 40, p. 2154-2162, 2007. 


[15] W. T. Mao, G. R. Yan, L. L. Dong and D. Hu, “Model 

selection for least squares support vector regression based 

on small-world strategy,” Expert Systems with Applications, 

vol. 38, p. 3227-3237, 2011. 

[16] T. V. Gestel, J. A. K. Suykens and B. Baesens, S. Viaene, 

J. Vanthienen and G. Dedene, et al., “Benchmarking least 

squares support vector machine classifiers,” Machine 

Learning, vol. 54, p. 5-32, 2004. 

[17] L. X. Li, Y. X. Yang, H. P. Peng and X. D. Wang, 

“Parameters identification of chaotic systems via chaotic 

ant swarm,” Chaos, Solitons and Fractals, vol. 28, p. 

1204–1211, 2006. 

[18] L. X. Li, Y. X. Yang, H. P. Peng and X. D. Wang, “An 

optimization method inspired by chaotic ant havior,” 

International Journal of Bifurcation Chaos, vol. 16, p. 

2351-2364, 2006. 

[19] J. J. Cai, X. Q. Ma, L. X. Li, Y. X. Yang, H. P. Peng and X. 

D. Wang, “Chaotic ant swarm optimization to economic 

dispatch,” Electric Power Systems Research, vol. 77, p. 

1373-1380, 2007. 

[20] L. X. Li, Y. X. Yang and, H. P. Peng, “Computation of 

multiple global optima through chaotic ant swarm,” Chaos, 

Solitons and Fractals, Vol. 40 (2009), p. 1399-1407. 

[21] Y. G. Tang, M. Y. Cui, Li L. X., H. P. Peng and X. P. 

Guan, “Parameter identification of time-delay chaotic 

system using chaotic ant swarm,” Chaos, Solitons and 

Fractals, vol. 41, p. 2097-2102, 2009. 

[22] L. X. Li, Y. X. Yang and, H. P. Peng, “Fuzzy system 

identification through chaotic ant swarm,” Chaos, Solitons 

and Fractals, vol. 41, p. 401-408, 2009. 

[23] H. Zhu, L. X. Li, Y. Zhao, Y. Guo and Y. X. Yang, “CAS 

algorithm-based optimum design of PID controller in 

AVR system,” Chaos, Solitons and Fractals, vol. 42, p. 

792-800, 2009. 

[24] Y. Y. Li, Q. Y. Wen, L. X. Li and H. P. Peng, “Hybrid 

chaotic ant swarm optimization,” Chaos, Solitons and 

Fractals, vol. 42, p. 880-889, 2009. 

[25] W. C. Hong, “Application of chaotic ant swarm 

optimization in electric load forecasting,” Energy Policy, 

vol. 38, p. 5830-5839, 2010. 

[26] A. Chatterjee, S. P. Ghoshal and V. Mukherjee, “Chaotic 

ant swarm optimization for fuzzy-based tuning of power 

system stabilizer”, Electrical Power and Energy Systems, 

in press. 

[27] B. J.Cole, “Is animal behavior chaotic? Evidence from the 

activity of ants.” Proc R Soc Lond Ser B biol Sco, vol. 244, 

p. 253-259, 1991. 

[28] R. V. Solé, O. Miramontes and B. C. Goodwill, 

“Oscillations and chaos in ant societies,” Journal of Theory 

Biology, vol. 161, p. 343-357, 1993. 

Chunli Xie Xie received his B.Sc. and M.Sc. degrees from 

Fushun Petroleum Institute and Liaoning Shihua University, 

Fushun, China, in 1995 and 2003, respectively. He is currently 

working toward the Ph.D degree with Dalian University of 

Technology, Dalian, China. 

His research interests include adaptive control, robust control, 

machine learning, nonlinear systems, artificial intelligence and 

application. 

Cheng Shao was born in Shenyang, P. R. China, on June 7, 

1958. Shao received his B.Sc. degree from Liaoning University, 

Shenyang, China, in 1982. Then he received the M.Sc and Ph.D. 

degrees from Northeastern University, Shenyang, China, in 

1987 and 1992. 

He is currently a full-time Professor and Ph.D. Advisor with 

School of Electronic and Information Engineering, Dalian


University of Technology, China. His research interest covers 

complex system modeling and intelligence control. 

Dandan Zhao was born in Fuxin, P. R. China, on March 4, 

1975. Zhao received her B.Sc. and M.Sc. degrees from Fushun 

Petroleum Institute and Liaoning Shihua University, Fushun, 


China, in 1997 and 2003, respectively. She is currently a fulltime 

lecturer of School of Computer Science and Engineering, 

Dalian Nationalities University, Dalian, China. Her research 

interests include electronic commerce, semantic network, 

swarm intelligent and information processing.


The Expected Value Model of Multiobjective 

Programming and its Solution Method Based on 


Mingfa Zheng 

College of Science, Air Force Engineering University, Xi'an, Shanxi, 710051, China 

Email: mingfazheng@126.com 

Bingjie Li and Guangxing Kou 

College of Science, Air Force Engineering University, Xi'an, Shanxi, 710051, China 

Email: {mingfa103, kouguangx }@ 163.com 

Abstract—In this paper, based on bifuzzy theory, we study 

the multiobjective programming problem under bifuzzy 

environment and present the expected value model to the 

problem. Furthermore, to the proposed model, the concepts 

of non-inferior solution are defined, and their relations are 

also discussed. According to practical decision-making 

process, a solution method, called the method of main 

objective function, has been studied, whose results can 

facilitate us to design algorithms to solve the bifuzzy 

multiobjective programming problem. Finally, a numerical 

example is given to explain the proposed method. 

Index Terms—credibility theory, bifuzzy variable, 

multiobjective programming, expected value model 


The multiobjective programming problems are studied 

by many researchers such as [3], [15], [16],[ [18]],[ [20]. 

For given multiobjective problem, its absolute optimal 

solutions which optimize each objective functions 

simultaneously usually don not exist, so we consider their 

non-inferior solutions, which are Pareto optimal solutions 

in common use. There are various types of uncertainties 

in the real-world problem. As we known, random 

phenomena is one class of uncertain phenomena which 

has been well studied. Based on the probability, 

stochastic multiobjective programming problems have 

been presented such as [1], [17]. Besides randomness, 

fuzziness is a basic type of subjective uncertainty 

initiated by [26]. Since the pioneering work of Zadeh, 

possibility theory was developed and extended by many 

researchers such as [2],[4],[7],[23],[21],[24]. Based on 

possibility theory, an axiomatic approach, called 

credibility theory [6], was studied extensively. From a 

measure-theoretic viewpoint, credibility theory provides a 

theoretical foundation for fuzzy programming [9] just like 

the role of probability theory in stochastic programming 

[5]. In a practical decision-making process, we often face 

a hybrid uncertain environment where linguistic and 

frequent nature coexist. For the examples of two fold 


doi:10.4304/jcp.6.9.1942-1948 

uncertainty, we may refer to Liu [6], Liu [8], [10], 

Liu[11], Liu and Liu [13], Yazenin[22], Zhou[25]. To 

deal with this two fold uncertainty, it is required to 

employ bifuzzy theory[7]. The multiobjective 

programming under bifuzzy environment has not been 

developed well, therefore, following the idea of stochastic 

multiobjective programming, this paper devotes the 

bifuzzy multiobjective programming (BMOP) problems 

based on the random fuzzy theory. For the parameters of 

bifuzzy, we consider their expectation which convert the 

BMOP problem into the expected value model of bifuzzy 

multiobjective (EVBMOP) which is a deterministic 

multiobjective problem. By the deterministic problem 

above, we can obtain the expected value efficient 

solutions or expected value weakly efficient solutions to 

the BMOP problem. In actual problem, we usually need 

to distinguish between primary and secondary of the 

objective functions to the BMOP problem, so the method 

of main objective function is presented to solve the 

BMOP problem in this paper, which can covert the 

EVBMOP problems corresponding to the BMOP 

problem into the deterministic single objective 

programming problems whose optimal solutions are 

expected value weakly efficient solutions to the BMOP 

problems. 

This paper is organized as follows. The next section 

provides a brief review on the related concepts and results 

in bifuzzy theory. Section 3 presents the BMOP problem 

and its expected value model. Furthermore, based on the 

expected value model, the expected value efficient 

solution and expected value weakly efficient solution to 

the BMOP are proposed, and their properties are 

discussed. To solve the BMOP problem, the method of 

main objective function is introduced in Section 4. 

Finally, Section 5 provides a summary of the main results 

of this paper.


II. BASIC CONCEPTS 

Given a universe Γ , ρ( Γ) 

is the power set of Γ , and 

a set function Pos defined on ρ( Γ) 

is called a possibility 

measure if it satisfies the following conditions[4]: 

(Pos1) P os( φ ) = 0, P os( 

Γ) 

= 1, and 

(Pos2) P os( ∪ i∈I Ai) = supi∈I P os( Ai) 

for any subclass 

{ Aii∈I}of ρ( Γ). 

The triplet ( Γ, ρ( 

Γ), 

Cr) 

is usually called a possibility 

space, which is called a pattern space by Nahimias [19]. 

In addition, a self-dual set function, called credibility 

measure, is defined as follows [12]: 

1 

c 

C r( A) = (1 + P os( A) − P os( A )). 

2 

for any A ∈ ρ( 

Γ), 

where A c is the complement of A . 

A fuzzy variable X is defined as a function from a 

credibility space ( Γ, ρ( 

Γ), 

Cr) 

to the set of real numbers. 

Based on credibility measure, the expected value of 

fuzzy variable X is defined as [12] 

∞ 

∫ ∫−∞ 0 

0 

E[ ξ] = Cr( ξ ≥ r) dr − Cr( ξ ≤ r) dr (1) 

provided that one of the two integrals is finite. 

Given a credibility space ( Γ, ρ( 

Γ), 

Cr) 

, which is 

complete, we obtain the definition of bifuzzy variable as 

follows: 

Definitions 2.1.[7] Let ( Γ, ρ( 

Γ), 

Cr) 

be a credibility 

space. A map ( 1, 2, 

, 

) : v is said to be 

an bifuzzy vector if for any Borel subset B of 

the function 

T 

n 

ξ = ξ ξ ξn 

T → F 

n−ary n 

' 

R , C{ r γ ∈Γ ξγ( γ ) ∈B} 

is measurable 

with respect to γ . As n = 1, ξ is called a bifuzzy 

variable. 

Definitions 2.2.[7] Suppose ξ is a bifuzzy variable, 

the expected value of ξ is defined as the mathematical 

expectation of the fuzzy variable E[ ξγ 

], i.e., 

E( ξ) = ∫ E[ ξγ] Crd( 

γ) 

(2) 

Γ 

provided that the integrand E[ ξ γ ] defined by Eq.(1) 

exists almost surely with respected to γ , and is integral. 

From Eq.(2), we can provide the expectation of 

bifuzzy variable, i.e., 

E( ξ ) = Eγ[ Eγ '[ 

ξγ( γ ')]]. 

Ⅲ BIFUZZY MULTIOBJECTIVE PROGRAMMING PROBLEMS 

A. Expected Value Model of Bifuzzy Multiobjective 

Programming 

Considering the bifuzzy multiobjective programming 

(BMOP) problem as follows: 

⎧min 

F( x, ξ) = ( f1( x, ξ), f2( x, ξ), ⋅⋅⋅, 

fj( x, 

ξ)) 

x∈R ⎪ 

(BMOP) ⎨ st . . G( x, ξ) = ( g1( x, ξ), g2( x, ξ), ⋅⋅⋅ , gn( x, 

ξ)) 


Similarly, by the linear properties of fuzzy variable, we 

can obtain: 

G( λx + (1 −λ) 

x , ξ ) 

1 2 

≤ λGx ( , ξ ) + (1 −λ) 

Gx ( , ξ ) 

1 γ 2 γ 

Using the same method above, we can obtain: 

EG [ ( λx+ (1 −λ) 

x , ξ)] 

1 2 

≤ λEG [ ( x1, ξ)] + (1 − λ) EGx [ ( 2, 

ξ)] 


F( x, ξ ) is strict convex vector function on D , and is 

also conmonotonic, by noting the inequality just given, it 

easy to know that 

* 

EF [ ( αx+ (1 −α) 

x, 

ξ)] 

* 

< α EF [ ( x, ξ)] + (1 −α) 

EF [ ( x, 

ξ)] 

* 

< EF [ ( x, 

ξ )], 

* 

which is a contradiction with x ∈ D . Thus, 

Dpa ⊃ Dwpa 

, 

wpa 

which proves the required theorem. 

Ⅳ SOLUTION METHOD 

A. Expected Value Model of Bifuzzy Multiobjective 

Programming 

In real-world problems, we just need to consider the 

main objective function to some real-life problems, 

therefore, a type of method, called the method of main 

objective function, is presented in the following. Without 

any loss of generality, let f1( x, ξ ) be regarded as main 

objective function to the BMOP problem, and wish the 

expectation of the other objective functions fj( x, ξ ) , 

j=2,3,…,p, satisfy the following constraint-conditions: 

E[ fj( x, ξ )] ≤ αi 

, j=2,3,…,p. Then the BMOP problem 

can be transformed into the following SOP problem 

where 

min E[ f1( x, 

ξ )] (8) 

x∈ D 

D 

= { x∈ D E[ f ( x, ξ)] ≤ α , j = 2, 3,..., p}, 

j i 

whose optimal solution set is denoted as D sab . 

Obviously, the constraint set D to problem (8) is a 

new set which is added into several constraint condition 

E[ fj( x, ξ )] ≤ αi, 

j = 2, 3, ⋅⋅⋅, 

p. 

Then we employ the 

method of solving the nonlinear programming which is a 

linear problem in particular to solve the transformed SOP 

problem whose optimal solution is the non-inferior 

solution to the BMOP problem verified by the following 

theorem. 

D ⊂ D 

Theorem 4.1. sab wpa 

* * 

Proof. If x ∈ Dsab 

, and x ∉ Dwpa 

, then, by the 

definition of expected value weakly efficient solution, 

there 

must exist some x∈ D, 

such that 

* 

E[ fj( x, ξ )] < E[ fj( x , ξ )] for all j=1,2,...,p. 

* 

Since x ∈ D, 

we have 

* 

E[ fj( x , ξ)] ≤ α j, 

j = 2,3, ⋅⋅⋅, 

p. 

It follows from inequality above that 

* 

E[ f j( x, ξ)] ≤ E[ f j( x , ξ)] ≤ α j, 

j = 2,3, ⋅⋅⋅, 

p, 

which illuminates x∈ D, 

i.e., x is the feasible solution 

to SOP problem, therefore, it is easy to know 

* 

E[ f ( x, ξ )] < E[ f ( x , ξ )] , 

j j 


which is a contraction with by the previous hypothesis 

* * 

that x ∈ Dsab 

. Hence, x ∈ Dwpa 

, which implies the 

required conclusion. 

Theorem 4.2. Without any loss of generality, assuming 

that f1( x, ξ ) is the main objective function, if H( x, ξ ) 

is linear vector function, F( x, ξ ) and G( x, ξ ) are strict 

convex vector function on x . Furthermore, for any 

given x1 and x2 , F( x1, t) 

and F( x2, t) 

(correspondingly, G( x 1, 

t) 

and G( x2, t) 

are comonotonic 

on t , then Dsab ⊂ Dpa 

. In addition, if Dab ≠ φ, 

we can 

obtain: 

Dsab ⊂ Dab 

. 

Proof. It follows from the assumed conditions that 

E[ f1( x, ξ )] is strict convex function, so the optimal 

* 

solution to SOP problem must be unique. If x is the 

unique optimal solution to the SOP problem, and 

* 

x ∉ D , there must exist x∈ D and x ≠ x 

* such that 

pa 

E[F( x, ξ )] ≤ E[F( x , ξ )] , 

i.e., 

* 

E[ fj( x, ξ)] ≤ E[ fj( x , ξ)], 

j = 1,2,3, ⋅⋅⋅, 

p. 

Obviously, there must exist some j0(1 =< j0 


the α i according to the actual demand. Furthermore, if 

the α i is not well-found, then the feasible sets D may 

be empty set, which can't get the optimal solution of SOP 

problem, that is, we can’t obtain the expected value noninferior 

solutions to the BMOP problem, so we can take 

the following measure which can avoid that D is empty 

set: 

0 α j=E[ fj( x , ξ )], j = 2,3, ⋅⋅⋅, 

p, 

0 

for any given x ∈ D, 

which can guarantee that one 

0 

solution at least, i.e., there exist x ∈ Dat 

least. 

Furthermore, the optimal solutions of the SOP problem 

by the measure proposed above must be the expected 

value weakly efficient solution of the BMOP problem, 

and it may be not satisfying, but it is the practical 

technique to deal with real-life problem frequently. 

B. Expected Value Model of Bifuzzy Multiobjective 

Programming 

In particular, if the bifuzzy variable ξ involved in the 

problem (8) is a discrete one, we will illuminate how to 

calculate the E[ f1( x, ξ )]. Assume that the bifuzzy 

variable ξ is a discrete one such that γ is a discrete 

fuzzy variable taking on finite number of values with 

possibility μ i , i = 1, 3, ⋅⋅⋅, 

N , respectively, and 

N 

satisfying maxi= 1 μi 

= 1, j = 1, 3, ⋅⋅⋅, 

N , and for each i , 

fuzzy variable ξ taking on the following values 

γ 

γ i 

ξ ( γ ) with possibility μ > 0; 

1 

' 

11 

' 

12 

ξγ ( γ 1 

…… 

) with possibility μ 12 > 0; 

' ξγ( γ 1 1N 

) with possibility μ 1 

1N1 

> 0; 

ξ ' 

( γ ) with possibility μ > 0; 

γ 

2 

21 

' 

22 

ξγ ( γ 2 

…… 

) with possibility μ 22 > 0; 

' ξγ( γ 2 2N 

) 2 

with possibility μ 2N 

2 

…… 

> 0; 

ξγ γ i i 

' ( 1 ) with possibility μ i1 

> 0; 

ξγ γ i i 

' ( 2 ) with possibility μ i2 

> 0; 

…… 

ξγγ i iN i 

' 

( ) with possibility μ iNi 

…… 

> 0; 

N N ' ξγ( γ 1) 

with possibility μ N1 

> 0; 

ξγγ N N 

' ( 2 ) with possibility μ N 2 > 0; 

…… 

ξγγ NN 

' 

( ) with possibility μ iN N > 0; 

N N 

It is easy to obtain the expectation of fuzzy variable 

f1( x, ξγ( γ ')) as follows: 


11 

21 

Ni 

' 

1 ξγ = γ ' 1 ξγ γ =∑ij 

1 ξγ γ i ij 

j= 

1 

f ( x, ) E [ f ( x, ( '))] p f ( x, 

( )) (9) 

where pij are the weights of fuzzy 

variable f x ξγγ i ij calculated by the following 

formulas [14]: 

' 

1( , ( )) 

j j−1 Ni Ni+ 

1 

1 1 

pij = (max μik − max μik ) + (max μik − max μik 

) 

2 k= 1 k= 0 2 k= j k= j+ 

1 

(10) 

where μi0= μ iNi+ 1 = 0, i = 1, 2, ⋅⋅⋅ , N, j = 1, 2, ⋅⋅⋅, 

Ni, 

and satisfies the following constrains: 

p ≥ 0, 

Ni + 1 

p 

Ni 

+ 1 = max μ = 1. 

∑ 

ij j ij j ij 

By the Eq.(2), the expectation of bifuzzy variable 

f1( x, ξγ 

) are given in the following 

E[ f ( x, ξ)] = E [ f ( x, ξ )] =∑p 

f ( x, 

ξ ) (11) 

1 γ 1 γ i 1 γ 

i= 

1 

where pi 

are the weights of fuzzy variable 

f1( x, ξγ 

) calculated similarly by the Eq.(10). 

Example 4.1. Solving the following bifuzzy 

multiobjective programming 

⎧ min( f1( x, ξ), f2( x, 

ξ)) 

x 

⎪ 

⎪ = (5x1 + 7x2 − 4x3 + 2 ξ, − 2x1 + 3x2 + 8x3 −3 

ξ) 

⎪ 

⎨ st .. x1 + 2x2 −3x3 ≤ 5 

⎪ 

−7x − x + x ≤ 

⎪ 1 3 2 7 3 3 

⎪ 11x1 + 5x2 −6x3 ≤ 10 

⎩ 

(12) 

where f1( x, ξ ) is the main objective function, and the 

limit value of E[ f2( x, ξ )] is 4.4, i.e., 

E[ f2( x, ξ )] ≤ α2 

= 4.4. 

In addition, ξ is the discrete bifuzzy variable defined 

as 

⎧ X , with possibility 3/5 

ξγ 

1 

⎪ 

= ⎨ X2 , with possibility 1/4 

⎪⎩ X , with possibility 1 . 

3 

Here the fuzzy variable X 1 assumes the value 3, 4, 

5with the possibility 1/4, 3/4, and 1, respectively; X 2 

assumes the value 1, 2, 3 with the possibility 5/12, 1 and 

7/12, respectively; and X 3 assumes the value 6, 8, 10 

with the possibility 2/7, 1/7 and 1, respectively. 

According to the method of main objective function 

discussed above, the problem (12) can be transformed 

into the following single objective problem 

⎧ min E[( f1( x, ξ)] = E[5x1 + 7x2 − 4x3 + 2 ξ] 

x 

⎪ 

⎪ st .. E[ − 2x1 + 3x2 + 8x3 −3 ξ ] ≤ 4.4 

⎪ 

⎨ x1 + 2x2 −3x3 ≤ 5 (13) 

⎪ 

−7x − x + x ≤ 

⎪ 1 3 2 7 3 3 

⎪ 11x1 + 5x2 −6x3 ≤ 10. 

⎩ 

We can obtain the following results 

N


' 

f1( x, 

ξγ( γ 1 11))] 

' 

= 5x1+ 7x2− 4x3+ 6, f1( x, 

ξγ( γ 1 12))] 

' 

= 5x1+ 7x2− 4x3+ 8, f1( x, 

ξγ( γ 1 13))] 

= 5x1+ 7x2− 4x3+ 10. 

It is easy to know 

' ' 

' 

f1( x, ξγ ( γ f x f x 

1 11))] ≤ 1( , ξγ ( γ 1 12))] ≤ 1( , ξγ ( γ 1 13))]. 

Therefore, we can obtain the distribution function of 

' f1( x, ξγ( γ j 

fuzzy variable 

1 1 )) 

as 

⎧5x1+ 

7x2− 4x3+ 6, with possibility 1/ 4 

' ⎪ 

f1( x, ξγ( γ j = ⎨ x + x − x + with possibility 

1 1 )) 5 1 7 2 4 3 8, 3/4 

⎪ 

⎩5x1+ 

7x2− 4x3+ 10, with possibility 1 

with weights p11 = 1/ 8, p 12 = 1/ 4, and p13= 

5 / 8, 

respectively, which are calculated by Eq.(10). 

It follows the Eq.(9) that 

1 ξγ1 3 

= ∑ 

j= 

1 

ij 1 ξγ1 ' γ1j 

f ( x, ) p f ( x, 

( )) 

1 1 

= (5x1+ 7x2− 4x3+ 6) + (5x1+ 7x2− 4x3+ 8) 

8 4 

5 

+ (5x1+ 7x2− 4x3+ 10) 

8 

= 5x1+ 7x2− 4x3+ 9, 

whose possibility is 3/5. 

Similarly, we can obtain 

f1( x, ξγ 

) = 5x + x − x + 

2 1 7 2 4 3 25/6, 

f1( x, ξγ 

) = 5x + x − x + 

3 1 7 2 4 3 132/7, 

with the possibility 1/4 and 1, respectively. 

Obviously, 

' ' 

' 

f1( x, ξγ ( γ f x f x 

1 12))] ≤ 1( , ξγ ( γ 1 11))] ≤ 1( , ξγ ( γ 1 13))]. 

Hence, without any loss of generality, the distribution 

function of fuzzy variable f1( x, ξγ 

) is the following 

i 

⎧ 5x1+ 7x2− 4x3+ 25/6, with possibility 1/4 

' ⎪ 

f1( x, ξγ( γ j = ⎨ x + x − x + with possibility 

1 1 )) 5 1 7 2 4 3 9, 3/5 

⎪ 

⎩5x 

1+ 7x2− 4x3+ 132/ 7, with possibility 1 

with weights 

p11 = 1/ 8, p12= 7 / 40, and p13 

= 7 /10, 

respectively, which are calculated by Eq.(10). 

By the Eq.(11), we can deduce 

ξ = γ ξγ = ∑ i ξγi 

i= 

3 

1 1 1 

1 

E[ f ( x, )] E [ f ( x, )] p f ( x, 

) 

1 7 

= (5x1+ 7x2− 4x3+ 25/ 6) + (5x1+ 7x2− 4x3+ 9) 

8 40 

7 

+ (5x1+ 7x2− 4x3+ 132 / 7) 

10 

= 5x1+ 7x2− 4x3 + 20.953. 

Using the same method, we can obtain the expectation 

of the bifuzzy variable f2( x, ξ ) as follows 

1 ξ = γ 2 ξγ 3 

= ∑ 

i= 

1 

i 2 ξγi 

E[ f ( x, )] E [ f ( x, )] p f ( x, 

) 

=− 2x + 3x + 8x −23.25. 

1 2 3 

Therefore, problem (13) is equivalent to the following 

problem: 


⎧min 

5x1+ 7x2− 4x3 + 20.953 

x 

⎪ 

⎪ st . . − 2x1 + 3x2 + 8x3 −23.25 ≤ 4.4 

⎪ 

⎨ x1 + 2x2 −3x3 ≤ 5 (14) 

⎪ 

−7x − x + x ≤ 

⎪ 1 3 2 7 3 3 

⎪ 11x1 + 5x2 −6x3 ≤ 10, 

⎩ 

whose optimal solution is 

( x1, x2, x 3) = ( -0.4286, 0,0) 

solved by LINGO software. Furthermore, we can obtain 

* 

that x = ( -0.4286,0,0) 

is the expected value weakly 

efficient solution to problem (12) by the Theorem 4.1. 

Ⅴ CONCLUSIONS 

In this study, we mainly concerned the expected value 

model and the solution method of the multiobjective 

programming problem under bifuzzy environment. We 

first presented a new type of bifuzzy multiobjective 

programming problem. As we known, the non-inferior 

solutions play important role to multiobjective problem, 

so the expected value non-inferior solutions to the BMOP 

problem are presented and their relations are also studied. 

In addition, a solution method, called the method of main 

objective function, was discussed, which facilitated us to 

design algorithms to solve the BMOP problem. 

REFERENCES 

[1] F. Benabdelaziz, P. Lang and R. Nadeau, “Pointwise 

Efficiency in Multiobjective Stochastic Linear 

Programming,” Journal of Operational Research 

Society,vol.45, pp. 11-18, 2000. 

[2] G. De. Cooman, E.E. Kerre and F. Vanmassenhove, 

“Possibility theory: an Integral Theoretic Approach,” 

Fuzzy Sets Syst, vol. 46, pp. 287-299, 1992. 

[3] Y.D. Hu, “The Efficient Theory of Multiobjective 

Programming,” China: Shanghai Since and Technology 

Press, 1994. 

[4] G.J. Klir, “On fuzzy-set Interpretation of Possibility 

Theory. Fuzzy sets and Systems,” vol. 108, pp. 263-273, 

1999. 

[5] P. Kall and S.W. Wallace, “Stochastic Programming,” 

Chichester: Wiley, 1994. 

[6] B.D. Liu, Uncertainty theory, “An Introduction to its 

Axiomatic Foundations,”Germany: Springer-Verlag, 2004. 

[7] B.Liu, “Toward Fuzzy Optimization without Mathematical 

Ambiguity,”Fuzzy Optimization and Decision Making,” 

vol. 1, pp. 43-63, 2002. 

[8] B.Liu,“Fuzzy Random Dependent-Chance Programming,” 

IEEE Trans. Fuzzy Syst, vol. 9, pp. 721-726, 2001. 

[9] B.D.Liu,“Theory and Practice of Uncertainty Programming, 

Heidelberg, Physica-Verlag, 2002. 

[10] B. Liu, “Uncertain Programming,” New York: Wiley, 1999. 

[11] B. Liu, “ Random Fuzzy Dependent-Chance Programming 

and its Hybrid Intelligent Algorithm,” Information 

Sciences,” vol. 141, pp. 259-271,2002. 

[12] B. Liu and Y.K. Liu, “Expected Value of Fuzzy Variable 

and Fuzzy Expected Value Models,” IEEE Trans. Fuzzy 

Syst, vol. 10, pp. 445-450,2002 

[13] Y.K. Liu and B. Liu, “Expected Value Operator of 

Random Fuzzy Variable Operator,” International Journal


of Uncertainty, Fuzziness, Knowlledge-Based Systems, vol. 

11, pp. 195-295, 2003. 

[14] Y.K. Liu and S.Wang,“Theory of Fuzzy Random 

Optimization (China Agricultural University Press, Beijing 

2006). 

[15] C.Y. Lin and J.L. Dong, “The Efficient Theory and 

Method of Multiobjective Programming,” China: Jilin 

Educational Press, 2006. 

[16] B.J. Ma, “The Efficient Rate of Efficient Solution to Linear 

Multiobjective Programming,”Journal of Systems 

Engineering and Electronic Technology, vol. 2, pp. 68-106, 

2000. 

[17] I.M. Stancu-Minasian, “Stochastic Programming with 

Multiple Objective Functions,” Buckarest, 1984. 

[18] M.M. Munoz and F. Ruiz, “An interval reference pointbased 

method for stochastic multiobjective programming 

problems,” European Journal of Operational Research, 

vol. 197, pp. 25-35, 2009. 

[19] S. Nahmias, “Fuzzy variables,” Fuzzy Sets and 

Systems.,vol. 1, pp. 97-101, 1978. 

[20] H.C.Wu, “Solutions of Fuzzy Multiobjective Programming 

Problems Based on the Concept of Scalarization,” J Optim 

Theory Appl, vol. 139, pp. 361-378, 2008. 


[21] Z. Wang, and J. Klir, “Fuzzy Measure Theory, New York: 

Plenum Press, 1992. 

[22] A.V. Yazenin, “Fuzzy and Stochastic Programming,” 

Fuzzy Sets Syst, vol. 22, pp. 171-180, 1987. 

[23] R. R.Yager, “A Foundation for a Theory of Possibility,” 

Journal of Cybernetics, vol.10, pp. 177-204, 1980. 

[24] Q. Zhang, Shigeya Maeda andToshihiko Kawachi, 

“Stochastic multiobjective optimization model for 

allocating irrigation water to paddy fields,” Paddy Water 

Environ, vol. 5, pp. 93-99, 2007. 

[25] J. Zhou and B. Liu, “Analysis and Algorithms of Bifuzzy 

Systems,” International Journal of Uncertainty. Fuzziness 

and Knowledge-Based Systems, vol. 12, pp. 357-376, 2004. 

[26] L.A. Zadeh, “Fuzzy Sets as a Basis for a Theory of 

Possibility,” Fuzzy Sets and Systems, vol. 1, pp. 3- 28, 

1978. 

.


A Method for Building Partially Connected 

Neural Network 

Gang Li 

Management Department, Shanghai University for Science and Technology, Shanghai, China 

Email: sdlig@163.com 

Xingsan Qian, Chunming Ye, 

Management Department, Shanghai University for Science and Technology, Shanghai, China 

Email: qxsqg@126.com, yechm6464@163.com 

Abstract - This paper focuses mainly on application of 

Partial Connected Back Propagation Neural Network 

(PCBP) instead of typical fully connected neural network 

(FCBP), as PCBP with less connections learns faster than 

FCBP. The initial neural network is fully connected, after 

training with sample data, a clustering method is employed 

to cluster weights between input to hidden layer and from 

hidden to output layer, and connections that are relatively 

unnecessary are deleted, thus the initial network becomes a 

PCBP network. PCBP can be used in prediction or data 

mining by training it with data that comes from database. 

At the end of this paper, several experiments are conducted 

to illustrate the effects of PCBP using the submersible pump 

repair data set. 

Index Terms - Neural Network; FCBP; PCBP; pruning 


Artificial neural networks have been proved to be a 

useful tool in pattern recognition and classification tasks 

in diverse areas like data mining, millions of databases 

are being used in business data management, scientific 

and engineering data management and other applications 

[1], and the most-widely used network is the standard 

Back Propagation (SBP) algorithm [2]. Indeed, the SBP 

learning algorithm has emerged as the standard algorithm 

for the training of multiplayer networks, and hence the 

one against which other learning algorithms are usually 

benchmarked. Actually, the SBP is fully connected, as 

called FCBP, and it has been commonly used as a matter 

of fact, since they usually do not need a priori 

information of data, of course, this is the feature of FCBP, 

but unfortunately, FCBP have several drawbacks, as 

reported by researchers [3]: it is extremely slow; training 

performance is sensitive to the initial conditions; it may 

become trapped in local minima before converging to a 

solution; oscillations may occur during learning (this 

usually happens when users increase the learning rate in 

an unfruitful attempt to speed up convergence); and, if 

the error function is shallow, the gradient is very small 


doi:10.4304/jcp.6.9.1949-1954 

Lin Zhao 

HP China, Shanghai, China 

Email: lemiozl@163.com 

leading to small weight changes. 

Also, as for FCBP, due to the learning style, the 

structure of the trained FCBP usually have unnecessary 

connections which induces the issue of the complexity of 

the networks and causes the slow training time, especially 

for large networks. The complexity problem has attracted 

the interest of researchers because of the advantages that 

would be obtained by solving it. One critical advantage is 

that the simpler the system, the better it is [4]. So, if these 

unnecessary connections can be removed from the 

network, then training times would be greatly reduced, it 

is especially important for data mining, where database 

usually contains large number of data records ranging 

from millions to even billions, without faster training 

time, data mining using neural network is mission 

impossible. 

One way to reduce the complexity of the networks is 

to reduce the number of redundant connections, nodes [5], 

or input features. The reduction of the connections or 

nodes can be achieved by removing the weights that 

contribute the least to the network outputs. To our best 

knowledge, most reduction methods have been done 

during training networks. And one important thing is to 

determine what kinds of connections are redundant? 

II. BUILDING A PCBP NETWORK 

As mentioned earlier, FCBP requires more training 

time than PCBP (see Fig1). 

Fig.1 Example of PCBPs


Generally, there are two ways to build a PCBP, one 

is manually; the other is automatic generated by starting 

from a FCBP and then prune FCBP to remove the 

unnecessary connections. The previous way is mandated 

and requires a deep insight into the data patterns involved, 

or else the network structure is not properly set, it may 

need more training time than FCBP; the latter way does 

not require user participated, and could determined the 

would-be removed connections automatically, the process 

is illustrated in Figure 2. 

Construct a FCBP 

Tr ai n FCBP wi t h 

sampl e dat a 

Pr une t he FCBP 

Sat i sf y 

accur acy? 

Yes 

The f i nal PCBP 

No 

Fig.2. Build a PCBP by pruning FCBP 

III. TRAINING FCBP 

Before training network, several things should be 

pre-defined, that is network structure including number of 

input and hidden and output nodes, generally speaking, 

numbers of input and output nodes depend on sample 

data, as for hidden nodes, it is usually determined by 

experience, some researchers have reported that a few 

number of hidden nodes is just enough. As for error 

function, the typical SBP employs Mean Squared Error 

(MSE) as follows: 

k o 1 

i i 2 

Error = ∑∑( 

S p − t p ) 

(1) 

2 i= 

1 p= 

1 

i 

Where S p stands for the actual output of output 

i 

node i, and tp for expected corresponding output value, 

while k is number of output nodes. 

Although MSE is the most widely used error function, 

it requires more training time and may become trapped in 

local minima before converging to a solution. It has been 

suggested by several authors, for example Lang [6] and 

Ooyen [7], that the cross-entropy error function improves 

the convergence of the training process, and can 

significantly reduce training time, the cross-entropy error 

function is as follows: 


Error= 

− 

k 

o 

∑∑ 

i= 

1 p= 1 

i i 

i 

(t log S + ( 1−t 

) log( 

1−S 

)) (2) 

i 

p 

p 

During our experiments, we also found that the cross 

entropy error function in equation 2 is pretty good, so we 

employ cross entropy as error function for both FCBP 

and PCBP. 

Before deriving the equations, we need to introduce 

the notations used in deriving the mathematical 

expressions of FCBP and PCBP training. 

Notations: 

n Numbers of input nodes 

h Numbers of hidden nodes 

o Numbers of output nodes 

x Sample input vector 

t Sample output vector 

j 

w l Weights between input node l to hidden node 

j 

j 

v p Weights between hidden node j to output 

node p 

σ () Activation function in hidden and output layer 

(here, we suppose it is sigmoid) 

j 

wc l Connection status between input node l to 

hidden node j 

j 

vc p Connection status between hidden node j to 

output node p (see definition 1) 

For FCBP, the components of the gradient of 

cross-entropy error function are given by equation 3 to 4: 

i 

i i 

i 

∂Error 

∂Error 

∂S 

p S p − t p ∂S 

p 

= × = 

× 

j 

i 

j i 

i j 

∂v 

∂S 

∂v 

S × ( 1− 

S ) ∂v 

i i 

S p − t p i 

i 

i j 

= 

× S p × ( 1− 

S p ) × σ ( ∑ x pwl 

) (3) 

i 

i 

S × ( 1− 

S ) 

p 

= ( S 

p 

i 

p 

− t 

(( S 

i 

p 

i 

p 

p 

p 

) × σ ( 

i 

p 

∑ 

x 

j 

p 

i 

p 

w 

i ∂σ 

( x 

∂Error 

∂Error 

∂S 

p 

= × 

× 

j 

i 

∂wi 

∂S 

p 

∂w 

∂σ 

( x w ) 

= 

∑ 

p 

j 

l 

) 

∑ 

− t ) × v ) × σ ( 

∑ 

p 

i 

p 

x 

i 

p 

p 

j 

l 

w 

j 

l 

p 

p 

∑ 

) × ( 1− 

σ ( 

j 

i 

p 

i 

p 

∑ 

(4) 

So, adjustment items of input to hidden weights and 

hidden to output weights can be calculated by equation 3 

and 4 plus learning rate. 

IV. PRUNING FCBP 

Before going on, we have to introduce a new 

definition: 

Definition 1: Connection Status: a vector that 

w 

j 

l 

) 

x w )) × x 

i 

p 

j 

l 

i 

p


represents how one node connects with its adjacent nodes 

in the following layer. 

From a macro perspective view, connection status 

represents network structure; while from a micro point of 

view, it is just a vector consisting of binary elements, that 

3 

is, zeros or ones. For example, if wc 2 = 1 , then we can 

say that the connection between second node in input 

layer and third node in hidden layer exists, while if 

wc 0 , then there is no connection between second 

3 

2 = 

node in input layer and third node in hidden layer. 

Actually, FCBP can be viewed as a particular PCBP 

whose connection status vectors are just ones. Hence, it 

can be concluded that by creating and maintaining such 

connection status vectors, FCBP can be easily defined 

and implemented. 

After training FCBP achieved predetermined accuracy, 

take 0.98 for example, unnecessary connections should be 

removed from the network in order to get a simple but 

efficiency PCBP. The pruning process consists of five 

steps: 

Step1: clustering weights between adjacent layers, it 

starts from the first node to the last one in the hidden or 

output layer, respectively. 

Step2: automatically determine a pruning bias that 

satisfied, if absolute clustered weights of connections that 

below this bias are all deleted, the pruning ratio can be 

met. 

Step3: removed all the connections that below the bias 

Step4: if network accuracy falls far below expected, 

then roll back pruning, set another pruning ratio, and go 

to step 2 

Step5: update connection status 

A. Algorithm for clustering weight 

The notations used in the algorithm are: 

β Clustering distance between connections 

w Weight vector of a node that contains all the 

connections that connect to it from its previous layer, 

either input to hidden layer or hidden to output layer 

num Length of w 

ClusterTyp e Represents which cluster type each 

w element belongs to 

ClusterVal ue A vector that represents clustered 

value 

ClusterSum Sum of weights that belong to a 

specific cluster type 

ClusterCou nt A vector that contains number of 

each cluster type 

Count Number of clustered type, that is the length 

of ClusterVal ue 

For each node of hidden or output layer, do the 

following: 

Step1: Initially, set 

ClusterSum ( 1) 

= w( 

1) 

ClusterVal ue( 

1) 

= w( 

1) 

Count = 1 

ClusterCou nt ( 1) 

= 1 


ClusterTyp e( 

1) 

= 1 

Step2: for each i = 2 to num , if there exists an 

index j that satisfies: 

max w( 

i) 

− ClusterVal ue( 

j) 

< β , 

j= 

1: 

Count 

Then it means that weight w( j) 

should be clustered 

with ClustedVal ue( 

j) 

, so set 

ClusterSum ( j) 

= ClusterSum ( j) 

+ w( 

i) 

ClusterCou nt ( j) 

= ClusterCou nt ( j) 

+ 1 

ClusterTyp e( 

i) 

= j 

Else, it means w( j) 

should be another cluster, then 

set 

Count = Count + 1 

ClusterSum ( count ) = w( 

i) 

ClustedVal ue( 

Count ) = w( 

i) 


i) 

= w( 

i) 

ClusterTyp e( 

i) 

= Count 

Step3: calculate the average clustered value for each 

cluster value; 

For i = 1 to Count 

Set 


i) 

= ClusterSum ( i) 

/ ClusterCou nt ( i) 

Step4: update w to relevant cluster value: 

For i = 1 to num 

Set 

w ( i) 

= ClusterVal ue( 

ClusterTyp e( 

i)) 

B. Deleting unnecessary connections 

We define the following criterion that evaluates which 

kinds of connections are unnecessary: connections that 

have relatively small weights. Small is an obscure word 

that it is difficult to determine exactly, especially facing 

the fact that the distribution of weights after training is 

unpredictable, as initial weights are random numbers 

usually ranging from zero to one. It is not practical to set 

pruning bias manually. In order to solve this problem, we 

propose a heuristic method to automatically generate such 

a pruning bias that depends on the distribution of weights. 

The algorithm is based on pruning ratio, which is defined 

as: 

Definition 2: Pruning Ratio: numbers of pruned 

connections divided by total connections of the previous 

FCBP, it ranges from zero to one. 

The algorithm is as follows: 

Step1: let µ ( 0 < µ < 1) 

be a predetermined 

pruning ratio that indicates how much connections should 

be pruned; let ϖ be the best pruning bias; num is the 

accumulated number of connections; 

Step2: sort ClusterVal ue in ascending order, at 

the same time, ClusterCou nt changes accordingly in 

order to make the two vectors still consistent with each 

other; set 

index = 1, num = ClusterCount( 

1) 

, ϖ = ClusterValue( 

1) 

Step3: do the following loop


WHILE( index < Count) do 

num 

IF( < µ ) THEN 

Count 

index = index + 1; 

num = num + ClusterCount( index); 

Continue 

ELSE 

ϖ = ClusterValue( index); Exit 

end 

After step 3, we can get the best pruning bias. When 

deleting unnecessary connections, we just set the 

corresponding elements in vector wc and vc to be 

zeros, and by doing this, a FCBP becomes a PCBP 

network. 

When pruning FCBP, one thing should be paid special 

attention to which may ignored by other researchers. To 

illustrate it, look at figure 3, on the left part is a pruned 

network which has three hidden nodes marked as A, B 

and C respectively. Notice that, node A has two 

connections with nodes in output layer (in bold line), 

while none with input layer, as they have been deleted as 

unnecessary connections in the above pruning process, 

though the probability of happening is small, but if it 

does happen, it must be handled properly, or else some 

error would happen, for example, for any kind of input 

pattern, the output of node A is always one if activation 

function of node A is sigmoid, because sum of input plus 

weight is zero, if node A has a bias, then the output of 

node A can be any float number as long as it varies with 

the value of bias, thus the actual output of output node is 

affected. For this kind of conditions, we propose that 

node A be deleted too, that is to delete connections 

between A and the output layer, then A is totally removed 

from the network as illustrated at the right part of figure 2. 

If node A has connections with its previous layer but none 

with following layer, then delete its connections with its 

previous too. By doing this, pruning connections and 

nodes can be handled together. 

Fig.3 Example of inconsistent connections 


V. TRAINING PCBP 

When apply PCBP in a specific domain, like data 

mining, usually, PCBP should be trained first just like 

FCBP does. The training process of PCBP is a slightly 

different comparing with FCBP, take a hidden node for 

example, not all the input nodes connect with it, so the 

actual input for hidden node is calculated by: 

n 

∑ 

i= 

1 

i i i 

( w × x × wc ) , notice that an additional items 

l 

l 

(connection status between input layer and hidden layer) 

is added, similarly, equation 3 and 4 change to equation 5 

and 6: 

∂Error i i 

i j j 

= ( S p − t p ) × σ ( ∑ x pwl 

wcl 

) (5) 

j 

∂v 

∂Error 

= 

∂w 

j 

i 

p 

∑ 

(( S − t ) × v ) × σ ( x w wc ) × ( 1− 

σ ( 

i 

p 

i 

p 

j 

p 

∑ 

i 

p 

j 

l 

VI. EXPERIMENTS 

j 

l 

∑ 

x w wc )) × x 

Enterprise data warehouse 

Sources Users 

Operation 

al 

database 

Operation 

al 

database 

Data files 

Sourcing Area 

Staging 

History 

Area 

Integrated Area 

Meta data and Security 

Finance 

Sales 

Marketing 

MPP (Massively Parallel Processing) 

Figure 4: data warehouse architecture 

In the data warehouse architecture, from Staging Area 

to Enterprise Report are considered as Enterprise Data 

Warehouse, because every one of them is integral part of 

a warehouse, and they can satisfy the current and future 

needs for all the business users across the enterprise. 

It is common understanding that data warehouse is 

basis for BI and DSS application, and implementing a 

successful data warehouse requires not only technologies 

but also methodology as well as culture and cooperation 

across the enterprise. 

The experiment data set which recorded submersible 

pump repair history contains four attributes classification 

codes, these attributes are separately: Single rotor electric 

power (kW Per Rotor), Cable Temperature Level(℃), 

Casing size(inch), Protector Length(m). In the following 

experiments, we use the data set to train FCBP and PCBP. 

The experiment data set which recorded submersible 

pump repair history contains four attributes classification 

codes, these attributes are separately: Single rotor electric 

power (kW Per Rotor), Cable Temperature Level(℃), 

Casing size(inch), Protector Length(m). In the following 

experiments, we use the data set to train FCBP and PCBP. 

i 

p 

j 

l 

(6) 

Enterprise Report 

j 

l 

i 

p 

`


First, we need to encode the numeric data into binary, 

as illustrated in table 1: 

Table 1 The data attribute encoding table 

Attributes Range Encoded 

Input 

Single rotor Single rotor electric power >8 1 1 1 

electric 

power 4


Fig 5: A PCBP after pruning 

VII. CONCLUSION 

We have presented here that how to construct a PCBP 

with cross-entropy as error function, and pruning 

algorithm was discussed in detail. Also the experiments 

showed that a PCBP had fewer connections, but still 

remained accuracy. When PCBP is applied in data 

recognition or other fields, like data mining, it learns 

faster than FCBP does, especially trained with a huge 

amount of sample data. 

ACKNOWLEDGMENTS 

The research work presented in this paper has been 

partially supported by science research foundation project 

of Shanghai bureau of education (NO.08YS103), and 

Shanghai important education subject project 

(NO.S20504). The authors would like to express their 

appreciation to the agency. 

REFERENCES 

[1] [1] Changchien, S.W.&Lu T. Mining association rules 

procedure to support on-line recommendation by 

customer and products fragmentation. Expert Systems 

with Applications, 20, 2001,pp:325-335. 

[2] [2] D. Rumelhart, G. Hinton, and R. Williams, Parallel 

Distributed Processing. MIT Press,Cambridge, MA, 1986. 

[3] [3] D. Sarkar, “Methods to speed up error back 

propagation learning algorithm,” ACM Computing 

Surveys, vol. 27, no. 4, pp. 519–542, 1995. 

[4] [4] Sanggil Kang, Can Isik. Patially Connected 

Feedforward Neural Network Structured by Input Types. 

IEEE Transactions on Neural Networks, Vol.16, No.1, 

January 2005. 

[5] [5] D. E. Duckro, D. W. Quinn, and S. J. Gardner, “Neural 

network pruning with Tuckey Kramer multiple 

comparison procedure,” Neural Computat., vol. 14, pp. 

1149–1168, 2002. 


[6] [6] Lang K.J and Witbrock M.J. Learning to tell two 

spirals apart. In proc. of the 1988 Connectionist Summer 

School, pp:52-59. Morgan Kaufmann, San Mateo, CA. 

[7] [7] A.Van Ooyen. Improving the Covergence of the 

Back-Propagation Algorithm. Neural Networks, Vol.5, 

1992, pp:465-471. 

Gang Li, Ph.D., born in Dong Yin 

city, Shan Dong Province , on 1970-08-04. 

He has a MS in computer science from 

Xi’an Jiaotong University and PhD in 

Information Management System from 

Donghua University. 

He is an associate professor in the 

Management Science and Technology at 

Management Department, Shanghai 

University for Science and Technology, China. In 2005/2009, 

he was an PhD candidate in management information system at 

Donghua University. His industrial career includes Information 

Management Department, Power Machinery Factory of Shengli 

oilfield. (1993-2004). Then he has been a teacher of Dongying 

Vocational College(2004-2005). 

Xingsan Qian ,Professor, Director of Shanghai branch of 

industrial engineering in Mechanical Engineering Society of 

China, president of industrial engineering teaching seminars in 

east China ,Shanghai registered consultants, Committee of 

Shanghai Institute of Electronics, Microelectronics.Professor 

Qian Xingsan is a famous expert in industrial engineering, 

logistics and engineering. He has undertake more than 40 

research subject about development, innovation and 

development, strategic over 20 years,has twice served as 

national information technology (or IC) policy drafting group 

members. 

His research areas:high (IC) industry development (strategy, 

reform, planning, industrial zone location); management science 

and engineering; Technology Management; regional innovation; 

Industrial Engineering . 

Awards: Has won third prize of Shanghai Science and 

Technology Progress Award 2 times, won second prize of 

Shanghai Science and Technology Progress Award.won second 

prize of National Science and Technology Progress Award. He 

has take charge of and participated more than 40 scientific 

research projects,has published 3 Book. 

Chunming Ye, Ph.D., Professor, Industrial Engineering 

expert of China, Secretary of the Shanghai Institute of Industrial 

Engineers, winner of Baosteel teachers in 2008. His research 

areas include: industrial engineering, intelligent algorithms, 

enterprise resource planning, supply chain management and 

evaluation of intellectual property, production planning and 

scheduling . He is the earlier of using cultural evolution of 

swarm algorithm applied to the field of production planning and 

scheduling researchers.He han won the first prize of scientific 

and technological progress of the State Administration of 

Machinery Industry (1999) ; third prize in 2005 in Shanghai 

teaching achievements; second Prize of 2009 State Education 

Commission Science and Technology Progress Award,and has 

published over one hundred and sixty papers at home and 

abroad. 

Lin Zhao Master, he received the degree from Dong Hua 

University in 2006. His research interests are data mining, 

artificial intelligence, knowledge management, etc.


A Cooperative Co-evolution PSO for Flow Shop 

Scheduling Problem with Uncertainty 

Bin Jiao 

Electric School, Shanghai DianJi University 

No.690, Jiang Chuan Rd., Min Hang District, Shanghai, 200240, China 

Email: abinjiaocn@163.com 

Qunxian Chen 

Electronic Information School, Shanghai DianJi University 

No.690, Jiang Chuan Rd., Min Hang District,Shanghai,200240, China 

Email: bchenqx@sdju.edu.cn 

Shaobin Yan 

School of Information Science and Engineering, East China University of Science and Technology 

No.130, Mei Long Rd., 200370, Shanghai, China 

Email: cyshaobin123@sina.com 

Abstract—Considering current situation of production 

scheduling with uncertainties in modern manufacturing 

enviroments, flow shop production scheduling model is 

established based on the theory of fuzzy programming, in 

which fuzzy processing time is considered and the duration 

time of intermediate is unlimited. The maximum 

membership function of mean value has been applied to 

solve the non-linear fuzzy scheduling model in order to 

convert the fuzzy optimization problem to the general 

optimization problem. Finally, a cooperative 

co-evolutionary particle swarm optimization algorithm 

based on catastrophe added to improve the diversity of the 

swarm (CCPSO) is adopted to solve flow shop production 

scheduling with uncertainty within infinite intermediate 

storage and the simulation results obtained are effective and 

satisfactory. 

Index Terms—uncertainty; fuzzy programming; Flow Shop 

scheduling problem; cooperative co-evolutionary particle 

swarm optimization algorithm; 


Production scheduling tackles effective allocation of 

production resources over time. Flow shop scheduling 

problem(FSSP), which represents nearly a quarter of 

manufacturing systems and information service facilities 

in use nowadays, is one of the most important issues in 

shop floor control of a manufacturing firm,. The first 

research conducted on the flow shop scheduling problem 

was proposed by Johnson (1954) [1], who developed an 

optimization algorithm to achive a minimum makespan 

for the n-jobs and 2-machines flow shop scheduling 

problem. Previously, researchers mainly studied flow 

shop scheduling problem based on ideal conditions such 

as processing time assigned or estimated as a fixed value 

etc.In many real world applications, however, there exist 

many uncertain factors including human intervention, 

incomplete information, and uncertain environment. 

Recently, considering lots of uncertain factors that appear 


doi:10.4304/jcp.6.9.1955-1961 

in operations, planning and other processes, researchers 

mostly conduct researches on uncertain processing times 

and due dates in the real world applications and use fuzzy 

number theory to describe this problem [2,3]. 

As to scheduling problems with fuzzy processing time, 

a few approaches have been developed. In fact, large 

quantities of uncertainties including fuzzy processing 

time and fuzzy due dates are always considered [4]. 

Fortemps [5], developing a fuzzy approach in job shop 

scheduling problem with imprecise durations, enrolled 

the important application of the uncertainty in time 

parameters. Chanas studied minimization of maximum 

lateness of jobs in a single machine scheduling problem 

[6]. Litoiu and Tadei [7] proposed some novel models for 

real-time task scheduling with fuzzy processing times and 

deadlines. Hong [8] applied triangular membership 

functions for flexible flow shop problem with two 

machine centers to examine uncertain processing times. 

Wu Chaochao [9] used an efficient genetic algorithm to 

solve single machine scheduling problems with fuzzy 

processing time and multiple objectives. Niu Qun [10] 

proposed a novel particle swarm optimization for flow 

shop scheduling problem with fuzzy processing time. Xu 

Zhenhao [11] established a scheduling model for flow 

shop problems with finite intermediate and adopted a 

fuzzy immune algorithm to optimize this problem. In this 

paper, flow shop scheduling problem with fuzzy 

processing time is considered and the time duration of 

intermediate is unlimited. Besides, a cooperative 


based on catastrophe (CCPSO) is adopted to verify the 

model and to solve the fuzzy scheduling problem. 

Ecological models and co-evolutionary architectures are 

effective methods to improve the performance of original 

particle swarm optimizer [12, 13]. And co-evolutionary 

scheme, which is inspired by the reciprocal evolutionary 

change driven by the cooperative [14] or competitive


interaction [15] between different species, can avoid the 

exponential increase in difficulty by dividing the search 

space into several smaller subspaces, and then conducting 

the overall optimization process over smaller regions. 

The remainder of the paper is designed as follows. In 

section 2, fuzzy scheduling problem is depicted briefly. 

The following section 3 introduces cooperative 


based on catastrophe (CCPSO) algorithm. Experiment is 

undertaken in section 4. Finally, we draw a conclusion in 

section 5. 

II. PROBLEM DESCRIPTION 

Flow shop scheduling problem (FSSP) is often 

expressed by the symbols 

n/ m/ P/ Obj . 

, in which n 

jobs J = {1, 2, . . . , n} have to be processed on m 

machines M = {1, 2, . . . , m}, P shows that only 

permutation schedules are considered and 

Obj . 

, the 

objective function, describes the performance measure by 

which the schedule is to be evaluated. Also, all machines 

should process all jobs according to the sequence of 

pre-defined permutation schedule. Hence a schedule is 

uniquely represented by a permutation of jobs. At any 

time, each machine can only process one job and each job 

can only be processed by one machine. 

A triangular fuzzy number is given to describe the 

uncertain processing time of products in this paper. The 

maximum membership function is defined as follows: 

⎧0, 

⎪ L 

⎪ c − x 

, 

⎪ M L 

x − x 

µ x( 

c) 

= ⎨ U 

⎪ x − c 

, 

⎪ U M 

x − x 

⎪ 

⎩0, 

A. Problem Definition 

x 

x 

c ≤ x 

L 

M 

< c ≤ x 

< c ≤ x 

c > x 

The following notation has been introduced to 

describe the problem more precisely. 

N ——a set of n products which must be processed, 

N = { 1, 

2, 

, 

i, 

n} 

; 

M ——a set of m processing units which are 

M = 1, 

2, 

, 

j, 

m 

; 

available for our purpose, { } 

T ~ 

ij 

——The processing time of products i on unit j, 

which includes the transfer time, the set-up time, the 

clean-up time, and so on. Because it is mutative and 

uncertain, it is represented by the triangular fuzzy 

number; 

S ~ 

ij 

——the starting time of product i processed on 

unit j, the parameter also is uncertain; 


L 

U 

M 

U 

(1) 

C ij 

~ 

——the completing time of product i processed 

on unit j , and it is represented by the triangular fuzzy 

number; 

S ie 

~ 

——the starting time of the last operation of 

product i; and it is represented by the triangular fuzzy 

number; 

T ie 

~ 

——the processing time of the last operation of 

product i, and it is represented by the triangular fuzzy 

number; 

In flow shop scheduling problem, every job has the 

same sequence of operating on all machines. All jobs are 

processed at time-zero. But the following constraints 

must be taken into account: 

Sequence Constraints 

T ~ 

S ~ 

S ~ 

ij ≥ i( 

j−1) 

+ i( 

j−1) 

, i ∈ N , j ∈ M 

(2) 

Equation (2) indicates that the operation of product i 

on unit j can start after completing its previous processing 

procedure, that is the starting time of each operation of 

product i can be more than or equal to the finishing time 

of the last operation. And the different procedure of the 

same product can not be operated at the same time. 

Resource Constraints 

S ~ 

ij 

≥ − 

S ~ 

( i−1) 

j + ( i 1) 

j i ∈ N , j ∈ M 

T ~ 

Equation (3) means that the product i on unit j can 

start after the completion of the previous product i-1, that 

is the same unit can’t process two or more different 

products at a time. 

Time Constraints 

(3) 

S i N , j M 

~ ij ≥ 0 ∈ ∈ 

(4) 

Equation (4) represents each product can be available 

at time zero. 

Moreover, we make the following assumptions 

regarding the process: there is no priority between 

products; once an operation has started, it can’t be 

interrupted unless having been finished; a unit can not 

process different products at one time, and a product 

can’t be processed by more than one unit simultaneously. 

There are many different optimal objectives, i.e. the 

maximum or average tardiness, the average flow time, the 

lateness and earliness and so on. In this paper, the 

scheduling goal is to find a feasible schedule which 

minimizes the maximum completion time, which is 

makespan: 

( ) ( ie ie ) T~ S ~ 

min makespan = min max + 

(5) 

In order to calculate the completion time of products 

with fuzzy durations, the addition and maximum 

operations are needed.


x~ = ( x ) 

Defining 1, 

x2 

, x3 

y~ = ( y ) 

and 1, 

y2 

, y3 

be 

the triangular fuzzy numbers, the addition and maximum 

operations are given in the form as follows: 

Fuzzy 

x~ + y~ = ( x 

) 

Addition: 1+ 

y1, 

x2 

+ y2, 

x3 

+ y3 

: 

x~ ∨ y~ = ( x 

) 

Fuzzy maximum: 1 ∨ y1, 

x2 

∨ y2 

, x3 

∨ y3 

B. Fuzzy Scheduling Model based on Triangular 

Fuzzy Number 

x~ = x , x , x 

( ) 

Triangular fuzzy number 1 2 3 

adopted to express uncertain processing time. Due to the 

resolvability of the fuzzy addition and maximum 

operations, the detail of solution can be described as 

follows: 

1) Ifi = 1 , 

ij 

ij 

j = 1: 

ij ij T ij 

~ 

T ~ 

S ~ 

C ~ 

S , 

~ 

= 0 

(6) 

= + = 

2) Ifi = 1 , j > 1: 

ij i( 

j ) ij ij ij i( 

j ) T ij 

~ 

C ~ 

T ~ 

S ~ 

C ~ 

C , 

~ 

S ~ 

(7) 

= −1 = + = −1 

+ 

3) If i > 1 , j = 1: 

ij ( i ) j 

ij ij ij ( i ) j T ij 

~ 

C ~ 

T ~ 

S ~ 

C ~ 

C , 

~ 

S ~ 

(8) 

= −1 = + = −1 

+ 

4) If i > 1 , j > 1: 

ij 

( i ) j C i( 

j ) , 

~ 

C , 

~ 

S max 

~ 

= −1 −1 

(9) 

5) Objective function is: 

( ) ( ie ie ) 

C ~ 

T 

min 

~ 

S ~ 

min makespan = min max + 

= 

( ) ij ij T ij 

~ 

S ~ 

C ~ 

= + 

( ie ) 

L M U ( C , C , C ) 

= min ie ie ie (10) 

As is shown above, the fuzzy programming problem is 

transformed into multi-objective programming model. 

C 

L 

ie 

, C 

M 

ie 

, C 

U 

ie 

Owing to 

being related to fuzzy 

L M U 

Tij 

, Tij 

, Tij 

processing time 

respectively, the solutions 

get by multi-objective programming model are also the 

worst solution, the most possible solution and the best 

solution of fuzzy programming model. So the following 

task is to apply the maximum membership function of 

mean value to manage to obtain a single objective model. 

C. Model Transformation 

Here, Zimmerman method is applied to transform ie C~ 

into two solutions including positive ideal solution (PIS) 

PIS 

NIS 

C ie 

C 

and negative ideal solution (NIS) ie , that is 

C ( k 1 , 2, 

3) 

PIS 

k = 

C ( k 1 , 2, 

3) 

and NIS 

k = 

respectively 


is 

formulated as follows: 

C 

C 

PIS 

1 

NIS 

1 

where, 

= min 

C 

= max C 

L 

ie 

L 

ie 

, 

, 

PIS 

C ie and 

C 

C 

PIS 

2 

NIS 

2 

NIS 

Cie 

= min C 

= max C 

M 

ie 

M 

ie 

, 

, 

C 

C 

PIS 

3 

NIS 

3 

= min C 

= max C 

(11) 

U 

ie 

U 

ie 

also represent the optimistic 

solution and the pessimistic solution, by which we can 

define another kind of membership function like this: 

µ 

Ck 

( x) 

⎧0, 

⎪ 

⎪ x − Cie 

= ⎨ NIS 

⎪Cie 

− C 

⎪ 

⎩1, 

PIS 

PIS 

ie 

, 

C 

PIS 

ie 

x > C 

≤ x ≤ C 

x < C 

NIS 

ie 

PIS 

ie 

NIS 

ie 

k = 1, 

2, 

3 

Then, the fuzzy scheduling model above can be 

transformed into the singular objective nonlinear 

objective model : 

L 

U 

max { Γα + ( 1 − Γ) 

α } 

L 

s.t. α ≤ µ 

U 

≤ α , k = 1, 3 

µ 

C 

k 

U 

L U 

α ≤ µ C , α ∈[ 

0, 1] 

C 

( x)( 

k = 1 , 2, 

3) 

2 

α (13) 

where k 

is satisfactory membership 

function of ie C~ L 

. α depends on the minimum value of 

µ C ( x)( 

k = 1 , 2, 

3) 

U 

k 

. α is determined by the 

µ C ( x)( 

k = 1 , 2, 

3) 

maximum value of k 

. During the 

actual decision-making process, the highest level of 

satisfaction of objective value is expected to gain in the 

most possible situation not in the worst or optimal cases. 

µ C ( x) 

Therefore, in the model above, let 2 be the 

maximum membership value while the minimum one is 

produced in the worst or the best circumstances. And the 

operator Γ is used to reflect the tendency degree of 

decision-maker choosing positive side and negative side. 

The smaller is the value Γ , the more positive is 

decision-making, on the contrary, the more negative is 

decision. 

III. COOPERATIVE PARTICLE SWARM OPTIMIZATION 

WITH CATASTROPHE (CCPSO) 

A. Review of Particle Swarm Optimization. 

Particle Swarm Optimization (PSO) is an evolutionary 

calculation technique proposed by Kennedy and 

Eberhart[16] in the mid 1990s. Different from other 

algorithm, PSO is simple and easily implemented due to 

having no operators such as crossover and mutation. It 

was inspired by the natural biologic phenomenon seen in 

a flock of birds attempting to find food through its own 

position as well as experience gained from others. The 

population of PSO is called a swarm and each individulal 

(12)


in the population is called a particle. PSO is an 

evolutionary computation technique through individual 

improvement plus population cooperation and 

competition. A particle’s status among the search space is 

characteristic with its position and velocity. Then, the 

position and the velocity are adjusted according to its 

own and its companions’ flying experience. 

Suppose that there is an d-dimensional search space for 

a swarm with m particles , and the i th particle is 

denoted by an d-dimensional vector 

X i = ( xi1, xi2, , 

xid) 

while its velocity is represented 

V 

by i = ( vi1, vi2, , vid) 

.Also, two key points directing 

particles moving to the best solution are i P and g P , of 

P 

which i = ( pi1, pi2, , 

pid) 

means the best 

previously visited position of the particle i and 

P = ( p , p , , 

p ) 

g g1 g2 gd 

means the position of the 

best individual of the whole swarm. The fitness value of 

each particle is evaluated by the objective function. 

During all the iteration, the velocity and position are 

updated according to the following equations: 

( k + 1) 

= vid 

( k) 

+ c1r1 

( pid 

( k) 

− xid 

( k)) 

+ c2r2 

( pgd 

( k) 

− x ( k)) 

(14) 

( k + 1) 

= x ( k) 

+ v ( k + 1) 

vid id 

xid id id 

( i= 1,2, , m; d = 1,2, , d) 

(15) 

where k is the iterative number, the variables 

c1, 

c2 

are 

learning factors, usually 

c1 

= c2 

= 2 

, which assign a 

fixed range for a particle’s moving and 1 r 

, 

r2 are 

elements from two uniform random sequences in the 

range (0, 1): 

r 1 ~U (0,1); 

r 2 ~U (0,1). 

B. Principle of Cooperative Co-evolution Algorithm 

Co-evolution mechanism, obviously a biologic process 

where population of interacting individuals challenge 

eachother in an ongoing of adaptation, can be classified 

into two main categories, cooperative co-evolution and 

competitive co-evolution. For cooperative co-evolution, 

in natural ecosystems, almost all species own appetence 

to interact with other species to improve the survival 

cooperatively. We can name the cooperative co-evolution 

as symbiosis, firstly introduced by German mycologist, 

Anton de Bary in 1879 [17]. 

As mentioned above, symbiosis is made up of three main 

categories including mutualism (both species benefit by 

the relationship), commensalism (one species benefits 

while the other species is not affected), and parasitism 

(one species benefits and the other is harmed) [18]. In this 

paper, we choose the mutualism and incorporate it into 

QPSO. 

(1) Form sub-swarms 

It is important to make up several sub-swarms for 

co-evolution algorithm. In this text, an initial main 


population is randomly generated, and each particle in the 

population has an initial main vector. Then according to a 

divide-up parameter set at the beginning, we separate the 

initial main vector into several sub-vectors. Through a 

kind of cooperative method introduced below, a 

newly-combined main vector is reached. Finally, the 

number of sub-swarm depends on the value of the 

divide-up parameter. 

(2) Design cooperative method 

Besides, how to design a cooperative method for 

sub-swarms is an important part of co-evolution 

algorithm. Generally speaking, cooperative method can 

be classified into three main categories namely, greedy, 

conservative and meta-heuristic methods [19]. Taking 

into account the advantage of fast convergence velocity, 

greedy method is applied to the co-evolution. In other 

words, the best particle of every sub-swarm is taken as 

the representative. Then we can gain a novel complete 

vector on condition that own current sub-vector and 

others’ representative are combined correctly. 

Fig. 1 An example for the cooperative method 

C. Catastrophe Operation in CCPSO 

In the process of searching best solution, velocity of a 

particle may be zero soon, which leads population to trap 

into local optimal. Therefore, on the basis of 

co-evolutionary particle swarm optimizer, catastrophe 

operation is brought into getting a novel algorithm. The 

catastrophe operation plays a part of judging every 

sub-swarm’s solution whether in a local convergence 

region and carries out some measures to ensure 

population’s global search. 

M 

Assuming that 1 = ( x1, x2,..., xn) 

and 

M 2 = ( y1, y2,..., yn) 

are two random individuals in 

n dimensional space, each variable in 1 M and 2 M is 

encoded as a m system, that is 

xi = ( xx i1 i2... xil 

) y 

i i = ( yi1yi2... yil 

) 

, i . So the 

Hamming Distance between 1 M and 2 M is defined 

as follows: 

n li 

H( M1, M2) = ∑∑ | xij − yij 

| (16) 

i= 1 j= 

1


xij = yij ( ∀i, 

j) 

From equation (3), when 

( , ) 0 

H M1 M 2 = 

, and the maximum value of the 

Hamming Distance between 

H( M1, M2) = ( m−1) ∑ li 

(17) 

n 

i= 

1 

1 M and 

Assuming that 1 1 2 

M = ( y , y ,..., y ) 

2 M is: 

M = ( x , x ,..., x ) 

2 1 2 n are two random individuals in 

n dimensional space, each variable in 1 M and 2 M is 

x = ( xx... x) 

encoded as a m i i1 i2 il 

system, that is i 

y = ( y y ... y ) 

i i1 i2 ili 

. So the dissimilarity factor is 

expressed as follows: 

( , ) 

H M1 M2 

µ = 

n 

( m−1) ∑l 

i 

i= 

1 

n 

(18) 

Obviously, the value of dissimilarity factor ranges from 

0 to 1.When two individuals are the same, the 

dissimilarity factor µ = 0 

. And the bigger µ is, more 

diversiform the population is. So it’s important to ensure 

the great difference between two different individuals. 

In the co-evolutional particle swarm optimizer, M 

sub-swarms are randomly separated into M/2 pairs of 

individuals. Then average dissimilarity factor µ is 

computed: 

M /2 

∑ µ i 

i= 

1 

(19) 

µ = 

M /2 

When average dissimilarity factor µ is very small, 

the local convergence appears. And catastrophe factor 

Ca is set. When µ < Ca 

, the catastrophe operation is 

adopted. In this paper, to reach a more ideal effect, the 

catastrophe operation, keeping best solution and 

reinitializing other particles, is chosen to increase 

diversity of population. 

IV. EXPERIMENTAL RESULTS 

To illustrate the effectiveness and performance of 

CCPSO for flow shop scheduling with fuzzy processing 

times to minimize makespan proposed in this paper, the 

scheduling problem of ten jobs on five machines has been 

selected to test. The fuzzy operating time of jobs on 

machines represented with triangular fuzzy number are 

listed in TABLE I. 


, 

, 

, 

During experiments, every run is repeated for 10 times 

and the population size is 60. The maximum iterative 

generation is 150. Also, the learning factors 1 c , 2 c is 2 

and weight parameter w is 0.3 in CCPSO. Catastrophe 

factor is 0.35 and the allowed catastrophe happen is 3 

times. 

In algorithm, two sub-swarms are given to conduct 

parallel evolutions, and then several experiments with 

different values of Γ are undertaken.Fig.2 and 3 are 

results for two sub-swarms when Γ= 0.3 . 

View from Fig. 2 and Fig. 3, two curve lines are 

depicted, in which the real-line represents the best 

solution is got by each generation and the broken one is 

the average objective solution of each generation. With 

the evolution of algorithm, the two lines travel towards 

the optimum point. It indicates that the novel algorithm 

CCPSO has a good convergence and the strong robust 

performance. 

objective value 

objective value 

Fig. 2 Evolutionary curve of species one 

Fig. 3 Evolutionary curve of species two 

To find how Γ effects on CCPSO, five experiments 

under different value of Γ are finished in the same 

condition and the results are taken down in TABLE Ⅱ. 

From TABLE Ⅱ, the smaller Γ is, the bigger 

objective fitness value is. Also, when Γ is 0.1, the


medium makespan is smaller than other cases. 

Meanwhile, a better scheduling scheme is obtained. On 

the contrary, when Γ is 0.9, the makesapn is the biggest 

among all results and the scheduling method is most 

negative. Therefore, Γ should be equaled a value 

ranging from 0 to 1 as small as possible. Besides, in any 

case, CCPSO algorithm can reach up to good effect and 

has a favorable convergence. 

The Fig. 4 is drawing for Γ = 0.3 . 

TABLE I. 

Uncertain processing time of productions 

Unit 1 Unit 2 Unit 3 Unit 4 Unit 5 

Job 1 (23 25 31) (11 15 21) (10 12 14) (34 40 46) (6 10 12) 

Job 2 (6 17 11) (37 41 47) (21 22 24) (28 36 40) (6 8 10) 

Job 3 (38 41 45) (137 155 167) (27 33 37) (111 121 141) (145 160 188) 

Job 4 (64 74 90) (8 12 16) (16 24 30) (40 48 58) (66 78 86) 

Job 5 (6 7 9) (69 95 107) (62 72 84) (51 52 56) (148 153 179) 

Job 6 (10 12 16) (8 14 16) (58 62 74) (26 32 38) (140 162 190) 

Job 7 (9 11 17) (5 7 12) (23 31 35) (20 26 30) (26 32 38) 

Job 8 (25 31 39) (35 39 43) (135 141 175) (4 6 10) (15 19 23) 

Job 9 (24 32 34) (84 92 98) (10 12 14) (8 14 18) (84 102 122) 

Job 10 (19 27 31) (109 114 128) (17 21 23) (78 90 102) (44 52 66) 


TABLE II 

Scheduling results under different Γ conditions 

Γ Job List Target Value Makespan L Makespan M Makespan U 

0.1 7 6 8 4 5 9 1 3 2 10 0.9950 792 887 1064 

0.3 7 1 6 5 4 3 8 9 2 10 0.9794 785 893 1061 

0.5 7 1 6 5 8 4 9 3 10 2 0.9769 785 893 1061 

0.7 7 2 6 4 5 8 3 10 9 1 0.9178 805 911 1082 

0.9 1 6 7 4 2 9 5 8 3 10 0.8438 828 941 1113 

Fig. 4 Gantt of the scheduling with optimal solution


V. CONCLUSION 

In this paper, Flow Shop production scheduling 

problem with uncertain processing time and infinite 

intermediate storage is researched on the basis of actual 

scheduling problem. The scheduling model is set up 

based on the theory of fuzzy programming, in which 

fuzzy processing is considered. The maximum 

membership function of mean value has been applied to 

solve the non-linear fuzzy scheduling model in order to 

transform the fuzzy optimization problem to the general 

optimization problem. Also, by simulating the 

phenomenon of the nature, Cooperative Particle Swarm 

Optimization with Catastrophe (CCPSO) is proposed, to 

which a cooperative and catastrophe operation is added. 

Finally, the novel algorithm is adopted to verify the 

model and satisfactory results are obtained. 

In my future work, this algorithm can be used to 

optimize more complex scheduling problem like 

multi-objectives and scheduling with uncertainty. 

Meanwhile, some other strategies can be taken into 

consideration to advance the algorithm so that 

performance can be enhanced. 


This work was supported in part by Shanghai Municipal 

Science and Technology Commission. (Grant No. 

10JC1405800), Project of Science and Technology 

Commission of Shanghai Municipality(08DZ1200505), 

Project of Shanghai Municipal Economic and Information 

Commission(09A118) ,and Key Discipline of Shanghai 

Municipal Education Commission(J51901). 

REFERENCES 

[1]Johnson SM. Optimal two- and three-stage production 

schedules with setup times included. Novel Research 

Logistics Quarterly 1954;1:61–8. 

[2] B.J.Lagewag, J.K.LEnstra and A.H.G.Rinnooy 

Kan, Job shop scheduling by implicit enumeration. 

Management Science, 1977, 24:441-450. 

[3] Prade H.Usmg Fuzzy Set Theory in a Scheduling Problem 

:a Case Study [J] .Fuzzy Sets and Systems, 1979, 

2(2):153-165. 

[4] Chanas S,Kasperski A.Minimizing Maximum Lateness in 

a Single Machine Sched uling Problem with Fuzzy 

Processing Times and Fuzzy Due Dates [J] . 

Engineering Applications of Artificial 


Intelligence,2001,14(3):377-386. 

[5] Fortemps P.Job shop Scheduling with Imprecise Duration: 

Fuzzy Approach [J] . IEEE, Transactions on Fuzzy 

Systems, 1997, 5(4): 557-569. 

[6] McCahon S, Lee E S. Job Sequencing with Fuzzy 

Processing Times [J]. Computers and Mathematics with 

Applications, 1990, 19(7):31-41. 

[7] Litoiu M, Tadei R. Real-time Task Scheduling with Fuzzy 

Deadlines and Processing Times [J]. Fuzzy Sets and 

Systems, 2001, 117(1):35-45. 

[8] TzungPei Hong,TzuTing Wang. Fuzzy Flexible Flow Shops 

at Two M achine Centers for Continuous Fuzzy Domains 

[J]. Information Sciences, 2000, 129(1-4):227-237. 

[9] Wu Chaochao, Gu Xingsheng. A Genetic Algorithm for 

Single Machine Scheduling with Fuzzy Processing Time 

and Multiple Objectives [J]. Journal of Donghua 

University, 2004, 21(3):185-189. 

[10] NIU Qun , GU Xing-sheng. A Novel Particle Swarm 

Optimization for Flow Shop Scheduling with Fuzzy 

Processing Time [J]. Journal of Donghua University, 

2008, 25(2):115-122 

[11] XU Zhen-hao, GU Xing-sheng, Earliness and tardiness 

flow shop scheduling problems under uncertainty with 

finite intermediate storage[J]. Control Theory & 

Applications, 2006, 23(3): 480-486. 

[12] M.A. Potter, The design and analysis of a computational 

model of cooperative co-evolution, Ph.D. Thesis, George 

Mason University, 1997. 

[13] G. Venter, R.T. Haftka, A two species genetic algorithm 

for designing composite laminates subjected to 

uncertainty, in: Proceedings of 37th 

AIAA/ASME/ASCE/AHS/ASC Structures, Structural 

Dynamics, and Materials Conference, 1996, 

pp.1848–1857. 

[14] M.A. Potter, K.A. De Jong, Cooperative co-evolution: An 

architecture for evolving coadapted subcomponents, 

Evolutionary Computation 8 (1) (2000) 1–29. 

[15] C.D. Rosin, R.K. Belew, New methods for competitive 

co-evolution, Evolutionary Computation 5 (1) (1997) 

1–29. 

[16] J Kennedy, R Eberhart.: Particle Swarm 

Optimization[C].In: Proc IEEE Int Conf on Neural 

Network (1995):1942-1948. 

[17] V. Ahmadjian, S. Paracer, Symbiosis: An Introduction to 

Biological Associations, Oxford University Press, New 

York, 2000. 

[18] A.E. Douglas, Symbiotic Interactions, Oxford University 

Press, Oxford, 1994. 

[19] Portter M A.: The design and analysis of a computational 

model of cooperative co-evolution [D]. Washington DC: 

George Mason University (1997).


A Double Margin Based Fuzzy Support Vector 

Machine Algorithm 

Kai Li 

School of Mathematics and Computer Science, Hebei University, Baoding, China 

Email: likai_njtu@163.com 

Xiaoxia Lu 

School of Mathematics and Computer Science, Hebei University, Baoding, China 

Email: yingli453@sina.com.cn 

Abstract—Although fuzzy support vector machine 

introduces the fuzzy membership degree in maximizing the 

margin and improves performance of classifier, it has not 

fully considered the position of training samples in the 

margin. In this paper, a double margin (rough margin) 

based fuzzy support vector machine (RFSVM) algorithm is 

presented by introducing rough set into fuzzy support 

vector machine. Firstly, we compute the degree of fuzzy 

membership of each training sample. Secondly, the data 

with fuzzy memberships are trained to obtain the decision 

hyperplane that maximizing rough margin method which 

contains the lower margin and the upper margin. In this 

algorithm, points in the lower margin have major penalty 

than those in the boundary in the rough margin. Finally, 

experiments on several benchmark datasets show that the 

RFSVM algorithm is very effective and feasible relative to 

the existing support vector machines. 

Index Terms—fuzzy support vector machine, double margin, 

classification, accuracy 


Support vector machine is firstly proposed by Vapnik 

et al for binary-class classification problem in 1995[1] [2] 

[3]. It has superior performance than traditional learning 

algorithms and has drawn the concern of many scholars 

in recent years. Support vector machine is based on 

Statistical Learning Theory (SLT) on (VC) dimension [4] 

deciding a confidence interval term and structural risk 

minimization (SRM) principle minimizing the upper 

bound of the generalization error. Support vector machine 

introduces a kernel tick to deal with non-separable 

problem. It maps points in the input space into a higherdimensional 

feature space such that the binary-class 

classification problem are indeed linearly separable or 

linearly approximately separable through a nonlinear map, 

and then finds an optimal separating hyperplane that 

maximizes the margin between two classes in the highdimensional 

feature space. However, there are still have 

Manuscript received October 1, 2010; revised December 1, 2010; 

accepted January 1, 2011. 

Corresponding author. Tel.:+86 0312 5079660 

Email: likai_njtu@163.com, yingli453@sina.com.cn 


doi:10.4304/jcp.6.9.1962-1970 

two questions needed to be further study which are how 

to effectively expand the binary-class classification 

problem to multiclass classification problem and how to 

overcome sensitivity or overfitting due to noises and 

outliers in optimal hyperplane. 

About the first problem, many scholars expand binaryclass 

classification to multiclass classification problem, 

wherein one-against-one (1-a-1) and one-against-all (1-ar) 

are common methods which transform multiclass 

classification problem into binary-class classification 

problem. Hsu and Lin studied a comparison of methods 

for multiclass support vector machines such as oneagainst-all, 

one-against-one, directed acyclic graph SVM 

(DAGSVM) [5]. To deal with unclassifiable region, 

Inoue and Abe proposed fuzzy support vector machine 

for multiclass problem [6]. This method uses fuzzy 

membership to resolve unclassifiable regions. In Ref. [7], 

the authors proposed a new fuzzy membership function in 

the nonlinear fuzzy support vector machine. Moreover, 

Yan and He propose a new method—multiclass fuzzy 

support vector machine of dismissing margin (DFSVM) 

based on class-center [8]. 

To the second problem, many scholars put forward a 

lot of variant SVM. In traditional support vector machine, 

each input point is fully assigned to one of two classes 

wherein some noises and outliers are ignored in training 

set. Therefore, it results in overfitting problem to some 

extent. In fact, only few input points can decide the 

hyperplane. In more and more real-world applications, 

the effects of the training points, especially noises and 

outliers, are different. Aimed at these problems, Lin and 

Wang introduced fuzzy set theory into support vector 

machine to overcome the sensitivity of noises and outliers 

to optimal hyperplane, called fuzzy support vector 

machine (FSVM) [9]. Fuzzy support vector machine 

associates a fuzzy membership with each input point such 

that different examples make different contributions to 

the learning of optimal surface. Other scholars combined 

FSVM with genetic algorithms (GA) [10] to improve the 

generalization performance of SVM. However, these 

need a prior knowledge of the distribution of training set. 

Wu and Law proposed fuzzy support vector regression


machine with Gaussian noises on triangular fuzzy number 

space to forecast fuzzy nonlinear system [11]. 

The rough set theory [12] is a powerful preprocessing 

tool to find out knowledge from an amount of uncertain 

and incomplete data and is applied to the support vector 

machines to reduce the features of data to process and 

eliminate redundancy. At the same time, it also improves 

performance of the classical support vector machines. To 

deal with the overfitting problem of the traditional 

support vector machine, Zhang and Wang proposed a 

rough margin based support vector machine [13]. In this 

paper, we propose a double margin based fuzzy support 

vector machine by combination of rough theory and 

fuzzy support vector machine, namely a double margin 

(rough margin) based on fuzzy support vector machine 

(RFSVM). The proposed method not only inherits the 

characteristic of the FSVM method, but also considers the 

effects of decision hyperplane depending on the position 

of training samples in the rough margin. So presented 

method further reduce overfitting due to noises or outliers. 

This paper is organized as follows. In section 2, a brief 

review of support vector machine is described. In Section 

3, we describe the proposed RFSVM in detail which 

contain both binary classification and multiple 

classification RFSVM. In the following section, we 

evaluate our method on benchmark data sets and compare 

it with the existing support vector machine. Some 

conclusions are given in the final section. 

II. SUPPORT VECTOR MACHINES ALGORITHM 

In this section, we briefly describe the support vector 

machines in binary classification problems. 

Given a dataset of labeled training points (x1, y1), (x2, 

y2),…, (xl, yl), where N 

( xi, yi) ⊆ R × { + 1, − 1} , i=1, 2…l. 

Supposed training data are linearly separable. That is to 

say, there is some hyperplane which correctly separates 

the positive examples and negative examples. The point x 

lying on the hyperplane satisfies +b = 0, where w 

is normal to the hyperplane. In this case, support vector 

machine algorithm finds the optimal separating 

hyperplane with the maximal margin. When the training 

data are linearly non-separable or approximately 

separable, it is needed to introduce the trade-off 

parameter. When the training data is not linearly 

separable, support vector machine learning algorithm 

introduces kernel strategy that maps the input data to a 

higher-dimension feature space z by using a nonlinearly 

mapping function ϕ () x and then the data in feature space 

z is indeed linearly or approximately separable. All 

training data satisfy the following decision function 

⎧+ 

1, if yi 

=+ 1 

f( xi) = sign( < w, x >+ b) 

= ⎨ 

. (1) 

⎩ − 1, if yi 

= -1 

All training points satisfy the following inequalities: 

⎧< 

wx , i > + b≥ + 1, ifyi = + 1 . (2) 

⎨ 

⎩ < wx , i > + b≤ − 1, ifyi = -1 

In fact, it can be written as yi( < w, xi > + b) 

≥ 1, 

i=1,2,…,l. above inequalities. It is seen that finding the 

hyperplane is equivalent to obtain the maximizing margin 


2 

by minimizing || w || subject to constraints (2). So the 

primal optimal problem is given as 

1 2 

min || w || 

wb , 2 

st .. y( < w, x >+ b) 

≥1. 

(3) 

i i 

i = 1, 2,..., l 

To solve optimal problem, we introduce Lagrange 

multiplier to transform the primal problem (3) into its 

dual problem that becomes the following quadratic 

programming (QP) problem: 

l l l 

1 

min ∑∑αiα jyy i j( xi⋅xj) −∑αi 

α 2 i= 1 j= 1 i= 

1 . (4) 

l 

∑ 

s. t. α y = 0, 0 ≤ α , i= 1,2,..., l. 

i= 

1 

i i i 

In classifier, the solution in feature space using a 

linearly mapping function ϕ ( x) 

only replaces the dot 

product x ⋅ x j by inner product vectors ϕ( x) ⋅ ϕ( 

x j ) . The 

mapping function ϕ( x) 

and ϕ ( xi 

) satisfy 

< ϕ( x), ϕ( 

xj) >= K( x, xi) 

, where K( x, xi) is called kernel 

function. In real world application, we would never need 

to explicitly know what ϕ is. A decision function with 

SVM is obtained by computing dot products of a given 

test point x, or more specifically by computing following 

sign: 

Ns 

* 

f( x) = α y ( s ⋅ x) + b 

∑ 

i= 

1 

i i i 

Ns 

* 

∑α 

i 

i= 

1 

iϕ i ϕ 

Ns 

* 

= ∑α 

i 

i= 

1 

i ( i, 

) + 

. (5) 

= y ( s ) ⋅ ( x) + b 

yK s x b 

Where the coefficient α is positive, i 

i 

s is support vector 

and Ns is the number of support vectors. 

In most cases, as the learning of a suitable hyperplane 

is too restrictive to be of practical use and causes a large 

overlap of classes, there is nonexistent some separable 

hyperplane. To deal with linearly non-separable data, it 

often allows that some points are misclassified, and 

introduces nonnegative slack variables ξ > 0 measuring 

the number of misclassifications and a punishment 

parameter C which is a cost trade-off between 

maximizing the margin and minimizing the classification 

error of training data. The sum of the slacks ∑ ξ is an 

i 

upper bound on the number of training errors. And, the 

original constraints (2) are relaxed to 

yi( < w, xi>+ b) ≥1 − ξi, 

i = 1,2,..., l. 

(6) 

Thus, constructing optimal hyperplane is equivalent to 

solve the following optimization problem: 

l 

1 2 

min || w|| + C∑ξi 

wb , , ξ 2 

i= 

1 

st .. yi( < w, ϕ( xi) >+ b) 

≥1−ξ (7) 

i 

ξ ≥ 0, i = 1, 2,..., l. 

i


The corresponding dual problem is as following: 

l l l 

1 

min ∑∑αiαjyyK i j ( xi⋅xj) −∑αi 

α 2 i= 1 j= 1 i= 

1 . (8) 

l 

∑ 

s. t. α y = 0, 0 ≤α ≤ C, i= 1, 2,..., l 

i= 

1 

i i i 

To overcome the difficulty of the value of chosen 

parameter C, an alternative classifier model, called ν- 

SVM was proposed and developed [14] [15] [16]. In this 

model, C is replaced by a more meaningful parameter 

ν ∈ (0,1) , which is the lower and upper bound on the 

number of training points that are support vectors and that 

lies on the wrong side of the hyperplane, respectively. 

The primal optimization problem of this algorithm 

becomes 

l 

1 2 1 

min || w || − νρ + ∑ ξi 

wb , , ξ , ρ 2 

l i= 

1 

st .. yi( < w, ϕ ( xi) > + b) 

≥ ρ −ξi 

, (9) 

ξi≥ 0, ρ ≥ 0, i = 1, 2,..., l 

where variables wbξ ,,, ρ are optimized. 

The dual problem of this primal optimal problem can 

be solved by the following quadratic optimization 

problem: 

l l 1 

min ∑∑αα 

i jyyK i j ( xi, xj) 

α 2 i= 1 j= 

1 

l 

1 . (10) 

st .. ∑αiyi 

= 0, 0 ≤αi ≤ , 

l 

i= 

1 

l 

∑ 

i= 

1 

α ≥ vi , =1,2,..., l. 

i 

In many practices, training points are not fully 

assigned to one class of two classes, so Lin and Wang 

proposed fuzzy support vector machine (FSVM) [8]. 

Given labeled each training point associate with a fuzzy 

membership, namely (x1,y1,s1),(x2,y2,s2),…,(xl,yl,sl), where 

N 

( x, y) ⊆ R × { + 1, − 1} , i = 1, 2,..., l and si ( 0 < s ≤ 1 ) is 

i i 

fuzzy membership corresponding to each training point, 

the parameter ξ is the measure of misclassification, the 

i 

term siξ is a measure of error with different weight. It is 

i 

equivalent to solve the following optimal problem: 

l 

1 2 

min || w|| + C∑siξi wb , , ξ 2 

i= 

1 

st .. yi( < w, ϕ ( xi) > + b) 

≥1 −ξi 

, (11) 

ξi 

≥ 0, i = 1, 2,..., l 

where C is a constant. 

Finding the optimal hyperplane can be solved by 

constructing a Lagrange function and transformed the 

primal problem into the following dual problem: 

l l l 1 

max ∑αi − ∑∑αα 


α 

i= 1 2 i= 1 j= 

1 

, (12) 

l 

∑ 

s. t. α y = 0, 0 ≤α ≤ sC, i= 1,2,..., l 

i= 

1 

i i i i 


i 

where α is the nonnegative Lagrange multiplier 

i 

associated with the inequality constraint. The points 

corresponding to α i > 0 are called support vectors. 

From the Kuhn-Tucker conditions, we can obtain 

αi( yi( < w, xi >+ b) 

− 1 + ξi) 

= 0 

. (13) 

(siC- αi) ξi 

= 0, i = 1, 2,..., l 

To deal with overfitting problem due to noises or 

outliers in support vector machine, Zhang and Wang 

proposed a rough margin based support vector machine 

(RMSVM) [13]. In this paper, they considered the 

training points with different effects on the learning of the 

separating hyperplane depending on their positions in the 

rough margin. It searches the optimal separating 

hyperplane that maximizes the rough margin which 

contains the lower and upper margin. The primal 

optimization problem of RMSVM can be defined as 

l l 

1 2 1 δ ' 

min || w|| 

−νρl − νρu + ∑ξi + ∑ξ 

wb , , ξξ , ', ρ, i 

l ρu 

2 

l i= 1 l i= 

1 

' 

st .. y( < w, ϕ( x) >+ b) 

≥ρ −ξ −ξ 

, (14) 

i i u i i 

ξi ρu ' 

ρl ξi ρl ρu 

0 ≤ ≤ − , ≥0, ≥0, ≥0. 

ρ ρ 

whereσ > 1 , l 

u and ( ρu > ρ ) are the width of 

l 

|| w || || w || 

the lower margin and upper margin, respectively. 

This primal optimal problem can be solved by its dual 

problem as follows 

l l 1 


i jyyK i j ( xi, x j) 

α 2 i= 1 j= 

1 

l 

δ . (15) 

st .. ∑ αiyi = 0,0 ≤αi ≤ 

l 

i= 

1 

l 

∑ 

i = 1 

α ≥ 2v 

i 

Ⅲ. DOUBLE MARGIN BASED FUZZY SUPPORT MACHINE 

A. Binary Classification Case 

Aimed at fuzzy support vector machine, to further 

overcome the overfitting problem and to reduce the 

effects of outliers or noises, we propose a double margin 

(rough margin) based fuzzy support vector machine 

(RFSVM), in which it not only associates a fuzzy 

membership with each training point, but also considers 

each training example’s position in the rough margin. 

According to rough theory, rough margin contains lower 

margin 2ρ 

l and upper margin 2ρ 

u ( ρu > ρ ). l 

|| w || 

|| w || 

The region of training points within the lower margins is 

equivalent to positive region in rough theory; the data in 

this region are noises or outliers. The regions of the 

training examples within the upper margins and outside 

the lower margins are equivalent to boundary regions. In 

addition, the data outside the upper margins 

corresponding to the negative regions are correctly 

classified and are non-support vectors that are not noises 

and outliers. Training examples in the lower margin have


major penalty but those in the boundary region of the 

rough margin have minor penalty to hyperplane. RFSVM 

finds the maximum rough margin in some highdimensional 

feature space. 

Similar to classical support vector machine, the primal 

problem of RFSVM can be described as: 

l l 

1 2 1 δ ' 

min || w|| −νρl − νρu + ∑siξi + ∑siξ 

wb , , ξξ , ', ρ, i 

l ρu 

2 

l i= 1 l i= 

1 

' 

st .. y( < w, ϕ( x) >+ b) 

≥ρ −ξ −ξ 

. (20) 

i i u i i 

' 

0 ≤ξi ≤ρu −ρl, ξ ≥0, ρ 0, 0 

i l ≥ ρu 

≥ 

Here, we set δ ≥ 1 and ν ∈ (0,1) . ρ l and ρ u construct 

the inner and the outer wall of the boundary regions, 

respectively. When δ = 1 , RFSVM is equivalent to 

contain the parameter ν of fuzzy support vector machine 

without ρ u andδ . When the data points locate on the 

positive region, those are regarded as outliers and noises. 

When training points lie on the negative regions, they are 

regard as non-support vector and are not noises and 

outliers. The slack variable ' 

ξ and ξ is introduced by 

datum locating in the positive region and boundary 

regions of rough margin, respectively. 

To solve this optimization problem, we construct the 

Lagrange function: 

l l 

1 2 1 δ ' 

L= || w|| −νρl− νρu+ ∑siξi+ ∑siξi 

2 

l l 

i= 1 i= 

1 

l 

' 

− ∑αi( 

yi( < w, ϕ( xi) >+ b) 

− ρu+ ξi+ ξ) 

i 

i= 

1 

l l l 

' 

−∑βξ i i−∑λi( ρu−ρl−ξi) −∑ηξ i −θρ i 1 l−θ2ρu i= 1 i= 1 i= 

1 

where i, i, i, i, 

1, 2 0 

, (21) 

α β λ η θ θ ≥ are Lagrange multipliers. 

According to KKT conditions, the parameters satisfy the 

following conditions: 

l ∂ L 

= w− ∑ αiyiϕ( xi) 

= 0 

∂w 

i= 

1 

l ∂ L 

=− ∑ α iyi = 0 

∂b 

i= 

1 

∂ L si 

= −αi − βi + λi 

= 0 

∂ξi 

l 

∂L 

δ si 

= −α 0 

' 

i − ηi 

= 

∂ξ 

l 

i 

l 

∂L 

=− v + ∑ λi− θ1= 

0 

∂ρ 

l 

i = 1 

l l 

∂L 

=− v + ∑α i −∑ λi − θ2= 

0 

∂ρ 

u 

i= 1 i= 

1 

' 

α i( yi( < w, ϕ ( xi) > + b) 

− ρ u + ξ i + ξ ) = 0 

i 

βξ i i = 0 

. (22) 

λi( ρ u − ρ l − ξ i) 

= 0 

' 

ηξ i = 0, θ1ρ0, 2 0 

i 

l = θ ρ u = 

Applying these equations into the Lagrange function 

(21), the primal problem (20) can be transformed into the 

Wolf dual problem: 


l l 1 



α 2 i= 1 j= 

1 

st .. 

l 

∑αiyi 

= 0, 0 ≤αi δ si 

≤ 

l 

. (23) 

i= 

1 

l 

∑ 

i= 

1 

α ≥ 2v 

i 

From the Kuhn-Tucker conditions, we obtain 

' 

αi( yi( < w, ϕ( xi) >+ b) 

− ρu + ξi + ξ ) = 0 

i 

si 

( − αi + λi) ξi 

= 0 

l 

δ si 

' 

( − αi) ξi 

= 0 

l 

. (24) 

The point xi with the corresponding to α i = 0 satisfies 

yi( < w, ϕ( xi) >+ b) 

> ρ lying outside the upper margin 

u 

of rough margin. The point xi with the corresponding to 

α i > 0 is called support vector. When 

si 

0 < α i < 

l 

, the 

one lying on the border of the upper margin of the 

hyperplane satisfies yi( < w, ϕ( xi) >+ b) 

= ρu. 

When 

si 

α i = , the one lying within the boundary of the rough 

l 

margin, 

satisfies yi( < w, ϕ( xi) >+ b) 

= ρu−ξi, where ξ i > 0. 

When si δ si 

< α i < , the one lying on the boundary of the 

l l 

lower margin satisfies yi( < w, ϕ( xi) >+ b) 

= ρl. 

When 

δ si 

α i = 

l 

, the one lying within the lower margin, is 

' 

misclassified and satisfies yi( < w, ϕ( xi) >+ b) 

= ρu − ξ . 

i 

From the optimal values α i ( i = 1, 2,..., l) 

of (21), we 

can obtain the decision function of RFSVM: 

f ( x) = sgn( α K( x, x ) + b) 

, (25) 

∑ 

i= RSV 

i i 

where RSV denotes the index set of data with 

α > 0 , 1 

b =− α iyi( K( xi, x j) + K( xi, xk)) 

,where 

∑ 2 i= RSV 

s j 

i∈{ j| α j ∈ (0, ), y j = 1} and 

l 

s j 

k ∈ { j | α j ∈ (0, ), y j = − 1} , or 

l 

sj δ sj 

i∈{ j| α j ∈ ( , ), y j = 1} and 

l l 

sj δ sj 

k∈{ j| α j ∈ ( , ), yj 

= − 1} . 

l l 

The design of fuzzy membership function is the key to 

the fuzzy algorithm using fuzzy technology. In this paper, 

we use class center method to generate fuzzy membership. 

Firstly, we denote the mean of class +1 as classcenter 

x + , and the mean of class -1 as class center x − . 

The farthest distance between the each class training


points and its class-center, the radius of class are 

r = max || x − x || and r = max || x − x || , 

+ + 

{ xi: yi= 

1} 

i 

− − 

{ xi: yi=−1} 

respectively. 

Fuzzy membership s i is a function of the mean and 

radius of each class 

⎧1 

−|| x+ − xi ||/( r+ + σ ), if yi 

=+ 1 

si 

= ⎨ 

(26) 

⎩ 1 −|| x- − xi ||/( r- + σ ), if yi 

= -1 

where σ > 0 is used to avoid the case s = 0 . 

B. Multiple Classification Case 

In this section, we extend binary-class classification of 

rough margin to multi-class classification and implement 

it on one-against-all and one-against-one methods in 

detail. 

The one-against-all method constructs p RFSVM, 

where p is the number of classes. The ith RFSVM is 

trained with all training point in the ith class with positive 

class, and all other training points are considered as 

negative class; at the same time, we computed fuzzy 

membership of each training point and the position of the 

rough margin of training set. p - RFSVM algorithm 

obtains linear decision function 

i i 

f ( x, y) =sgn(g ( x, y)) 

wb , 

i i 

=sgn(< w , ϕ( 

x, y)>+ b ) 

i 

class of x = max ( g ( x, y)). 

i= 1,2,, p 

i 

When the value of g ( xy , ), i= 1, 2,..., l is equivalent 

or very little different, unclassifiable region exists. We 

use rough margin and fuzzy membership functions to 

resolve unclassifiable regions to realize the same 

classification results with that of traditional multiclass 

support vector machine and multiclass fuzzy support 

vector machine. 

The one-against-one method constructs p(p-1)/2 

RFSVM where p is the number of classes. The ith 

RFSVM is trained with the training point in the ith class 

with positive class and the jth class with negative class 

( i≠ j ). At the same time, we computed fuzzy 

membership of each training point and according to each 

example’s position of training set in the rough margin to 

further reduce the effects of the outliers and noises. The 

decision function for class i against class j, with the 

ij 

maximum margin, is fw, b () x =< wij , ϕ()>+ x b where ij 

ij w 

is the l-dimensional vector, bij is a scalar and 

ij ji 

f ()=- x f (). x For test vector x, we calculate 

wb , wb , 

p 

g ()= x sgn( f ()), x class of x = 

∑ 

i ij 

wb , 

j= 1, j≠i i 

i 

max ( g ( x)). 

i= 1,2,, p 

Ⅳ. EXPERIMENTAL RESULTS AND ANALYSIS 

We conduct some experiments on benchmark datasets 

to test performance of RFSVM algorithm and compare it 

with other related approaches which include rough 

margin based support vector machine (RMSVM), ν-SVM, 


i 

fuzzy support vector machine (FSVM) and Standard 

SVM. Experiments are conducted on 16 different data 

sets from UCI [17], Statlog [18] and TKH96a [19]. 

Details about these data sets are given in Table I. In 

selected data sets, the number of features has a large 

range. In experiments, we use randomly selected 

techniques to evaluate the performance of an algorithm. 

In random selecting approach, dataset is divided into two 

parts: training and testing set. 

For each dataset of all experiments, the experiments 

are repeated ten times using randomly selected training 

and testing sets (70% of the examples for training and 

30% for testing) from each dataset. At the same time, we 

compute the predicted accuracy of each testing set every 

time. The parameter C is fixed on 100 and 10. For each 

dataset, the best parameter value ν is used for training. In 

most cases, the selected optimal parameter ν of ν-SVM 

was between 0.3 and 0.6. For RFSVM, the ν value is also 

within the range of 0.3-0.6 and δ is within 3.0-15. In our 

experiments, RFSVM and RMSVM use exactly the same 

parameter values on each dataset. In the experiments, we 

2 

use Gaussian kernel, Kxx (, ') = exp( −γ|| x− x'||), 

where 

γ = 1.0 . 

The experimental results are shown in Table II and Fig. 

1. The average classification accuracies of each algorithm 

are presented in Table II. The best result in each dataset 

using different algorithm is shown in boldface. It is seen 

by Table II that RFSVM outperforms the other support 

vector machine learning algorithm in most cases. In 

addition, RFSVM usually improves the classification 

results of the fuzzy support vector machine. In some 

cases, the improvement is very large such as Fourclass, 

German and Heart. Especially Australian, Diabetes, 

Fourclass, German, and Liver-disorders have smaller 

standard deviation. 

The experimental results in Table II demonstrate that 

in most cases RFSVM beats ν-SVM. Similarly, RFSVM 

usually outperforms RMSVM and standard SVM. These 

conclusions are further validated by Fig. 1. This means 

that for given dataset, introducing the rough margin and 

fuzzy membership is a good choice. 

TABLE I. 

DATASETS AND THEIR CHARACTERISTICS 

Dataset Data 

items 

Features Class Source 

Australian 695 14 2 Statlog 

Breast-cancer 683 10 2 UCI 

Bupa 345 6 2 UCI 

Cancer 683 7 2 UCI 

Diabetes 768 8 2 UCI 

Fourclass 862 2 2 TKH96a 

German 1000 24 2 Statlog 

Glass 214 9 6 UCI 

Heart 270 13 2 Statlog 

Iris 150 4 3 UCI 

Liver-disorders 345 6 2 UCI 

Sonar 208 60 2 UCI 

Splice 1000 60 2 UCI 

Vowel 528 10 11 UCI 

Wdbc 569 30 2 UCI 

Wine 178 13 3 UCI


The classification accuracies of RFSVM, RMSVM, ν- 

SVM, FSVM-100 (C=100), SVM-100(C=100), FSVM- 

10(C=10) and SVM-10(C=10) on each dataset with 10 

times are shown in Fig. 2, where the digit 1-7 in the xaxis 

represents different classifiers respectively and the yaxis 

presents the classification accuracy of repeated 10 

times. 

To test whether the new proposed algorithm is superior to 

current algorithm, two-pairs t-test is performed among 

RMSVM and other algorithms which contain RMSVM, 

ν-SVM, FSVM-100, SVM-100, FSVM-10, and SVM-10. 

Results are presented in Table III. It was shown that for 

all datasets the differences between the results obtained 

by two compared classifiers were statistically significant 

(significance p < 0.05). From the point of view of 

statistics, win means that RFSVM algorithm is 

Accuracy 

Accuracy 

significantly better than any other algorithm; when tie 

appear, it shows that there is no obvious difference 

between two algorithms; but when significance is loss, 

the performance of RFSVM algorithm is inferior other 

binary classification support vector machine algorithm. 

0.86 

0.84 

0.82 

0.8 

0.78 

0.76 

TABLE II. 

EXPERIMENTAL RESULTS WITH DIFFERENT METHOD 

RFSVM RMSVM ν-SVM FSVM-100 SVM-100 FSVM-10 SVM-10 

Figure 1. Average value of classification accuracy for all datasets. 

Dataset RFSVM RMSVM ν-SVM FSVM-100 SVM-100 FSVM-10 SVM-10 

Australian 0.8606 0.7675 0.8575 0.8282 0.8466 0.8575 0.8593 

Breast-cancer 0.9606 0.9446 0.9619 0.9543 0.9494 0.9559 0.9641 

Bupa 0.7168 0.5918 0.6897 0.6696 0.7063 0.5559 0.6678 

Cancer 0.9659 0.9428 0.9623 0.9636 0.9619 0.9685 0.9685 

Diabetes 0.7623 0.6757 0.7458 0.7528 0.7584 0.7670 0.7690 

Fourclass 0.9912 0.9902 0.9414 0.7996 0.8125 0.7989 0.8080 

German 0.7400 0.7036 0.7291 0.6894 0.7009 0.7158 0.7127 

Heart 0.8272 0.7778 0.8092 0.7789 0.7767 0.8081 0.7957 

Liver-disorders 0.7386 0.6128 0.6538 0.6976 0.7343 0.5621 0.6906 

Sonar 0.8773 0.8846 0.8889 0.8773 0.8846 0.8889 0.8846 

Splice 0.7945 0.7688 0.7688 0.7585 0.7688 0.7115 0.7688 

Wdbc 0.9793 0.9756 0.9596 0.9734 0.9708 0.9788 0.9777 

Different algorithm 

(1) Australian 


(5) Diabetes 


Accuracy 

Accuracy 


(2) Breast-cancer 


(6) Fourclass 

Accuracy 

Accuracy 


(3) Bupa 


(7) German 

Accuracy 

Accuracy 


(4) Cancer 


(8) Heart


Accuracy 

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, 

7-SVM-10). 

TABLE III. 

EXPERIMENTAL RESULTS OF TWO-PAIRS T-TEST FOR ALL DATASETS 

Accuracy (70%) Win Tie Loss 

RFSVM RMSVM 10 2 0 

RFSVM ν-SVM 5 7 0 

RFSVM FSVM-100 8 4 0 

RFSVM SVM-100 2 7 0 

RFSVM FSVM-10 4 8 0 

RFSVM SVM-10 5 7 0 

Besides above experiments with binary classification, 

we also perform some experiments on multi-class 

datasets. RFSVM is also compared with multiclass fuzzy 

support vector machine (FSVM) and multiclass support 

vector machine (SVM) on one-against-all method and 

Accuracy 


(9) Liver-disorders 


(1) Glass 

Accuracy 


(10) Sonar 

Accuracy 


(11) Splice 

TABLE IV. 

EXPERIMENTAL RESULTS OF MULTI-CLASS CLASSIFICATION PROBLEM 

one-against-one method. We set C=100 in default. For 

each algorithm, we estimate the generalized accuracy 

using same kernel function, kernel parameters γ and cost 

parameters C in multiclass FSVM and multiclass SVM. 

Experimental results are shown in the Table IV. It can be 

seen that the accuracy obtained by RFSVM is same or 

even better compared with FSVM and SVM aimed at 

both the one-against-one method and one-against-all 

method. 

Similarly, we give the average accuracy and standard 

deviation as shown In Fig. 3 and Fig. 4. The x-axis 

represents the classifiers, namely RFSVM, FSVM and 

SVM. The y-axis represents the average accuracy and 

standard deviation of ten times on random selecting 

method. Multi-class RFSVM improves the generalization 

ability compared with multi-class support vector machine 

and fuzzy support vector machine, although it has larger 

standard deviation than the others. 

Dataset 

One-against-All method 

RFSVM FSVM SVM 

One-against-One method 

RFSVM FSVM SVM 

Glass 0.6358 0.6006 0.6104 0.6431 0.6155 0.6286 

Iris 0.9667 0.9600 0.9536 0.9697 0.9688 0.9536 

Vowel 0.9748 0.9740 0.9731 0.9887 0.9868 0.9851 

Wine 0.9668 0.9649 0.9652 0.9812 0.9706 0.9689 


(2) Iris 


(3) Vowel 

Figure 3. Experimental results with different algorithms on one-against-all method (1-RFSVM, 2-FSVM, 3-SVM). 


Accuracy 

Accuracy 

Accuracy 

Accuracy 


(12) Wdbc 


(4) Wine


Accuracy 


(1) Glass 

Figure 4. Experimental results with different algorithms on one-against-one method (1-RFSVM, 2-FSVM,3-SVM). 

Ⅴ. CONCLUSIONS 

The support vector machine is a powerful tool for 

classification. However, the final decision function 

obtained by the support vector machine depends on few 

extreme value points, which makes the support vector 

machine sensitive to outliers or noises in the training set. 

In this paper, following the rough theory, we propose a 

double margin (rough margin) based fuzzy support vector 

machine that combines the notion of rough set with the 

fuzzy support vector machine to deal with the outlier 

sensitivity problem of fuzzy support vector machine, and 

then we design a classifier building method based on 

fuzzy support vector machine. The key idea of building 

the classifier is to find suitable fuzzy membership 

function and controlled parameter. This combination 

allows us adaptively consider more data information in 

the construction of the optimal hyperplane. The double 

margin (rough margin) based fuzzy support vector 

machine depends on the number of training set and the 

position of training data in rough margin. In this RFSVM, 

it consists of three regions: positive region, negative 

region and boundary region. It makes the original crisp 

margin become rough margin, the lower margin and the 

upper margin. The user can control the parameter 

ν and δ . One advantage of this method is that the 

classifier RFSVM is effective and robust with respect to 

misclassification and it considers the position of rough 

margin in fuzzy support vector machine. The 

experimental results on 16 datasets demonstrate that the 

generalization performance of RFSVM is better than the 

other SVM classifiers. 


This work is supported by Natural Science Foundation 

of China (No. 60773062, No. 61073121) and Nature 

Science Foundation of Hebei Province (No. 

F2009000236). 

REFERENCES 


(2) Iris 

[1] V. N. Vapnik, “The Nature of Statistical Learning 

Theory.” New York: Springer-Verlag New York. 1995. 

“ISBN:0-387-94559-8” 

[2] C. Cortes, and V. N. Vapnik, “Support-Vector 

Networks.” Machine Learning, vol. 20, pp. 273-297, 

1995. “doi:10.1023/A:1022627411411” 


Accuracy 

Accuracy 


(3) Vowel 

Accuracy 


(4) Wine 

[3] C. J. C. Burges, “A tutorial on support vector machines 

for pattern recognition.” Data Mining and Knowledge 

Discovery, vol. 2, no. 2, pp. 121-167, June 1998. 

“doi:10.1023/A:1009715923555” 

[4] A. Blumer, A. Ehrenfeucht, D. Haussler, and M. K. 

Warmuth, “Learnability and the Vapnik-Chervonenkis 

Dimension.” Journal of the Association for Computing 

Machinery, vol. 36, no. 4, pp. 929-965, 1989. 

“doi:10.1145/76359.76371” 

[5] C. Hsu, and C. Lin, “A Comparison of Methods for 

Multiclass Support Vector Machines.” IEEE transactions 

on neural networks, vol. 13, no. 2, pp. 415-425, March 

2002. “doi:10.1109/72.991427” 

[6] T. Inoue S., and Abe, “Fuzzy Support Vector Machines 

for Pattern Classification.” International Joint 

Conference on Neural Networks, pp.1449-1454, July 

2001. “doi:101109/IJCNN.2001.939575” 

[7] X. F. Jiang, Z. Yi, and J. C. Lv, “Fuzzy SVM with a new 

fuzzy membership function.” Neural Computing and 

Applications, vol. 15, no. 3-4, pp. 268-276, 2006. 

“doi:10.1007/s00521-006-0028-z” 

[8] W. Yan, and Q. He, “Multi-class Fuzzy Support Vector 

Machine Based on Dismissing Margin.” Proceedings of 

the Eighth International Conference on Machine 

Learning and Cybernetics, vol. 2, pp. 1139-1144, July 

2009. “doi:10.1109/ICMLC.2009.5212368” 

[9] C. F. Lin, and S. D. Wang, “Fuzzy Support Vector 

Machines.” IEEE transactions on neural works, vol. 13, 

no. 2, pp. 464-471, March 2002. 

“doi:10.1109/72.991432” 

[10] B. Jin, Y. C. Tang, and Y. Q. Zhang, “Support vector 

machines with genetic fuzzy feature transformation for 

biomedical data classification.” Information Sciences, vol. 

177, pp. 476-489, 2007. “doi:10.1016/j.ins.2006.03015” 

[11] Q. Wu, and R. Law, “Fuzzy support vector regression 

machine with penalizing Gaussian noises on triangular 

fuzzy number space.” Expert Systems with Applications, 

vol.37, no. 12, 2010. “doi:10.1016/j.eswa.2010.04061” 

[12] Z. Pawlak, “Rough sets.” International Journal of 

Parallel Programming, vol. 11, no. 5, pp. 341-356, 1982. 

“doi:10.1007/BF01001956” 

[13] J. H. Zhang, and Y. Y. Wang, “A Rough Margin based 

Support Vector Machine.” Information Sciences, vol. 178, 

pp. 2204-2214, 2008. “doi:10.1016/j.ins.2007.12.012” 

[14] B. Scholkopf, A. J.Smola, R. C. Williamson, and P. L. 

Bartlett, “New Support Vector Algorithms.” Neural 

Computation, vol. 12, no. 5, pp. 1207–1245, 2000. 

“doi:10.1162/089976600300015565” 

[15] C. C. Chang, and C. J. Lin, “Training v-Support Vector 

Classifiers: Theory and Algorithms.” Neural 

Computation, vol. 13, pp. 2119–2147, 2001. 

“doi:10.1162/089976601750399335”


[16] P. H. Chen, C. J. Lin, and B. Scholkopf, “A Tutorial on 

[nu]-Support Vector Machines.” Applied Stochastic 

Models in Business and Industry, vol. 21, no. 2, pp. 111- 

136, 2002. “doi:10.1002/asm.537” 

[17] C. L. Blake, and C. J. Merz, UCI Repository of machine 

learning databases 

[http://www.ics.uci.edu/~mlearn/MLRepository.html]. 

Irvine, CA: University of California. Department of 

Information and Computer Science, 1998. 

[18] R. D. King, C. Feng, and A. Sutherland, “Statlog : 

Comparison of classification algorithms on large realworld 

problems.” Applied Artificial Intelligence, vol. 9, 

no. 3, pp. 289-333, 1995. 

“doi:10.1080/08839519508945477” 

[19] T. K. Ho, and E. M. Klernberg, “Building projectable 

classifiers of arbitrary complexity.” Proceeding of the 

13th International Conference on Pattern Recognition, 

vol. 2, pp. 880-885, 1996. 

“doi:10.1109/ICPR.1996.547202” 


Kai Li, born in Baoding, China, 1963. He received Bachelor 

Degree and Master Degree in mathematics and education 

technology from Hebei University, Baoding, China in 1986 and 

1995, respectively. In 2005, he received PhD degree in 

computer from Beijing Jiaotong University, Beijing, China. His 

research interests include machine learning, neural network, 

pattern recognition, data mining, etc. 

Currently, he is a Professor at school of mathematics and 

computer science, the Hebei University. He has published over 

fifty papers on machine learning, clustering, ensemble learning, 

support vector machine, and pattern recognition. 

Xiaoxia Lu, Born in Shijiazhuang, China, 1984. She received 

Bachelor Degree in computer science from Hebei University, 

Baoding, China in 2009. 

Currently, she is a Master student in the school of 

mathematics and computer science at Hebei University. Her 

research interests are in the areas of support vector machine, 

fuzzy sets and rough sets theory.


A Modified Technique for Analysis of 

Synchronous Counters Constructed with 

Flip-flops 

Dangui Yan 1 

1 College of Mathematics and Physics 

Chongqing University of Post and Telecom 

Chongqing, P.R. China 

Email: yandg@cqupt.edu.cn 

Ruijun Tong 2 , Chengchang Zhang 3 , Changyong Li 4 

2 Deptartment of Electronic Engineering, Chongqing College of Electronics Engineering, P.R. China 

tongrj@163.com 

3 College of Communication Engineering, Chongqing University, Chongqing, P.R. China 

Email:zcc_918@163.com 

4 Chongqing Communication Acadimic of P.L.A. ,Chongqing, P.R. China 

Email: lll_ccc_yyy@163.com 

Abstract—Some methods of fabrication make it 

economically attractive to construct counters (and other 

devices) by connecting sets of identical flip-flops(FFs), if the 

FFs have a common clock input, the state transitions of the 

whole counters are as rapid as the state transitions of each 

FFs, so that the counter is further restricted to be 

synchronous. In order to simplify the process for analysis of 

synchronous counters constructed with flip-flops, a simple 

and successful method is proposed. Using this method, the 

state transition equations obtained from logic diagram of 

counter are converted to standard sum-of-products 

forms(SOPs). By finding out the logic principle for 

achieving the value of logic function based on the standard 

SOPs, the values of next state can be directly obtained 

without any Boolean calculation. Analysis for a 3-bits 

counter shows that this method eliminates complex 

calculations, and makes the process of obtaining next state 

value and developing truth table more rapid and 

convenient. 

Index Terms— counter, synchronous, flip-flops, 

equation, calculation, sum-of-products forms 


We call a device that accepts clock pulses as input and 

that exhibits periodic behavior as output a counter. Some 

methods of fabrication make it economically attractive to 

construct counters (and other devices) by connecting sets 

of identical flip-flops(FFs), such as D FFs, J-k FFs, and 

so on[1-6]. If the FFs have a common clock input, the 

state transitions of the whole counters are as rapid as the 

state transitions of each FFs, so that the counter is further 

restricted to be synchronous[7-10]. These counters may 

be clocked at a maximum rate counter for they have no 

gates or ripple effects to introduce delays. Thus, a 


doi:10.4304/jcp.6.9.1971-1975 

synchronous counter can operate at a much higher input 

frequency and have numerous well-known uses in digital 

apparatus[11,12]. 

For a given synchronous counter constructed with FFs, 

in order to know its logic function, there are two basic 

messages should be obtained by analyzing the logic 

diagram, one is to obtain modulus value of the counter, 

the other is to know whether the counter can self-starting. 

The first step for analysis is to get the state transition 

equations from counter’s logic diagram, the following 

step is to develop truth table and state transition table, 

from which one can obtain information about modulus 

value and self-starting. The key to develop truth table is 

how to obtain the next state values. Present method uses 

substitution method[13-15], in which the values of 

present state are substituted into the state transition 

equations and the next state values are obtained by 

Boolean calculation. The disadvantage of this method is 

that it needs a large number of calculations with timeconsuming 

and error-prone. 

In this paper, we propose a modified method, which 

converts the state transition equations to standard SOPs, 

by finding out the principle for the values of next state, 

the next state values should be directly obtained based on 

the standard SOPs without the need for any calculation. 

Analysis for a 3-bit synchronous counter constructed with 

three J-k FFs flip-flops shows that this method eliminates 

the complex calculations of current method, and makes 

the process for analyzing synchronous counters more 

rapid and convenient. 

II. PRESENT METHOD 

Consider a 3-bits synchronous counter whose logic 

diagram is shown in Fig.1. It uses three J-K FFs-FF0, 

FF1and FF2-and each one has a J and a K input.


& 

& 

CLK 

J 

K 

Q 

Q 

Q0 

& 

K 

J 

Q 

Q 

Q1 

Fig 1. The logic diagram 

& 

K 

J 

Q 

Q 

Q2 

In order to obtain the modulus value of the counter and 

to know whether the counter can self-starting, a general 

procedure is applied in three steps when analyzing this 

counter. 

Step 1: Obtaining the state transition equations 

From the logic diagram of Fig .1, we can obtain the 

following expressions for the J and K inputs of each FFs: 

J = Q + Q, K = Q + Q 

0 2 1 0 2 1 

J = QQ , K = Q 

1 2 0 1 2 

J = QQ , K = Q 

2 1 0 2 1 

According to the characteristic equation of J-K FFs 

[16,17], which is shown as equation (2). 

(1) 

Q′ = J Q + k Q 0 ≤ i ≤ n− 

1 (2) 

i i i i i 

where i Q′ denotes next state, Q denotes present state, 

i 

n is the number of FFs. The expressions of J and i 

K ( 0≤i≤ 2) 

are separately substituted into equation(2), 

i 

we can obtain the state transition equations as shown in 

equations(3): 

⎧Q′ 

= QQQ + Q Q 

2 2 1 0 2 1 

⎪ 

⎨Q′ 

= QQQ + QQ 

1 2 1 0 2 1 

⎪ 

⎩ Q′ = QQQ + QQ + QQ 

0 2 1 0 2 0 1 0 

Step 2: The state truth table 

We develop a truth table for the equations(3) as shown 

in Tab.1, the present state includes three variables 

Q2、 Q and 1 Q in the domain, so there are eight 

0 

possible combinations of binary values of the variables as 

listed in medium columns, which are 000, 001, 010, 011, 

100, 101, 110 and 111, the decimal digit corresponding to 

each binary value is listed in left column, which are from 

0 to 7, the next state is listed in right columns. 

After a CLK pulse input, the FFs enter a new 

state(from present state to next state), assuming initial 

present state is 000, that is, 

Q = 0, Q = 0, Q = 0 ( 2 1 0 000 

QQQ = ), substituting it 

2 1 0 

into equation 2 Q′ of equations(3), we can get 

corresponding value of next state, the process is as follow: 


(3) 

Tab. 1 The truth table 

Q2 Q1 Q0 

' ' ' 

Q 2 Q 1 Q 0 

Q′ = 0⋅0⋅ 0+ 0⋅ 0= 0 

(4) 

2 

Here, we define such process of obtaining value of i Q′ 

as one time next state equation calculation, obviously, it 

perhaps includes NOT, AND and OR operations. Next, 

we can calculate the values of 1 Q′ and 0 Q′. 

Q′ 

= 0⋅0⋅ 0+ 0⋅ 0= 0 

1 

Q′ 

= 0⋅0⋅ 0+ 0⋅ 0+ 0⋅ 0= 1 

(5) 

0 

So, after three times calculations, we obtain the next 

state QQQ ′ ′ ′ = 001, 

and fill the values in correspondence 

2 1 0 

positions in right-hand column of the truth table, which is 

shown in Tab.1. 

Regarding a new present state is 001 and substituting it 

into equations(3), the new next state can be obtained, 

which is 011. Regarding the third present state is 010 and 

substituting it into equations(3) again, the new next state 

can be obtained, which is 110. This process is continued, 

when the next state corresponding to the present state 111 

is calculated, the process is ended. 

It is clearly to know, in order to complete the truth 

table(Tab.1), we need accomplish 24(8*3) times next 

state equation calculations. 

Step 3: The state transition table and the state 

transition diagram 

After obtaining the truth table, the next step is to 

construct the state transition table and the state transition 

diagram. 

Firstly, we construct the state transition table. The table 

is constructed with three columns, which are comments, 

present state and nest state, which is shown in Tab.2. 

Regarding initial present state is 000, observing Tab.1, 

the next state is 001, filling 000 and 001 to corresponging 

column of Tab.2. Regarding the next state 001 is as a new 

present state, the new next state is 011. Regarding the 

third present state is 011 and the third next state is 010. 

This process is continued, in the end, we can observe 101 

is as a present state, which next state is 000, the counting 

cycle is finished, there are six counting states, which are 

000, 001, 011, 010, 110 and 101, we call these states as 

counting states. For next two states, which are 100 and


111, for they are not in counting cycle, we call them as 

offset states. 

Comments 

Counting 

states 

Offset 

states 

Tab.2 The state transition table 

Present State Next State 

Q2 Q1 Q0 

' ' ' 

Q 2 Q 1 Q 0 

0 

0 

0 

0 

1 

1 

0 

0 

1 

1 

1 

0 

0 

1 

1 

0 

0 

1 

0 

0 

0 

1 

1 

0 

0 

1 

1 

1 

0 

0 

1 

1 

0 

0 

1 

0 

1 

1 

0 

1 

0 

1 

0 

1 

0 

0 

1 

0 

Secondly, we can construct the state transition diagram 

based on Tab.2 easily, which is shown in Fig.2. 

Fig.2. The state transition diagram 

Observing Tab.2 and Fig.2, there are six counting 

states, which are 000, 001, 011, 010, 110 and 101, and 

two offset states, they are 111 and 100. These offset 

states can return to counting states after one or two clock 

cycles, obviously, it is a Mod-6 counter and has selfstarting 

performance. 

Summarizing above steps, we can see that the key step 

of this method is to construct the state truth table. It is 

clear that the disadvantage of this method is its large 

calculations during constructing the state truth table . If n 

is the number of state variable, there are n state transition 

equations,which are Q′ , Q′ ,…, n−1 

n−2 

0 Q′ , and 2 n present 

states (from 00…0 to 11…1). In order to create the truth 

table, one must calculate 2 n n times next state equation 

calculations, and every calculation includes NOT, AND 

and OR operations. This procedure seems to be simple 

but in fact it is time-consuming and easy to make 

mistakes. 

III. PROPOSED METHOD 

A. Logic principle 

Any n-variable state transition equation can be 

expressed canonically by the standard SOPs as follows 

[18]. 


n 

j=2 −1 

Q′ i ( Q ... Q ) = a Q Q ∑ ... Q 

n−1 0 j n−1 n−2 

0 

j=0 

= 

= 

where { 0,1} 

n 

j=2 −1 

∑ 

j=0 

n 

j=2 −1 

∑ 

j=0, a j =1 

a m 

j j 

 

j 

, 0≤ p ≤ n − 1, 

p p p 

Σ 

is the OR operator, m j is a minterm, j is the minterm 

number. 

m 

a ∈ , Q ∈{ 

Q , Q } 

n−1 

k 

2 ( Qk 

) 

k=0 

j = ∑ d 

where if Q = Q that dQ ( ) = 1, 

if Q = Q that 

k k 

k 

k k 

d ( Q 

) = 0. 

k 

For any present state Qn−1... Q0 = x n− 1... x k... 

x0 

, 

x ∈ 0,1 , 0≤k ≤ n− 

1. 

Definiting q, 

k 

{ } 

q =2 x + ... 2 x 

n -1 

0 

n− 

1 + 0 

j 

(6) 

The value of next state i Q′ is calculated as follows: 

Q′ ( Q ... Q )= Q′ ( x ... x ) 

i n−1 0 i n−1 

0 

n 

j=2 −1 

∑ 

= a m + a m 

q q j j 

j=0, j≠ q 

n 

j=2 −1 

∑ 

= a .1 + a .0 = a 

q j q 

j=0, j≠ q 

Where if the minterm m is included in the standard 

q 

SOPs that a q is equal to 1, otherwise aq is equal to 0. 

That is, if the minterm corresponding to a present state is 

included in the state transition equation that the next state 

value is equal to 1, otherwise equal to 0. 

As an example, taking into account a 2-bits state 

transition equations as equations(8): 

⎧ 

⎨ 

⎩ 

0 1 0 1 0 

(7) 

Q′ = QQ = ∑m( 00 ) = ∑m( 

0) 

1 1 0 

Q′ = QQ + Q Q = ∑m( 01,10 ) = ∑m( 

1, 2) 

(8) 

For expression Q′ = QQ , which includes one 

1 1 0 

minterm QQ (m0), when present state values 00, Q′ is 

1 0 

1 

equal to 1, for any other present state, such as 01,10,11, 

Q′ = QQ + Q Q , it 

Q′ is equal to 0. For expression 1 

0 1 0 1 0 

includes two minterms QQ (m1) and QQ (m2), when 

1 0 

1 0 

present state values 01 or 10, the value of next state 0 Q′


equals to 1, otherwise equals to 0. All values are showm 

in Fig.3. 

B. Analysis for Fig.1 

Q1 Q0 

' ' 

Q1 Q0 

Fig.3 All values for equations(8) 

By the proposed method, the analysing process for the 

3-bit synchronous counter shown in Fig.1 includes main 

three steps. 

Step 1: The standard SOPs 

Each state transition equation is converted to standard 

SOPs, taking 2 Q′ as an example, the converting processes 

are as shown in Equation.9, this processes can also be 

realized by Karnaugh Maps[19,20]. 

Q′ = Q Q Q + Q Q 

2 2 1 0 2 1 

= QQQ + QQ( Q + Q) 

2 1 0 2 1 0 0 

= QQQ + QQQ + QQQ 

= ∑ m (010,110,111) 

= ∑ m (2,6,7) 

2 1 0 2 1 0 2 1 0 

The conversion to obtain standard SOPs By similar 

processes, the equations Q′ and 1 

0 Q′ can be obtained as 

shown in equations(10). 

Q′ = QQQ + Q QQ + Q QQ 

1 2 1 0 2 1 0 2 1 0 

= ∑ m( 

001,010,011) 

= ∑ m( 

1,2,3) 

Q′ = Q QQ + Q QQ + Q QQ + Q Q Q 

0 2 1 0 2 1 0 2 1 0 2 1 0 

= ∑ m( 

000, 001,100, 110) 

= ∑ m( 

01,4,6 , ) 

Step 2: The truth table 

(9) 

(10) 

(11) 

We develop a truth table for the equations(9), (10) and 

(11), eight possible combinations of binary values are 

listed in the medium columns, the decimal digit 

corresponding to each binary value is listed in the left 

column. According to the principle for the value of next 

state, the present state values that make the next state 2 Q′ 

equal to 1 are 010(2), 110(6), and 111(7). For each of 

these values, a 1 is filled in each corresponding position 

in the first column of three right columns. 


Using the same principle, when the present state values 

are 001(1), 010(2), and 011(3), the next state Q′ equal 

1 

to 1, a 1 is filled in each corresponding position in the 

second column of right columns. When the present state 

values are 000(0); 001(1); 100(4); and 110(6), the next 

state 0 Q′ equals to 1, a 1 is filled in each corresponding 

position in the third column of right columns. After all 1 

are filled, the view of truth table is shown as Tab.3. 

Tab. 3 The truth table 

Q2 Q1 Q0 

' ' ' 

Q 2 Q 1 Q 0 

All the remaining positions in right columns are placed 

by a 0, we can get the same truth table as Tab. 1. 

The third step is same as the present method, detailed 

analysis is no longer given here. 

IV. CONCLUSIONS AND DISCUSSIONS 

This paper deduces the principle for obtaining next 

state values in state transition equations. Based on the 

principle, a modified method has been proposed to 

analyse synchronous counters constructed with flip-flops. 

The method can develop the truth table directly from state 

transition equations with SOPs. This method facilitates 

analysis of synchronous counters constructed with Flip- 

Flops by eliminating large number of Boolean 

calculations. 

The number of state variables n is three in the example, 

if there have more state variables, the calculation will be 

increased, such as n=4, the number of minterm is 16 and 

the state transition equation is 4, it is needed 64 times 

next state equation calculations to accomplish the truth 

table. Clearly, the advantages are more obvious with the 

increase of n. Certainly, if n is enough large, for example, 

n=10, although the proposed method is still more rapid 

and convenient than present one, the obtaining for 

minterm is very complex, so we advise that one should 

analyze with computer. 

Obviously, the method is also suitable for synchronous 

counters constructed with other FFs, such as D FFs, and 

so on. 


The authors thank their colleagues at College of 

Communication Engineering for fruitful discussions. This 

work was supported by Natural Science Foundation 

Project of CQ CSTC under contract no: 2010BB2240.


REFERENCES 

[1] Stan, M.R.; Tenca, A.F.; Ercegovac, M.D. Long and fast 

up/down counters. IEEE Trans on computers, Vol. 47, 

No. 7, July 1998, pp. 722-735. 

[2] ZENG Xiao-pang; WANG Peng-jun. Design of four valued 

synchronous reversible counter based on the theory of 

three essential circuit elements. Journal of Zhejiang 

University(Science Edition), 2009, vol.36(5), pp. 553- 

556. 

[3] YE Xi-en, TAO Wei-jiong , WANG Lun-yao. A low power 

terminary D type flip-flop design based on clock gating 

technique . Journal of Circuits and Systems, 2006, 11(3), 

pp. 106-109. 

[4] Vaquero, A.R.; Aguilo, J.. Gateless synchronous counters 

with D flip-flops. Electronics Letters Vol.14 (16) , 1978 , 

pp. 496-498. 

[5] F. B. Manning. Autonomous, Synchronous Counters 

Constructed Only of J-K Flip-Flops. S.M. thesis, 

available in microfiche from MIT Barker Engineering 

Library or in paperback as Project MAC, Massachusetts 

Inst. Technol.,Cambridge, MA, Tech. Rep. Period 

Counlter Comment TR-96, 1972. 

[6] YE Xi-en, TAO Wei-jiong, WANG Lun-yao. A low power 

ter nary D type flip-flop design based on clock gating 

technique. Journal of Circuits and Systems, 2006, 11(3), 

pp.106-109. 

[7] WU Zhong guang 1; YANG Yu zhi. The Method of Real 

Time and Synchronous Counting for High Speed Multi- 

Event by CPLD. Journal of Sichuan University (Natural 

Science Edition). 2002,vol.39(1), pp. 62-64. 

[8] WU Zhong guang; YANG Yu zhi. The Method of Real 

Time and Synchronous Counting for High Speed Multi- 

Event by CPLD. Journal of Sichuan University (Natural 

Science Edition). 2002,vol.39(1), pp. 62-64. 

[9] Misra, S.K.; Kolagotia, R.K.; Srinivas, H.R.; Mo, J.C.; 

Diamondstein, M.S. VLSI implementation of a 300-MHz 

0.35-µm CMOS 32-bit auto-reloadable binary 

synchronous counter with optimal test overhead delay . 

VLSI Design, 1998. Proceedings. Eleventh International 

Conference on , 1998, pp. 326-329. 

[10] Aguirre-Hernandez, M.; Linares-Aranda, M. A Clock- 

Gated Pulse-Triggered D Flip-Flop for Low-Power High- 

Performance VLSI Synchronous Systems. Devices, 

Circuits and Systems, Proceedings of the 6th International 

Caribbean Conference on, 2006, pp. 293-297. 

[11] Misra, S.K.; Kolagotia, R.K.; Srinivas, H.R.; Mo, J.C.; 

Diamondstein, M.S. VLSI implementation of a 300-MHz 

0.35-µm CMOS 32-bit auto-reloadable binary 

synchronous counter with optimal test overhead delay . 

VLSI Design, 1998. Proceedings., 1998 Eleventh 

International Conference on , pp. 326-329. 

[12] Aguirre-Hernandez, M.; Linares-Aranda, M. A Clock- 

Gated Pulse-Triggered D Flip-Flop for Low-Power High- 

Performance VLSI Synchronous Systems. Devices, 

Circuits and Systems, Proceedings of the 6th International 

Caribbean Conference on , 2006 , pp. 293-297. 


[13] Thomas L.Floyd. Digital Fundamentals. 9th ed. P. 

cm.2004, pp. 398-403. 

[14] Wang Yuyin. Digital circuit and logic design. 3rd ed. 

Higher education press, 2008, pp.181. 

[15] WANG Shi-yuan, XIE Kai-ming, SHI Ya-wei, CHEN 

Meng-gang, LONG Zheng-ji. Implementation of a New 

FPGA-Based Controllable Frequency Divider. Journal of 

Southwest University (Natural Science Edition). 2007, 

vol. 29(1), pp. 89-93. 

[16] Frank B.Manning AND Rober R. Fenichel. Synchronous 

Counters Constructed Entirely of J-K Flip-Flops. IEEE 

Trans on computers, March 1976, pp. 300-306. 

[17] Manning, Frank B.; Fenichel, Robert R. Synchronous 

Counters Constructed Entirely of J-K Flip-Flops . IEEE 

Transactions on Computers, Vol: C-25 (3), 1976, pp. 300- 

306. 

[18] L. Wang and A.E.A. Almani. Fast conversion algorithm 

for very large Boolean functions. Electronics letters, Vol. 

36, No. 16, August 2000, pp. 1370-1371. 

[19] Michel E. Holder. A Modified Karnaugh Map Technique. 

IEEE Trans on education, Vol. 48, NO. 1, February 2005. 

pp. 206-207. 

[20] Dean, K.J.. An extension of the use of karnaugh maps in 

the minimization of logical functions. Radio and 

Electronic Engineer. Vol.35(5) ,1968 , pp.294-296. 

Dangui Yan was born in Luotian, China, 1975. She received 

the BS degree in Department of Mathematics from Hubei 

institutes for nationalities in 1997. She received the MS degree 

in Department of Mathematics from Hubei University in 2000. 

She is currently lecturer of Chongqing University of Post and 

Telecom. Her research interest is logic algebra. 

Ruijun Tong was born in Shanxi, China, 1976. She received 

the MS degree in College of Communications Engineering from 

Chongqing University in 2005. She is currently lecturer of 

Chongqing College of Electronics Engineering. Her research 

interests are digital system and FPGA design. 

Chengchang Zhang was born in Lichuan, China, 1975. He 

received the BS degree in automation engineering from the 

Wuhan Institute of Chemical Technology in 1997. He received 


Chongqing University in 2005. He is currently PhD candidate of 

Chongqing University majoring in Communication and 

information systems. His research interests are software radio 

and FPGA design. 

Changyong Li was born in Chongqing, China, 1971. His 

research interests are software radio and ultra-wide band radar.


A New Method of Detecting Multi-component 

LFM Signals Based on Blind Signal Processing 

1 Qiang Guo 

1 Space Control and Inertial Technology Research Center,Harbin Institute of Technology, P.R. China 

Email: guoqiang292004@163.com 

2 Yajun Li and 1 Changhong Wang 

2 College of Information and Communication Engineering, Harbin Engineering University,P.R.China 

Email: liyajun1985happy@163.com 

Abstract—To effectively detect and recognize multicomponent 

Linear Frequency-Modulated (LFM) emitter 

signals, a multi-component LFM emitter signal analysis 

method based on the complex Independent Component 

Analysis(ICA) which was combined with the Fractional 

Fourier Transform(FRFT) was proposed. The idea which 

was adopted to this method was the time-domain separation 

and then time-frequency analysis, and in the low SNR cases, 

the problem which is generally plagued by noised of feature 

extraction of multi-component LFM signal based on FRFT 

is overcame. Compared to the traditional method of timefrequency 

analysis, the computer simulation results show 

that the proposed method for the multi-component LFM 

signals separation and feature extraction was better. 

Index Terms—multi-component LFM emitter signals, timefrequency 

analysis, feature extraction,ICA 


Radar emitter signal detection is a key problem which 

is demanded to be resolved in modern electronic 

reconnaissance system. With large new complex radar 

systems in practice, a large number of pulses overlap and 

form the multi-component emitter signals(MCES)[1]. 

MCES analysis is a prerequisite and primary task for 

detecting and identifying emitter signals. Multicomponent 

Linear Frequency Modulated (LFM) emitter 

signals is a non-stationary signal which is commonly 

used in active sonar, radar imaging, fuse of missile and so 

on. As a new time-frequency analysis tool, FRFT is a 

generalization of the Fourier transform (FT). It not only 

has a natural link in classical FT, but also provides some 

characteristics which FT do not have. So FRFT is 

specially suitable for processing LFM class (chirp-like) 

signal. At present, regardless of the traditional parameter 

estimation or detection methods of multi-component 

LFM signal, most of them are based on time-frequency 

analysis or all finds of FT method[2]. The parameter 

estimation methods mainly through two-dimensional 

Manuscript received January 2, 2011; revised February 1, 2011; 

accepted February 28, 2011. 

Qiang Guo, Yajun Li,Changhong Wang. 


doi:10.4304/jcp.6.9.1976-1982 

object function, and combined with two-dimensional 

search to estimate, such as the maximum likelihood 

method, the time-frequency analysis methods, FRFT, 

match fourier transform(DCFT)[3,4] ,S-method and so 

on. But when the low SNR, together with the existence of 

weak signal, the traditional detection method is often 

difficult to effectively detect the MCES, even lead to 

misjudgment of signal and noise ,so the result of crossterm 

suppression is not good. Above-mentioned issue has 

been a relatively difficult problem. In this paper, a multicomponent 

LFM emitter signal analysis method based on 

complex FastICA which was combined with FRFT was 

proposed. Firstly, complex ICA algorithm was used as 

time-domain separation for multi-component LFM 

emitter signals with noise. Secondly, determine the signal 

and noise by the automatic identification method of 

second central moment of FRFT. Lastly, the noise was 

removed and the LFM signals were detected by FRFT. 

The effect of noise can be greatly reduced. 

Simultaneously, the cross-terms are effectively deduced 

with higher time-frequency resolution. It is a good 

method for the multi-component LFM signals. 

Simulation results verify the effectiveness of this new 

method. 

II. MODEL OF MULTI-COMPONENT LFM EMITTER SIGNALS 

In modern electronic reconnaissance system, receiver 

often intercepted to pulses which emitted by multiple 

sources at the same time. A stream of pulses was formed 

through these pulses interleaved together. As the pulse 

signals density increases, the pulse formed MCES 

x() t ,the signal model is expressed as follows. 

k −1 

∑ 

i= 

0 

i 

2 

j2 π ( fit+ ( µ it 

/2)) , −∆t/2 ≤ t ≤ ∆ t/2 

x() t = Ae + n() t 

(1) 

Where Ai is the amplitude of each signal, fi is initial 

frequency and µ i is chirp rate. nG() t is White Gauss 

Noise with zero mean and variance 2 

σ . 

III. MULTI-COMPONENT LFM EMITTER SIGNALS 

ANALYSIS


In this paper, the flowchart of proposed multicomponent 

LFM signals analysis method is shown in 

Fig.1. Firstly, complex FastICA algorithm is used as the 

pretreatment of detection of multi-component LFM 

emitter signals. The emitter signals and noise were 

separated by it. Secondly, discriminate LFM signal and 

noise based on the automatic identification method of 

second-order central moment of FRFT. Then detect the 

separated LFM signals through FRFT after removing the 

noise. Through the above steps, the impact of noise was 

effectively reduced and the cross terms were effectively 

restrained. Last, the LFM signal parameter estimation can 

be completed through the conversion formula [1,5]. 

Figure 1. Multi-component LFM emitter signals analysis with low SNR 

A. Pretreatment 

1)Problem Description 

As the number of mixed-signal (decided by the number 

of channel) received by the radar signal receiver and do 

not necessarily match the number of radar emitter signals, 

and LFM signal is non-stationary signal, so the complex 

ICA of the blind signal processing techniques was used in 

multi-component LFM signals with low SNR to do timedomain 

separation pretreatment. 

At present, most of the FastICA algorithm is mainly 

concentrated in the real domain in blind separation of 

mixed signals[7], but in practice, many real signal model 

is represented by linear mixed model of complex signals. 

Blind separation algorithm for complex signals is 

summarized into the methods based on higher order 

statistics and the methods based on second order statistics. 

It is actually an optimization problem, namely, how to 

make the separated independent component to maximum 

approach the source signal. Here we will extend the real 

domain variable to the complex domain[8,9]. 

Taking into account the observed mixed-signal is 

instantaneous linear mixing of the each source signal, the 

standard ICA linear model with noise as follows [3] 

n 

xi() t = ∑ aijsj() t + ni() t ( i = 1,2, ⋅⋅⋅ , m) 

(2) 

j= 

1 


iii 

iii 

iii 

Expressed in matrix form, ie. 

X = A S+ n 

(3) 

where X = ( x1, x2,..., xm) 

is the vector of observed 

random variables, S = ( s1, s2,..., sn) 

is the vector of 

statistically independent latent variables called the 

independent components, and A is an unknown constant 

complex mixing matrix. The above model is identifiable 

under the following fundamental restrictions: 

① At most one of the independent components s j may 

be Gaussian. 

② The matrix A must be of full column rank. The 

number of observing signals m is more than the number 

of source signals n . ( m≥ n) 

,here m=n. 

③ The various components of the source signals 

si ( i = 1,..., m) 

and observed signals x i are zero-mean and 

unit variance. 

In addition, the noise itself can be regarded as a source 

of signal to use BSS, and thus make the algorithm have a 

wider scope and greater robustness. 

Can be proved that we can find a matrix W by linear 

transformation to do m mixed-signal X , making between 

the each component of new vector Y obtained by 

transformating the X as independent as possible in the 

case m≥ n , that is 

H 

Y = W X 

(4) 

where, Y is the separated vector signal, that is, the 

estimated value of source signal vector S . 

Complex FastICA algorithm is a fixed-point algorithm 

using the batch processing. Compare with the ordinary 

ICA algorithm its convergence speed is more quickly. In 

this paper, we adopt complex FastICA algorithm in the 

pre-processing, in order to separate the multi-component 

LFM signals of low SNR in time-domain. 

2)Time-domain separation technique of multi-component 

LFM signals with low SNR based on complex FastICA 

Multi-component LFM time-domain separation 

technique of basic complex FastICA algorithm includes 

two steps: first, preprocess the chosen mixed-signal X 

which was composed of multi-component LFM signals 

and noise, that is, using whitening treatment. Whitening 

treatment can be used to remove the correlation between 

signals, which simplifies the process of follow-up to the 

extraction of independent component, second is the 

extraction of independent components, namely, the 

completion of the mixed signals separation. Complex 

FastICA algorithm flow chart is shown in Fig.2. 

Figure 2. Algorithm flow chart of complex FastICA 

Complex FastICA is a fast optimization iterative 

algorithm using the batch approach which has a large 

number of samples of data involved in computation in 

each iteration. According to the central limit theorem, 

linear sum of a number of independent random variables 

will tend to Gaussian distribution, so complex FastICA 

mainly achieve the purpose of separation by measuring


the largest non-Gaussian. For complex random vector y , 

its kurtosis is defined as 

4 * * * * * * 

kurt( y) = E{ y } −E{ yy } E{ yy } −E{ yy} E{ y y } − E{ yy } E{ y y} 

(5) 

As 

kurtosis can be converted to[2] 

* 

y is y conjugate transpose, the definition of 

4 2 2 2 

kurt( y) = E{ y } −2( E{ y }) − E{ y } 

4 

= E{ y } − 2 

(6) 

where y is white, i.e., the real and imaginary parts of 

y are uncorrelated and their variances are equal. 

We usually use some other suitable non-linear 

function Gy ( ) instead of y for (6), which makes the 

convergence calculation of kurtosis is more robust. The 

expectation function of separation matrix is expressed as 

2 

H 

2 

J ( W) = E{ G( Y )} = E{ G( 

W X )} (7) 

G 

1 

where G ( y) = log( a + y) 

, g ( y) 

= was choosen 

a+ y 

as non-linear function[3,15]. Where a is arbitrary constant 

for which values a ≈ 0.1 were chosen in this work. 

Because the above non-linear function give more robust 

estimators, g ( y) is derivative of G ( y ) . 

Now give the fixed-point algorithm for complex 

signals under the ICA data model(3). In this paper ,we 

obtain separation matrix W , which makes the separation 

H 

components Y = W X , so the estimation Y of source 

independent component were obtained. The need for 

preprocessing to reduce the difficulty of analysis due to 

less known information in blind separation. So the 

whitening process of the observation signal will greatly 

simplify the analysis difficulty. Firstly, centralize the 

mixed-signal X of multi-component LFM signal and noise, 

that is X = X−E{ X} 

so that the mean of X is 0. Then we 

can obtain zero mean vector X by observational data and 

whitening matrix Q ,i.e. X = Q Xold 

, 

H 

X = ( x + ix ,..., x + ix 

) ,and therefore E { XX } = I . 

1r1i nr ni 

Whitening can be accomplished by principal component 

analysis (PCA). 

The complex FastICA algorithm searches for the 

2 

extrema of EG { ( )} 

H w X .Details of the derivation are 

presented in the appendix[3] . Supposing the separation 

matrix W , first select an initial separation vector w 

(random). The fixed-point algorithm for one unit is 

+ H * H 

2 

H 

2 

w = E{( x w x) g( w x )} −E{( 

g w x ) 

w 

H 

2 

H 

2 

' 

+ w x g ( w x )} w (8) 

new 

= 

w 

+ 

+ 

w (9) 

The one-unit algorithm can be extended to the 

H 

estimation of whole ICA transformation S = W X .To 

prevent different neurons from converging to the same 


2 

H H 

maxima, the outputs W X,..., X 

1 W are decorrelated after 

n 

every iteration. A simple way to accomplish this is a 

deflation scheme based on a Gram-Schmidt-like 

decorrelation: When we have estimated p independent 

components, or p vectors w ,..., w ,we run the one-unit 

1 p 

fixed-point algorithm for w new ,and after every iteration 

step subtract from wnew the projections of the previously 

estimated p vectors, and then renormalize wnew as follows 

= 

p−1 

−∑ 

j = 1 

j j pnew 

H 

w w w w w (10) 

w 

pnew p 

pnew 

= w 

w 

pnew 

pnew 

(11) 

where w ( j = 1,..., p−1) 

is previous p − 1 separation 

j 

vector, wnew denotes the p new value of separation vector. 

Determining whether w is convergence. If not 

pnew 

convergence, the w obtained by (9) instead of the w in 

pnew 

(9) and instead of the w in (10) up to the time when 

p 

w convergence, therefore the p separation vector is 

pnew 

obtained. Sometimes it is preferable to estimate all the 

independent components simultaneously, and use a 

symmetric decorrelation. This can be accomplished e.g., 

by 

1/ 2 

( ) − H 

w = w w w (12) 

where W = ( W,..., W ) is the matrix of the vectors. 

1 

n 

At this time we can get mixing matrix Α and separation 

matrixW . Then the separation signal y1, y2, ⋅⋅⋅, 

yncan be 

H 

calculated according to Y = W X . Until now, we have 

completed the mixed-signal time-domain separation 

process of multi -component LMF signals with 

noise[10,11]. 

B.The automatic identification method based on second 

central moment of FRFT 

Multi-component LFM emitter signals after blind 

source separation, the signal has a random arrangement, 

in the case of unknown a priori information can not 

determine which way is the signal or noise. If use the 

FRFT of noise to estimate the parameter of LFM signal, 

the measuring results will be wrong. To solve the 

problem, we use the second central moment of FRFT 

method to achieve the above separated signal and noise 

auto-discrimination. 

FRFT is the promotion of the traditional Fourier 

transform. In recent years, it has attracted increasing 

attention in signal processing field. Setting one of the 

output signals after above-mentioned complex FastICA 

algorithm is yi() t . Its FRFT is defined as 

X ( u) = ∫ y ( t) K ( t, u) dt 

(13) 

α i α


where K (, tu) 

is kernel function of FRFT, is defined as 

α 

2 2 

⎧ 1−jcotαt + u ut 

⎪ exp( j cot α − j ), α ≠nπ 

⎪ 2π 2 sinα 

⎪ 

Kα(, t u) = δ( t− u), α = 2nπ 

(14) 

⎨ 

⎪ δ( t+ u), α = (2n+ 1) π 

⎪ 

⎪⎩ 

where α is rotation angle. With a view of timefrequency 

plane rotation to explain, then the following 

equations are established 

0 

R ( u) = y ( t) 

yi π 

i 

yii R ( u) = y ( − t) 

π /2 

R ( u) = FT( y ( t)) 

(15) 

yii π /2 

where Ry( u) 

corresponds to the FT of signal y () 

i 

i t . 

The traditional FT can be seen as the time-frequency 

distribution of signal in the projection of frequency axis, 

while FRFT can be seen as the time-frequency 

distribution of signal in the projection of the rotated 

frequency axis. The representation of signal in the 

fractional Fourier domain includes both the time domain 

and frequency domain information, so FRFT is also 

considered a generalized time-frequency analysis[13,14]. 

By the definition of FRFT, a LFM signal only at the 

appropriate fractional domain is an impulse function. 

Therefore, FRFT in a fractional Fourier domain has the 

best gathering characteristics for the given LFM signal. In 

the time-frequency plane, a limited length of the LFM 

signal appears as the distribution of dorsal fin shape of 

diagonal line, but FRFT is essentially the "rotating" of 

signal. If choose the appropriate rotation angle, it will 

show the energy aggregation and apparent peak in the 

fractional Fourier domain of signal. It was shown in Fig.3. 

| Xp( u)| 

Figure 3 .The distribution of time-frequency and in the projection of 

fractional Fourier domain of LFM signals 

The bandwidth of signal in time domain and frequency 

domain can be estimated by the second-order central 

moments[15], and the bandwidth of signal in the 

fractional Fourier domain can be obtained by the secondorder 

central moments of FRFT[16]. The second-order 

central moments(SCM) of FRFT Pα is defined as 

∞ 

2 

α 

2 

Pα = ∫ Ry( u) ( u−m ) du 

−∞ i 

α 

(16) 


where 

α 

Ry( u) 

i 

is FRFT of yi() t , mα is first-order 

moments of FRFT 

m 

∞ 

= ∫ 

2 

α 

R ( u) udu 

(17) 

α 

−∞ 

yi 

As FRFT is a periodic function with the period of 

α+ π α 

2π about α , and meet R ( u) = R ( − u) 

, so the 

yi yi 

second-order central moments of FRFT Pα has a 

maximum or minimum value in the range of α ∈ [0, π ) . 

As Pα represents the bandwidth of signal in the fractional 

Fourier domain, when the rotation angle of timefrequency 

planeα = αe 

, the bandwidth has a minimum. 

We can find spindle direction of time-frequency 

distribution α by searching the minimum point of Pα , 

namely, the best fractional Fourier transform domain. The 

bandwidth of noise is wide in the fractional Fourier 

transform domain, α = αe 

corresponds to the minimum of 

bandwidth (the minimum of FRFT) also very large, 

namely, Pα = α corresponds to the minimum. So we can 

e 

determine signal or noise by the bandwidth of fractional 

Fourier transform domain. 

The noise can be removed from the separated signals 

after the LFM signal and noise discrimination method 

based on second-order central moments of FRFT. Then 

only detect the remaining LFM signal. 

C. FRFT detection for LFM signal 

As shown in Fig.3, the observed signal was 

continuously proceed FRFT for rotation angle variable 

α , the two-dimensional distribution of signal energy was 

formed in the parameter ( α , u) 

plane [14]. And the 

detection and parameter estimation of LFM signals can 

be realized by two-dimensional search of peak point 

threshold in this plane. For type (1), the process of this 

model can be described as[16] 

∧ ∧ 

2 

{ α0, u0} = arg max X ( u) 

α 

α , u 

(18) 

⎧ 

⎪ 

⎪ 

∧ ∧ 

⎪ 

µ 0 =−cot 

α 0, 

⎪ 

⎪ ∧ ∧ ∧ 

⎨ f0 = u0 

csc α 0, 

⎪ 

∧ 

⎪ X ∧ ( u0 

) 

∧ 

⎪ 

α0 

⎪ 

Ai 

= 

⎪ ∆tA 

∧ 

⎩ 

α0 

(19) 

IV. SIMULATION VERIFICATION 

Select a group of mixed-signal in order to verify the 

validity of the method proposed in this paper. Mixedsignal 

composed of two-component LMF signal and 

noise, the first LFM signal is 

2 

− j5π t 

x1= e (initial 

f = 0 and chirp rate k 10 =− 10 ), the second is 

frequency 10


x2 2 

− j 2π 

t 

0.5* e 

f 20 = 0 and chirp rate 20 4 

weak LFM signal 

= (initial frequency 

k =− ), the noise is Gaussian 

white noise, the SNR of the weakest component LMF 

signal(the second LFM signal and noise ratio) is SNR=- 

15dB. Nonlinear function selected G ( y) = log( a + y) 

, 

here a ≈ 0.1 , sampling point N=1601, the order of 

FRFT is 0

Using the new method to simulate and analyse the 

above assumed mixed-signal. The multi-component 

signals with noise were separated by complex FastICA 

algorithm. The time-domain separation results were 

shown in Fig.4(a-b). 

(a) The results of mixed-signal by the complex FastICA (taking the real 

part of the signal) 

(b)The convergence of complex FastICA algorithm for Multicomponent 

LFM emitter signals 

Figure.4. Multi-component LFM emitter signals separated by complex 

FastICA 

As can be seen from Fig.4(a), the effect of signal 

separation was very good in the low SNR(-15dB) by the 

complex ICA algorithm. Fig.4(b) shows the convergence 

of the fixed-point algorithm using contrast function 

G( y) = log( a+ y) 

, average result over ten runs. About six 

iteration steps were needs for convergence. Fig.5 (a-b) 

shows the second-order central moments of FRFT of the 

each separated signal( 0

α ∈ [0, π ) ). 


(a) Second-order central moments of time-domain separated output 

signal 1 

(b) Second-order central moments of time-domain separated output 

signal 2 

(c) Second-order central moments of time-domain separated output 

signal 3 

Figure.5. Second-order central moments of FRFT of the each separated 

signal 

According to (16) , the change results of second-order 

central moments P α with FRFT angle α were shown in 

Fig.5 (a-c) . The minimum of second-order central 

moments (ie, bandwidth of signal) of three separated 

5 

signals is respectively Pα 1 = 4.671*10 , 

6 

6 

Pα 2 = 1.2001*10 and Pα 3 = 1.3201*10 by computer 

calculation. So the third signal is Gaussian white noise 

which to be filtered. Next only the remaining two LFM 

signals were detected by FRFT. Fig.6(a-b) shows the 

FRFT time-frequency map of remaining two signals. 

Fig.7 shows the time-frequency map of traditional FRFT


of the same group of mixed-signal without the complex 

FastICA(SNR=-15dB). 

(a) The FRFT of one LFM signal after automatic identification method 

(b) The FRFT of the other LFM signal after automatic identification 

method 

Figure.6. The FRFT time-freqency map of remaining two signals after 

automatic identification method 

Figure.7. Distribution of traditional FRFT of the same group of mixedsignal 

without the complex FastICA algorithm(SNR=-15dB) 

From Fig.6 and Fig.7 can be seen in SNR=-15dB, 

when the mixed-signal was processed by the new method, 

the FRFT distribution of the separated signal was 

influenced by noise slightly and cross-term has also been 

suppressed. However, when the same mixed-signal was 

processed by traditional FRFT, the signals were 

influenced by noise and cross-term largely, weak signal 

has been drowned by the noise and the noise would cause 

great difficulties in the extraction and detection of 

signals. Especially,the weak signal is more affected by 

it[13]. 


Using the method described in section (C) to 

estimate the parameters. Parameter estimation results in 

the following Tab.Ⅰ: 

Table Ⅰ 

PARAMETER ESTIMATION RESULTS OF TWO METHODS 

SNR=- 

15dB 

Signal 

1 

Signal 

2 

Real Value 

10 

The new method 

(FastICA&SCM& 

FRFT) 

Test results 

Traditional 

method 

(FRFT) 

f =0 0.0012 0.0102 

10 =-5 µ 

-5.00757 -5.9894 

f 20 =0 0.0007 2.0691 

20 =-2 µ -2.00133 -3.311 

Above the table can be seen the new method can 

correctly estimate the parameters of the LFM signal in 

SNR=-15dB. For the signal 1, the relative error of the 

estimate value was η f = 0.24% and η 0.342% 

10 

µ = (the 

10 

relative error of the estimate value was expressed as 

η f and η µ ). For the signal 2, the relative error of the 

estimate value was η f = 0.35% and η 0.356% 

20 

µ = . 

20 

However, the traditional FRFT method is no longer 

correctly estimate the signal parameters in SNR=-15dB. 

V. CONCLUSION 

To effectively detect and recognize multi-component 

LFM signals in low SNR, a new multi-component LFM 

signals analysis method which was based on the complex 

FastICA&FRFT was proposed. First, multi-component 

LFM signals were processed by complex ICA to obtain 

the time domain separate signals in low SNR. Second, the 

time domain separated signals were respectively 

discriminated by automatic identification method based 

on second central moment of FRFT. Then the LFM 

signals were processed by the FRFT. In this paper, the 

new method was compared with the traditional FRFT to 

prove the validity of the new method. The simulation 

results show that the new method can effective analysis 

multi-component LFM emitter signals in low SNR. 


We thank the National Natural Science Foundation 

Project (No.:60872108), China Postdoctoral Science 

Foundation Special Support Project (No.:200902411) , 

the financial support from China Postdoctoral Science 

Foundation (No.:20080430903), Heilongjiang Postdoctor 

Financial Assistance (LBH-Z08129), the Scientific and 

Technological Creative Talents Special Research 

Foundation of Harbin Municipality (2008RFQXG030) 

and Central University Basic Research Professional 

Expenses Special Fund Project (No.:HEUCFZ1015) for 

this paper support. We thank members of College of 

Information and Communication Engineering, Harbin


Engineering University and Space Control and Inertial 

Technology Research Center, Harbin Institute of 

Technology for technical support. 

REFERENCES 

[1] Liu feng, Sun dapeng, Huang yu, Tao ran, Wang yue,in 

:“Multi-component LFM signal feature extraction based on 

improved Wigner-Hough transform,” Journal of Beijing 

Technology University.(2008.10) 

[2] Ashok Narayanan V, Prabhu K M M. The fractional 

fourier transform:theory, implementation and error 

analysis[J]. Elsevier Micorprocessors and Microsystems, 

Vol.27(2003),p.511-521 

[3] Ella Blngham and Aapo Hyvarinen. “A Fast Fixed-Point 

Algorithm For Independent Component Analysis Of 

Complex Valued Signals”. International Journal of Neural 

Systems, Vol.10, No.1(February,2000),p.1-8 

[4] J.Herault and C.Jutten.Blind separation of sources, part: an 

adaptive algorithm based on neuro mimetic. Signal 

Processing, Vol.24(1)(1991),p.1-10. 

[5] LIU Q S , LU H Q , MA S D, “A Non-parameter Bayesian 

Classifier for Face Recognition [J] ,”Journal of Electronics 

(China), Vol.20(5)(2003),p.362 -370. 

[6] Shimizu S., Hyvarinen A., Kano Y.. A generalized least 

squares approach to blind separation of sources which have 

variance dependencies[J].Statistical Signal Processing, 

IEEE/SP 13th Workshop on(2005),p.1080-1083 

[7] Tachibana K., Saruwatari H., Mori Y.. Efficient Blind 

Source Separation Combining Closed-Form Second-Order 

ICA and Nonclosed-Form Higher-Order ICA. IEEE 

International Conference on Acoustics,Speech and Signal 

Processing. ICASSP 2007. Vol. 1(2007), p.I-45-I-48 

[8] Chee-Ming Ting, Salleh S.-H., Zainuddin Z.Z.. Spectral 

Estimation of Nonstationary EEG Using Particle Filtering 

With Application to Event-Related Desynchronization 

(ERD) [J]. IEEE Transactions on Biomedical Engineering. 

Vol. 58(2011) p.321-331 

[9] Zou Hong-xing, LU Xu-guang, DAI Qiong-hai. 

Nonexistence of cross-term free time-frequency 

distribution with concentration of Wigner-ville 

distribution, Vol.3(2002) 

[10] Yuan junquan,Sun minqi,Sun xiaoxu, “LFM signal 

parameters estimation method based on Wigner Hough 

Transform”,Aerospace Electronic Countermeasures, Vol.6 

(2004). 

[11] Solvang H.K.,Nagahara Y.,Araki S.. Frequency-Domain 

Pearson Distribution Approach for Independent 

Component Analysis (FD-Pearson-ICA) in Blind Source 

Separation[J]. IEEE Transactions on Audio,Speech,and 

Language Processing . Vol.17,No.4(2009),p.:639-648 

[12] Liu ju, He zhenya, Zhang xianda. Blind Source Separation 

and Blind Deconvolution. Electronics Journal, Vol.30(4) 

(2002),p.570-576 

[13] Li xiaoju,Zhu xiaolong,Zhang xianda. Blind source 

separation classification and prospects. Journal of Xi'an 

University of Electronic Science and Technology, 

Vol.31(3) (2004),p.399–404 

[14] Zhang xianda, Bao zheng. Blind signal separation. E- 

Journal. Vol.29(12)( 2001),p.1766-1771. 

[15] Zou hong. Time-frequency analysis of multi-component 

LFM signals [D]. Xi'an Electronic Science and Technology 

University, 2000. 

[16] Liu Jiancheng, Wang Xuesong, Xiao Shunping, et a1. 

Radial acceleration estimation based on Wigner-Hough 

transform[J]. Acta Electronica Sinica, Vol.33(12) 

(2005),p.2236-2238. 


Qiang Guo was born in 1972. He 

received the B.S., M.S. and Ph.D. degree 

from Harbin Engineering University in 

information and communication 

engineering in 1994, 2003, and 2007, 

respectively. He is now an associate 

professor of information and 

communication engineering at Harbin 

Engineering University and a postdoctoral 

fellow of control science and engineering 

at Harbin Institute of Technology (HIT), China. His current 

research interests include radar signals sorting and recognition. 

More complex and dense pulses environments in modern 

electronic warfare present a severe challenge to the problem of 

radar signal sorting. Based on fractal theory and Hilbert-Huang 

Transform (HHT), he presented a new feature extraction 

method for radar pulse signal sorting. It used structure function 

and empirical mode decomposition to process 2-dimension 

feature information, which constituted carrier frequency and 

time-of-arrival. The same scheme also applied to the analysis 

and extraction of hidden periodically changing features—G 

features. Experiment results show that the method can 

effectively identify the agile frequency in periodically changing 

radio frequency signals of complex pulse environment, 

therefore provides a new feature for signal sorting of interleaved 

radar pulse sequences. 

He received the national 100 excellent doctor degree 

dissertation candidate nomination in 2009. He is now Academic 

Degree & Graduate Education Evaluating expert of MOE. 

Yajun Li was born in 1983. He 

respectively received the B.S., M.S. degree 

from YaTai University and Harbin 

Engineering University in information and 

communication engineering in 2008, 2011. 

He is now an Ph.D. at Harbin Institute of 

Technology (HIT), China. 

His current research interests include 

radar signals sorting and detection. At 

present, he has already published seven articles(EI index). 

Changhong Wang was born in 1961. He 

received the B.S. the M.S. and Ph.D. degree 

from Harbin Institute of Technology in 

1983, 1986, and 1991, respectively. He is 

currently a professor and Ph.D. student 

supervisor of Harbin Institute of 

Technology. 

His research interests are mainly in 

inertial navigation, precise servo control 

system, and robust control.


Research on Self-built Digital Resource Backup 

Systems 

Abstract – This paper discussed the characteristics of the 

self-built digital resources and the requirement for 

long-term preservation. A backup system for self-built 

digital resources has been proposed based on the software 

and hardware features of the resources. Furthermore, 

simple analysis has been carried out on the proposed 

system. 

Index Terms – self-built digital resource, backup system, 

long-term preservation, data duplication elimination 


Digitalized preservation, organization and sharing are 

very important in digital resources construction. 

Self-built digital resources is an important part in digital 

resources construction. Self-built digital resources often 

appear in the form of feature databases self-created by 

libraries, for example, numerous well developed feature 

databases, dissertation databases and textbook databases 

are planned in library projects. Furthermore, different 

schools and universities also construct their own 

databases based on their discipline and regional 

characteristics, technical specialty, as well as financial 

budget [1]. 

In order to preserve self-built digital resources for long 

time, two aspects have to be considered, namely, how to 

prevent unauthorized modification and breach to digital 

information, and how to maintain long-term readability 

and authenticity of digital information. Technology is 

readily available to tackle the first problem, as a number 

of mature techniques have been proposed world widely to 

prevent illegal modification and breach of digital 

information; therefore, it is possible to solve the first 

problem to some extent if technical measures can be 

scientifically integrated with management practice. 

However, how to effectively maintain long-term 

readability of the digital information is still an open 

research area, no perfect solution has been proposed so 

far. The major difficulties lie on the deep involvement of 

numerous issues in which the most important one is the 

adoption of standards. Adopting standards can ease the 

conflicts between the technological update and 

readability of digital information. Nevertheless, problems 

still remain as some standards, particularly industrial 

standards are commonly outdated; and it is also difficult 

to completely comply with standards in practice. 

Currently, techniques used for long term preservation 


doi:10.4304/jcp.6.9.1983-1987 

Li-zhen Shen 

Wenzhou University, Wenzhou, Zhejiang, China 

a wzu-slz@wzu.edu.cn 

include migration technique, updating technique, 

conversion technique, simulation technology, and 

digitizing technique using graphic tablets, etc. [3]. 

II. CHARACTERISTICS OF BACKUP SYSTEMS FOR 

SELF-BUILT DIGITAL RESOURCES 

Self-built digital resources are diverse, including such 

resource types as WEB resources, electronic publications, 

scientific data, multimedia resources and electronic 

dissertations, etc. Furthermore, all self-built digital 

resources use internet to provide resource services, 

therefore, in addition to fully back up the content of 

databases, it is often required to backup server systems 

and data publishing environment (both will be referred as 

servers in the sequel). The servers which need to be 

backed up may run multiple operating systems such as 

SUN Solaris, LINUX, Windows NT, and Windows 2000, 

and some may have Microsoft SQL Server 2000 database 

and ORACLE running on them. Considering all these 

aspects, backup systems should have the following 

functions: 

Backup across operating systems. Backup systems 

should support data backup and recovery across different 

operating systems such as Microsoft Windows, Unix, and 

IBM Aix, that is, a backup server can back up data from 

multiple operating platforms, thus reduce operational 

complexity and lower total cost of the backup work. 

Automatic Backup. Backup uses system resources. In 

practice, a running backup job may take 60% of the CPU 

resource of a mini-computer server with average 

configuration. Besides, backup jobs will also occupy 

network bandwidth as well as other resources. Therefore, 

backups should be performed when the load of the 

servers is minimal, should avoid casting extra load on the 

servers in the peak hours. Obviously, it is essential to use 

unattended, automated backup systems to avoid the 

human interactions during the backup time which, most 

likely, happens at late nights or in public holidays. 

Support multiple backup strategies. The famous 

Pareto principle holds in the backup area, namely, 20% of 

the data is updated more frequently with a back up 

probability of 80%. If every time a full backup is 

performed, it will inevitably waste resources and time in 

some cases, thus full backup is sometimes not viable. 

What we need the daily backup to do is to backup the 

delta of two consequential full backups. Therefore, we 

should adopt the so-called incremental backup strategy


and a combination of several other backup strategies. 

Meanwhile, it is necessary to consider the requirement of 

long-term preservation when backing up data. 

Efficient and safe recovery. The fundamental purpose 

of backup is for recovery, a backup which cannot be 

restored is meaningless. An important factor which the 

end users will use to determine the quality of a backup 

system is whether the system can restore the backed up 

data in a safe, convenient and efficient way. 

Easy upgrading. It is necessary to consider possible 

future extension of functions when designing the system 

initially. For example, the designed system should be able 

to support database online backup, and should be easy to 

add functions on the client-side. 

Long-term preservation. Long-term preservation of 

electronic resources is an important task for library 

resources construction in the new information 

environment. Long-term preservation is not only a new 

mission for the libraries, but also a major challenge as 

many technical, economic, legal and other problems 

emerge. Regardless of changes in the external 

environment, it is an essential characteristic for modern 

backup systems to effectively preserve data over long 

term, and to guarantee readability of the preserved data at 

any time. 

Based on the analysis of requirement for self-built 

backup system resources, and considering multiple 

factors such as unified backup management and support 

for future storage infrastructure, we propose a backup 

system which will be described in the sequel. 

A. Backup system structure 


III. DESIGN AND ANALYSIS 

Fig. 1 A self-built backup system for digital resources 

The proposed backup system includes self-built 

featured database reservoir, primary storage, application 

server farms, backup / media servers, file management 

application servers, and remote archive storage for 

disaster recovery. 

The backup system works as follows: A server is 

configured as a backup server which is responsible for 

system backup operation; a large capacity backup storage 

device, which consists of low-end storage, tape drive or 

tape array, is connected to the backup server. Other 

servers within the network which may need to have 

managed data backup will run backup client software 

which enable centralized data backup via LAN to the 

primary backup storage device connected to the backup 

server. Prior to backup operation, digital resources are 

classified as backup type and archive type. Furthermore, 

data which is classified as backup type will go through 

duplication elimination equipment or software to further 

reduce size before performing actual backup operation. 

Meanwhile, a comprehensive backup plan and associated 

backup strategy will be established using the planning 

functions of the backup software, and all data will be 

backed up through centralized management. It should be 

noted that the local backup storage is in fact the primary 

storage in our case, and the primary storage will map a 

copy in the remote archive storage, then use the backup 

software to provide safe disaster recovery measures. The 

proposed backup system can greatly shorten time 

required to perform backup and disaster recovery, and is 

capable to achieve high security and usability for 

network-based data backup. Structure of the proposed 

backup system for self-built resources is shown in Fig. 1.


B. System Backup 

Server backup. Self-built featured library consists of 

several processing computers and a processing 

management server. In order to ensure the smooth 

operation of the processing work, the operating system of 

all processing computers is backed up to a compact disc 

media. In general, all the processing computers are the 

same model and were purchased in the same batch, hence 

mo individual backup of the operating system of each 

processing computer is required. It is only necessary to 

do an individual backup for the process management 

server. Using server virtualization techniques such as 

virtual machines, the processing management server and 

the application server cluster which is responsible for the 

web database publishing and reader service can not 

only achieve highly efficient server performance, but also 

perform file backup on the server operating system. In 

addition, restore is simple and fast in this case. 

Data backup. In the construction of self-built 

databases, the originally sampled or collected data need 

to be backed up, and the final data production also has to 

be backed up. However, the original and the final data 

types bear different usage frequency and lifecycle, hence, 

the archive backup of the original data often happens 

after the completion of the processing operation, while 

the product data is backed up afterwards. 


Fig.2 the Backup data flow diagram 

Implementation of long-term preservation. Long- 

term data preservation is expected in the proposed backup 

system, therefore, migration and simulation techniques 

are employed to explore the characteristics of the 

self-built data resources in the archiving process. 

Specifically, archiving data and files used for reading 

environment are archived to the tape storage using the 

media server, and the archived data is regularly restored 

to verify its readability. 

Duplication elimination. Many backup strategies 

can reduce the backed up data size, however, the results 

are still not satisfactory. On the other hand, duplication 

elimination techniques can achieve a data compression 

ratio of 1:20, therefore, in the backup system, it is viable 

to use data duplication elimination combined with 

incremental backup to greatly reduce the storage space 

required for a whole system backup. Namely, prior to 

backup, data files are divided into several blocks of data 

to store. In principle, the same data block in different data 

files is backed up only once, thus significantly eliminate 

duplication in data, and reduce data redundancy[4-5] . 

IV. BACKUP DATA FLOW DIAGRAM 

A backup data flow diagram is shown in Fig. 2 which 

describes in detail how to back up the digital library data.


V. REMOTE BACKUP DATA STORAGE FOR DISASTER 

RECOVERY 

Unlike commercial data such as bank transactions, data 

in digital library is not necessary to be error-free, hence 

the main task for remote backup is to guarantee the 

consistency between the remote recovery data and the 

existing data in the local environment. Therefore, the 

proposed backup system uses the idle time period such as 

the period from 24:00 to 08:00, and employs 

asynchronous PPRC (Peer to Peer Remote Copy) 

combined with FLASHCOPY to ensure a secured copy of 

the whole backup. In the case when data in the main 

corrupts, remote disaster recovery backup can be used to 

quickly restore data service.. 

Currently, two basic types of disk-based remote copy 

are commonly used in industry, namely, synchronous 

PPRC and asynchronous PPRC. The major problem with 

synchronization PPRC is that it will occupy more 

bandwidth when transferring through network which 

influences the normal system performance. As a result, 

the performance of the whole system will be degraded 

when disaster recovery is carried out. 

Though asynchronous PPRC data may cause data lost 

problem, and asynchronous PPRC may cause 

inconsistency in the data if it fails to complete 

synchronization successfully, there is no doubt that 

asynchronous PPRC is far more efficient than its 

synchronous counterpart. As digital library has massive 

data, we have to select asynchronous PPRC to complete 

the daily remote backup. 

Therefore, we propose to do remote backup through 

asynchronous PPRC combined with Flashcopy, as 

asynchronous PPRC can resolve the performance 

problem while Flashcopy can resolve the data lost 

problem. Afterwards, data is synchronized. This is a 

quick way to ensure that backup data can be rolled back 

after the data loss. In fact, two techniques complement 

with each other, the combined application of two 

techniques result in more efficient, faster and safer 

disaster recovery than sole application of synchronous 

PPRC [6]. 

VI. BACKUP DATA ARCHIVE FLOW DIAGRAM. 

If remote disaster recovery backup data already exists, 

the major task for backup and archiving is to ensure data 

security and to decrease data storage requirement. 

Firstly, for those digital resources which are under 

construction or which are used in online services, part of 

the data do not require to be stored in the storage array in 

the form of long-term storage, these data can be archived 

directly; Furthermore, if some data do not have 

downloading or visiting for long time, these should also 

be archived according to the information life-cycle 

management theory. Therefore, the first step in Fig. 2 

is to evaluate data to determine whether it should be 

backed up or to be archived, and to decide whether to use 

disk media or tape in backup. In the case of archiving, 

using tape is relatively affordable. 

Secondly, duplication elimination techniques can be 

used to shrink size of the data that will be backed or 


archived; as a result, this also improves the efficiency of 

backup or archive operation. For the same data type, it is 

obvious that coping small amounts of data is much faster. 

Here we use duplication elimination techniques to 

process the source data. However, there are some 

exceptions, namely, duplication elimination techniques 

which are based on either hashing or content 

identification are effective only when content of data 

blocks are duplicated. For example, it is quite effective 

to apply duplication elimination techniques to the virtual 

machine files of service systems to greatly reduce the 

amount of backup or archiving data. Nevertheless, if the 

document is already in a compressed format such as 

DJVU or compressed video files, the situation becomes 

less optimistic. As duplication elimination takes a lot of 

time in comparison and calculation, it is best to do 

backup or archiving directly if the benefit obtained from 

data compression is not obvious. 

Thirdly, it is necessary to determine whether the data 

has been backed up or archived. If data has been backed 

up, only incremental or differential backup should be 

performed; otherwise, full backup has to be done. The 

decision to choose differential backup or incremental 

backup replies on the properties for data recovery which 

also use the same way to restore data. Differential backup 

is a backup all files has changed since the last full backup. 

Advantages of this method are that it performs well when 

a full recovery is demanded, as it only involves restoring 

a full backup and the latest differential backup. 

Disadvantage is that the size of differential backups 

grows quickly within a week. Hence the backup data can 

grow to a considerable scale before the next full backup. 

Incremental backup only backs up the data changed in 

files since the last backup, regardless whether the last 

backup is full or incremental. The main advantages of 

this approach are that files backed up each day between 

two full backups significantly reduce the backup window 

and are more concise. The disadvantage is that in order to 

perform a full recovery, the latest full backup has to be 

restored, together with all subsequent incremental 

backups, thus incremental backup is more time 

consuming. Technical explanation for the backups can be 

summarized as follows: full backup and incremental 

backup can be used to reset the archive bit in a file to 

indicate the file has been backed up, while differential 

backup cannot perform the archive bit resetting [7]. In 

addition, it is also necessary at this stage to choose a 

backup or archive data storage media in the hope to be 

more cost-effective. 

Finally, it is important to ensure the readability of 

backup or archive data, therefore, it is very important to 

perform data recovery verification. Based on the amount 

of backup or archiving data, as well as the differences in 

data backup storage hardware and archive material, it is 

proposed that the backup cycle should be a week or a 

month, while the archiving cycle should be three or six 

months. Of course, the six month archive cycle also takes 

into consideration that lower visiting amount for libraries 

happen at the winter and summer semester breaks for 

Chinese universities.


VII. SUMMARY 

With the rapid development of digital library, data 

volume in digital resources grow rapidly, and hence data 

storage and backup become more and more difficult. For 

the self-built digital resources which focus on the 

unstructured data, the proposed backup system can 

protect data well, and reduce backup storage consumption. 

However, due to the imperfection of long-term 

preservation techniques, it is still insufficient to achieve 

long-term preservation of data in the system. 

Storage media. Currently, a variety of digital 

resources are based on a binary 0 or 1 stored in some 

physical carrier, the digital information life depends on 

the physical carrier of life. The life of the disk is 

generally believed to be an average of 10 to 15 years, 

CD-ROM about 30 years, durable CD-ROM up to 100 

years, but the durable CD-ROM is expensive, hence 

cannot be widely used. Even the most durable discs 

cannot be used to save printed literature for thousands 

years, it is also that much difference, the long-term 

preservation of digital information need strengthening. In 

addition, because the digital document carrier prone to 

physical or chemical change, so demanding on the 

storage environment. The results of the disappearance of 

the event information will be disastrous about digital 

document. Therefore, solving the long-term preservation 

of digital resources stored vector problem is an important 

challenge of digital resource conservation; it needs to be 

studied carefully. 

Diverse formats of information resources. Library 

digital resources, including digital resources, whether self 

or mirroring database which be provided by database 

vendors, their format is diverse: TXT, PDF, CAJ, PDG, 

JPG, TIF, DJVU, MP3, MPG, RMVB and so numerous, 

as the storage technology of digital resources continuous 

development, many of the digital storage format makes a 

variety difficulty of network data exchange between 

information resources, It affects the long-term use of 

digital resources. 

Technology obsolete. With the network information 

technology and its products constantly upgrading, the one 

hand, enhanced information processing capability makes 

the network cost reduction, on the other hand, the using 

of stored digital resources has new difficulties. As digital 

resources are digital electronic information resources, it 

needs computer equipment with certain software. With 

these software and hardware technology continues to 


upgrade, making the old and new versions of the software 

is not compatible, use older versions of technology to 

store digital resources can not read, it makes the loss of 

resources, loss of use value. It affects the long-term 

preservation of digital resources. 

Network information security. In recent years, with 

the rapid popularization and development of the Internet, 

a variety of network information in the net, people can 

easily and freely on the Internet to read, browse, search, 

download a variety of network information, access to 

information to people has brought great convenience . 

However, a large number of computer viruses on the 

Internet, seriously affecting the information resources 

security on the network transmission and storage. 

Meanwhile, hackers also took the opportunity to 

infiltration, they use the computer system itself, there are 

a lot of flaws and weaknesses to attack, the light can not 

use the computer, the serious is causing severe paralysis 

of the network, so that preservation of digital resources is 

increasingly serious security problem, it cause permanent 

loss of digital information, it became one of the biggest 

threats on preservation of digital resources. 

Further research will be carried out on long-term 

preservation technology to establish a more 

comprehensive, more reliable and efficient backup 

system. 

REFERENCES 

[1] CHEN You hua, ZHENG Qiao ying, YANG Zong ying, 

WANG Shao ping, SUN Hua. Self-Developed Digital 

Resources of Chinese Academic Libraries [J]. Journal of 

Shanghai Jiaotong University, 2003, (S1). 

[2] http://baike.baidu.com/view/2119114.htm 

[3] Li kezheng. Long-term preservation of digital information 

technology analysis [J]. Library Work and Study, 2006, (2). 

[4] You L L,Pollack K T,Long D D E.Deep Store:An Archival 

Storage System Architecture[C]//Proc.of the 21st 

International Conference on Data 

Engineering.California,USA:[s.n.],2005:804-815. 

[5] Bhagwat D,Pollack K T,Long D D E.Providing High 

Reliability in a Minimum Redundancy Archival Storage 

System[C]//Proc.of the 14th IEEE International Symposium 

on Volume.New York,USA:[s.n.],2006:413-421. 

[6] Zhou Jian Feng.RESEARCH OF FAST DISASTER 

RECOVERY BETWEEN DIFFERENT SITES[D]. 

Shanghai Jiao Tong University,2009. 

[7] http://www.searchstorage.com.cn/ShowContent_14945_26. 

htm


Configuration Scheme for Small Scale 

Multi-FPGA Systems 

Chengchang Zhang 1 

1 College of Communication Engineering 

Chongqing University, Chongqing, P.R. China 

Email:zcc_918@163.com 

Lisheng Yang 2 , Dangui Yan 3 , Changyong Li 4 

2 College of Communication Engineering, Chongqing University, Chongqing, P.R. China 

Email: yls@ccee.cqu.edu.cn 

3 College of Mathematics and Physics, Chongqing University of Post and Telecom, Chongqing, P.R. China 

Email: yandg@cqupt.edu.cn 

4 Chongqing Communication Acadimic of P.L.A. ,Chongqing, P.R. China 

Email: lll_ccc_yyy@163.com 

Abstract—Multi-FPGA systems have tremendous potential, 

providing a high-performance computing substrate for 

many different applications. These systems harness multiple 

FPGAs, connected in a fixed pattern, to implement complex 

logic structures. In order to use such a system effectively, 

it is a key for constructing a good performance hardware 

platform. The configuration scheme is an important part in 

hardware design. This paper aims at small scale 

Multi-FPGA systems composed of SRAM-based FPGAs 

developed by Xilinx Corporation, proposes a novel 

configuration technique by using Platform Flash PROM 

XCF32P. Using this scheme, only adopting one XCF32P and 

one Complex Programmable Logic Device (CPLD) we can 

configure four FPGAs with monolithic configuration data 

smaller than 8Mbit. When the number of FPGA is more 

than four, Design revisioning allows the user to cascade 

more XCF32P PROMs to realize. Since Xilinx Platform 

PROM and Xilinx FPGA/CPLD are used to get a 

single-vender solution, the design for hardware and 

software is simplified. 

Index Terms—Multi-FPGA systems, XCF32P, design 

revision, configuration, 


There is currently tremendous interest in the 

development of computing platforms from multiple 

standard FPGAs [1,2,3,4]. One reason is that the digital 

system is too large to be achieved with only one FPGA, 

another, the growth rate of the FPGA capacity is far 

behind that of the ASIC(Application Specific Integrated 

Circuit) chip scale [5,6]. These systems harness multiple 

FPGAs [7], connected in a fixed pattern, to implement 

complex logic structures. In order to use such a system 

effectively, it is a key for constructing a good 

performance hardware platform. The configuration 

method plays important role for hardware platform 

because of two major factors. First, the configuration 

chips affect layout and wiring for printed circuit 

board(PCB). Second, the initialization and 


doi:10.4304/jcp.6.9.1988-1993 

reconfiguration for a multi-FPGA system is usually 

needed after the PCB developed, especially in system 

debug. A good design of configuration can optimize 

construction of PCB, and also make the configuration and 

debug processes more convenient and effective. 

In this paper, we focus on SRAM-based FPGAs 

developed by Xilinx Corporation. In SRAM-based 

FPGAs, the contents of the internal configuration 

memory are reset after power-up. As a result, the internal 

configuration memory cannot be used for storing 

configuration data permanently. SRAM-based FPGAs 

require external devices to initiate and control the 

configuration process. 

For Multi-FPGA systems configuration, if the number 

of FPGA chip and monolithic FPGA configuration files 

are both very large in a system, such as the DN9000K10 

System [8] developed by Dini Company, the Xilinx 

Company launched a special configuration solution, that 

is: System ACE (System Advanced Configuration 

Environment), in this solution, CF(Compact Flash) Card 

and ACE Control Chip are used to configure the multiple 

FPGAs automatically [9,10], but the system is costly. For 

general application system (such as the number of FPGA 

isn’t larger than four, and the configuration files is less 

than 8Mbit), self-made configuration scheme is usually 

adopted, for example, literatures [11,12,13,14] use the 

configuration scheme based on CPLD and general 

FLASH, a special FLASH drive device is needed to 

program configuration file to FLASH, and a group of 

output pins corresponding with FLASH capacity are 

needed to be distributed as address bus. And, designers 

must be clear with the first and the end address in the 

FLASH corresponding with configuration files of each 

FPGA, so that they can make sure that the counter in 

CPLD can start the control signal of next FPGA 

configuration after completing the last configuration, 

which is in fact very troublesome. Besides, the access 

speed of general FLASH is relatively slow to the FPGA 

and affects the system configuration speed. Literature


[15] adopted the DSP + CPLD + general FLASH 

configuration scheme, which is based on processor, the 

design and debug of the circuit and program cost 

considerable time, and processor usually bears arduous 

task in addition to completing the FPGA configuration, 

so bus contention is appear easily. 

In this paper, we propose a novel configuration scheme 

based on Xilinx Platform Flash PROM XCF32P 

The chip supports FPGA serial or parallel interface 

configuration, basically have the following typical 

characteristics [16,17,18]: 

The embedded data decompressor compatible with 

Xilinx senior compression technique can decompress 

PROM compressed files with a highest 50% data 

compression ratio, and the compressed file is generated 

from target FPGA bit stream file. When decompression is 

enabled, FPGA must be in slave configuration mode and 

PROM first decompress the stored data then drive the 

clock and data to FPGA interface. 

There is an optional oscillator in interior and can 

provide a 20MHz or 40MHz clock which is output by 

CLKOUT pin. Among them, the 40MHz clock is used to 

start the internal decompressor. 

Design revisioning allows the user to create up to four 

unique design revisions on a single PROM or stored 

across multiple cascaded PROMs. Design revisioning can 

be used with compressed PROM files, and also when the 

CLKOUT feature is enabled. The 32Mbit storage 

capacity of monolithic XCF32P can be divided into 

several independent spaces, with 8Mbit as a unit, and 

each independent space can store an independent 

configuration file, which is called a storage version. 

There are many methods to manage storage versions. 

Shown as Fig. 2, one XCF32P can be divided into only 

one 32Mbit storage version, two independent 16Mbit 

storage versions, one independent 8Mbit storage version 

and one independent 24Mbit storage version, two 

independent 8Mbit storage versions and one independent 

16Mbit storage version or four independent 8Mbit 

storage versions, and so on. During the PROM file 

creation, each design revision is assigned a revision 

number: Revision 0 = '00', Revision 1 = '01', Revision 2 = 

'10', Revision 3 = '11'. 


Fig.1 Structure diagram of XCF32P 

to simplify the design of hardware and software. 

II. XCF32P STRUCTURE CHARACTERISTICS 

XCF32P is the programmable high capacity Platform 

Flash PROM developed by Xilinx Company, its storage 

capacity is 32Mbit. The structure diagram is shown as 

Fig.1. 

Fig.2 Design Revision storage examples for a single XCF32P PROM 

After programming the Platform Flash PROM with a 

set of design revisions, a particular design revision can be 

selected using the external REV_SEL[1:0] pins or using 

the internal programmable design revision control bits. 

The EN_EXT_SEL pin determines if the external pins or 

internal bits are used to select the design revision. When 

EN_EXT_SEL is Low, design revision selection is 

controlled by the external revision select pins, 

REV_SEL[1:0]. When EN_EXT_SEL is High, design 

revision selection is controlled by the internal 

programmable revision telect control bits. During power 

up, the design revision selection inputs(pins or control 

bits) are sampled internally. After power up, when CE is 

asserted (Low) enabling the PROM inputs, the design 

revision selection inputs are sampled again after the 

rising edge of the CF pulse. The data from the selected 

design revision is then presented on the FPGA 

configuration interface. 

Xilinx company develops the Multiple versions design 

function of Platform Flash PROM is to realize the 

dynamic reconfigure of system or for some special 

application of changeable configuration when start the


FPGA each time. The work in this paper uses the 

multiple independent design versions to achieve multiple 

FPGAs configuration. 

III. CONFIGURING FOUR VIRTEX XCV200 FPGAS 

A. System components. 

The system includes one Platform Flash PROM 

XCF32P, one CPLD XC9572 and four XCV200 FPGAs 

to be configured, the system structure diagram is shown 

in Fig. 3a, and the circuit board is shown in Fig. 3b. 

Fig.3a. Structure diagram of configuration system 

B. Configuration principle. 

Fig.3b. The circuit board 

The configuration interface circuit is shown in Fig. 4. 

The circuit is designed with the help of OrCAD 

software. Because the software can't identify the sign of 

NOT operation, low-level effective is expressed as "/" 

(same in the following text). 

Virtex XCV200 FPGA supports the following four 

configuration modes[19]: master serial mode, slave serial 

mode, slave parallel (Slave SelectMAP)mode and 

boundary scan mode. In this work, high-speed 

slave-parallel mode is used and configuration clock 

CCLK is supplied by exterior. The frequency is 

determined by the formula followed: 

In equation (1), t is the access time of XCF32P 

ACC 

with a minimum of 25ns, t is the setup time of 

SMDCC 

input data of the SelectMAP interface with a minimum of 

2ns.Thus, the maximum frequency of CCLK is about 

37MHz. 


f 

CCLK 

1 

= 

t + t 

ACC SMDCC 

Fig.4 Configuration circuit interface 

Make a Parallel connection for the control signal 

CCLK, /PROGRAM, /INIT and data D[7..0] of the four 

FPGAs. And configure all devices orderly by setting chip 

selected signal /CS [4..1] respectively. When one of the 

three FPGAs(FPGA1, FPGA2 and FPGA3) configuration 

completed, it will enter its start-up stage, and send out its 

instructions signal DONE, set the version selection signal 

corresponding to the next configuration program and start 

configuration for next FPGA. It means that configuration 

is completed when the forth FPGA(FPGA4) release its 

signal DONE. This signal is connected to /CE, XCF32P 

is no longer effective and configuration process ends. The 

configuration flow is shown in Fig. 5. 

The data configuration timing diagram is shown in Fig. 

6. When /PROGRAM is in low state, four FPGAs begin 

to initialize synchronously. After initialization completed, 

the signal DONE turns to be low. Because the signal /CE 

of XCF32P is connected with the signal DONE of the 

forth FPGA (DONE4), the chip enable signal of XCF32P 

is effective. Meanwhile, the signal /INIT turn to be low 

automatically and begin to clear configuration memory. 

When the low level of the signal /INIT is input to the 

OE/(/RESET) interface of XCF32P, the chip XCF32P 

begins to reset and address pointer points to the first 

address of memory space. After configuration memory is 

emptied, the signal /INIT is set to high again, and device 

samples mode pins to make sure that configuration data is 

loaded in parallel mode. 

When multi-version design function is started, the 

internal logic of configuration PROM samples the design 

version selected input(pin /SEL) when power up. When 

/CE is set to low, the design version selected input signal 

(1)


is sampled again at the rising edge of /CF pulse to 

determine which design version to provide configuration 

monolithic 

configuration 

process 

Therefore, the signal /INIT is regarded as the initial 

trigger signal of /CF, and /CF is triggered at the rising 

edge of /INIT. /CF is set to be low and the low level 

should delay more than 300ns duration. The version 

selected input signal /SEL is triggered and set to be “00” 

at the same time, namely, configuration data is output 

form the zero version. Trigger piece selected signal /CS1 

is effective at the rising edge of /CF signal. The zero 

version data of XCF32P is output to the first FPGA and 

begin to configure FPGA1 at the affection of CCLK. 

When the first FPGA is configured, it releases the signal 

DONE, by this way, DONE1 turns to be high level. /CF 

signal is triggered by the rising edge of DONE1 and is 

reset to be low level, at the same time, /SEL signal is set 


data for the FPGA. The version selected pin should be set 

before sampling is triggered at least 300ns. 

Start Clear the configuration 

memory and set DONE 

to be low 

/PROGRAM is low? 

Clear the configuration 

memory again 

/INIT is low? 

Begin to configure FPGA1 

Set the chip selected /CS to be low; Set the version code 

REV_SEL corresponding to this chip; Set the version 

initialization signal /CF to be low, and the low level stay 

for longer than 300ns. 

Write data in 

BUSY is low? 

DONE is high? 

Set the chip selected /CS 

to be high, and enter the 

starting process 

Repeat monolithic configuration 

process, and configure FPGA2, 

FPGA3 and FPGA4 

DONE4 is high? 

Set /CE to be high 

End 

N 

Y 

Y 

Y 

Y 

Fig.5 Configuration flow 

N 

N 

N 

N 

to be “01”. When the rising edge of /CF signal arrives, 

configuration data is sent out by the first version of 

XCF32P. By this time, /CS2 is set to be effective and the 

second FPGA is selected to begin receive configuration 

data. Besides, /CS1 is set to be ineffective and starts to 

configuration the second FPGA. The configuration of the 

third and forth FPGA is similar to above. After the forth 

FPGA configured, /CE of XCF32P is set to be high level 

by DONE4 signal released by this FPGA. That is to say, 

the chip enable signal of XCF32P is ineffective and the 

whole configuration process ends. 

C.The software design of CPLD 

The design of internal control circuit in CPLD is a key


of the system. Providing the needed timing sequence 

when configuring, coordinating the configuration process, 

and ensuring that multi-FPGA configuration completed 

Design is realized by combining the hardware 

description language with schematic diagram. Control 

circuits are made up of a delay model, a counter and a 

shift register, as shown in Fig. 7. 

Delay module tests the rising edge of /INIT, 

DONE(1), DONE(2) and DONE(3) and trigger internal 

delay circuit to produce the negative pulses longer than 

300ns which is need by /CF signal. It is difficult to detect 

rising edge of four signals simultaneously, so, there are 

four independent delay circuits in delay model to detect 

four trigger signals respectively and to produce four 

negative pulses which can get /CF signal when they are 

done the AND operation. The shift register is triggered 

by the rising edge of /CF and produces the chip selected 

signal /CS(4:1). The falling edge of /CF triggers counter 

and produce version selected signal /SEL(1:0). 

The simulation results of control circuit are shown in 

Fig. 8. 

IV. CONCLUSIONS 

A new configuration scheme for small scale 

multi-FPGA systems based on XCF32P is given. In this 

scheme, a XCF32P and a CPLD are used to configure 


Fig.6 Configuration timing 

Fig.8 Control circuit simulation results 

as the predetermined process are the main functions of 

this work. 

Fig.7 Control circuit block 

four Virtex XCV200 FPGAs. The design has certain 

universality, and can be used to configure multiple Xilinx 

FPGAs with monolithic configuration data smaller than 

8Mbit. When starting the internal decompression in 

XCF32P, monolithic FPGA configuration data can reach 

16Mbit. When the number of FPGA is more than four,


Design revisioning allows the user to cascade more 

XCF32P PROMs to realize. 

Due to the XCF32P is special configuration chip 

developed by XILINX Company, the chip access time is 

short, and the configuration speed is fast. Meanwhile, 

Xilinx Platform and Xilinx FPGA/CPLD are used to get a 

single-vender solution to make the design for hardware 

and software simplified. 


The authors thank their colleagues at College of 

Communication and Engineering for fruitful discussions. 

This work was supported by Natural Science Foundation 

Project of CQ CSTC under contract no: 2010BB2240. 

REFERENCES 

[1] Panella, A.; Santambrogio, M.D.; Redaelli, F.; Cancare, 

F.; Sciuto, D.. A design workflow for dynamically 

reconfigurable multi-FPGA systems. VLSI System on 

Chip Conference (VLSI-SoC), 2010 18th IEEE/IFIP, pp. 

414-419. 

[2] Jain, S.C.; Kumar, S.; Kumar, A.. Evaluation of various 

routing architectures for multi-FPGA boards. VLSI 

Design, 2000. Thirteenth International Conference on , 

pp.262-267. 

[3] Khalid, M.A.S.; Rose, J.. A novel and efficient routing 

architecture for multi-FPGA systems. Very Large Scale 

Integration (VLSI) Systems, IEEE Transactions on, Vol. 8 

, Issue. 1, 2000 , pp.30-39. 

[4] Zhang Cheng-chang; Yan Dan-gui; Yang Li-sheng; Qi 

Huai-long; Li Chang-yong. DLL-based multi-FPGA 

systems clock synchronization. Industrial Electronics and 

Applications (ICIEA), 2010 the 5th IEEE Conference on, 

pp. 1420-1423. 

[5] Krupnova, H. Mapping multi-million gate SoCs on 

FPGAs:industrial methodology and experience. Design, 

Automation and Test in Europe Conference and 

Exhibition, Proceedings, Volume 2, 16-20 Feb.2004 Vol.2, 

pp.1236-1241. 

[6] Melnikova, O.; Hahanova, I.; Mostovaya, K.. Using 

multi-FPGA systems for ASIC prototyping. CAD 

Systems in Microelectronics, 2009. CADSM 2009. 10th 

International Conference-The Experience of Designing 

and Application of 24-28 Feb.2009, pp.237-239. 

[7] Scott Hauck. Multi-FPGA Systems, Doctor of Philosophy, 

University of Washington. 1995. 

[8] http://www.dinigroup.com/index.php/. 

[9] Yang Sen; Chen Jian-jun; Wang Jian-guo. System ACE 

CF Technology--A New Configuration Solution for 

FPGAs. Radar & Ecm, 2002(4), pp.72-77. 

[10] Alonso, R.; Barbara, D.; Cova, L.L.. A file storage 

implementation for very large distributed systems. 


Workstation Operating Systems, 1989., Proceedings of 

the Second Workshop on , pp. 1-5. 

[11] Li Peng; Lan Ju-long. The configuration method for 

FPGA based on CPLD and Flash. Application of 

Electronic Technique, 2006(6), pp.101-103. 

[12] Guo Tian-tian. Interface Circuit of Configuring Virtex 

FPGA Through SelectMAP. Microprocessors, 2000(4), 

pp.17-19. 

[13] Zhang Hong-gang; Xin Fan-ge; Wang De-shi. The fast 

configuration circuit design for FPGA based on CPLD. 

Application of Electronic Technique, 2006(2), 

pp.123-125. 

[14] Xiao Jin-qiu; Liu Chuan-yang; Feng Yi; Zhong Jia-lin. 

Design of FPGA Initialization Configure System at High 

Speed with LPC Bus. Computer Engineering, 

2005(13):176-178. 

[15] She You-jun; Wang Dan. Design for double FPGA 

configuration based on TMS320C61416 EMIF bus 

[J].Microcontrollers & Embedded Systems, 

2007(7):29-31. 

[16] Platform Flash In-System Programmable Configuration 

PROMS. http://www.xilinx.com, DS123 (v2.6) March 

14, 2005. 

[17] LI Yan-bin; LI Yan-chun. Fast Dynamic Reconfiguration 

of FPGA with XCF32P. Telecommunication Engineering, 

2006(6):199-202. 

[18] Platform Flash PROM User Guide. http://www.xilinx.com, 

UG161(v1.5) October 26, 2009. 

[19] Virtex 2.5V Field Programmable Gate Arrays. 

http://www.xilinx.com, DS003-1(v2.5) April 2, 2001. 

Chengchang Zhang was born in Lichuan, China, 1975. He 

received the BS degree in automation engineering from the 

Wuhan Institute of Chemical Technology in 1997. He received 


Chongqing University in 2005. He is currently PhD candidate 

of Chongqing University majoring in Communication and 

information systems. His research interests are software radio 

and FPGA design. 

Lisheng Yang was born in Chongqing, China, 1972. He is an 

professor of Chongqing University. His research interest 

includes software radio, radar ,TT&C, etc. 

Dangui Yan was born in Luotian, China, 1975. She is an 

lecturer of Chongqing University of Post and Telecom. Her 

research interest is logic algebra. 

Changyong Li was born in Chongqing, China, 1971. His 

research interests are software radio and ultra-wide band radar.


Order Bi-spectrum For Bearing Fault Monitoring 

and Diagnosis Under Run-up Condition 

Hui Li 

Department of Electromechanical Engineering, Shijiazhuang Institute of Railway Technology, Shijiazhuang, China 

Email: Huili68@163.com 

Abstract—Varying speed machinery condition detection and 

fault diagnosis are more difficult due to non-stationary 

machine dynamics and vibration. Therefore, most 

conventional signal processing methods based on time 

invariant carried out in constant time interval are 

frequently unable to provide meaningful results. This paper 

deals with the detection of bearing faults in gearbox under 

non-stationary run-up of gear drives. In order to process the 

non-stationary vibration signals such as run-up or rundown 

vibration signals effectively, the order bi-spectrum 

technique is presented. This new method combines 

computed order tracking technique with bi-spectrum 

analysis. First, the vibration signal is sampled at constant 

time increments during run-up of gearbox and then uses 

numerical techniques to resample the data at constant angle 

increments. Therefore, the vibration signals are 

transformed from the time domain transient signal to angle 

domain stationary one. Second, the re-sample signal is 

processed by bi-spectrum analysis method. The procedure is 

illustrated with the experimental vibration data of a gearbox. 

The experimental results show that order bi-spectrum 

technique can effectively diagnosis and diagnosis the faults 

of bearing. 

Index Terms—fault diagnosis, gearbox, bearing, vibration, 

signal processing, order tracking, bi-spectrum 


Rotating machine fault diagnosis is typically based on 

vibration. The spectral contents of emitted vibration 

signals are analyzed to ascertain the current condition of 

the monitored process. At present, for the fault diagnosis 

of rotating machinery, many research outcomes have 

been obtained in the stationary process. However, little 

research has been done for monitoring the vibrations of 

varying speed condition such as the run-up or run-down 

process. The reason why we stress the run-up or rundown 

process is that non-stationary vibrations signals 

from varying speed machinery may include more 

abundant information about its condition. Some 

phenomena, which are usually not obvious at constant 

speed operation, may become more apparent under 

varying speed conditions. Therefore, the behavior 

characteristics of the run-up or run-down process have a 

distinct diagnostic value, and the fault diagnosis of run-up 

Manuscript received November 5, 2010; revised December 23, 

2010; accepted January 28, 2011. 

Corresponding author: Hui Li. 


doi:10.4304/jcp.6.9.1994-2000 

or run-down process has owed its distinct standing in the 

fault diagnosis of rotating machinery. In the last decade 

vibration analysis and condition monitoring techniques 

for varying speed machinery have attracted the attention 

of scientists and engineers. Lopatinskaia et al. [1,2] 

presented the application of recursive filtering and angle 

domain analysis to non-stationary vibration analysis. The 

approach is implemented and validated through computer 

simulation and experiments. Meltzer [3,4] dealt with the 

recognition of faults in gear tooth during non-stationary 

start-up and run-down of planetary gear drives using the 

time-frequency approach and the time-quefrency 

approach. Wu et al. [5] presented the application of 

adaptive order tracking fault diagnosis technique based 

on recursive Kalman filtering algorithm to gear-set defect 

diagnosis and engine turbocharger wheel blades damaged 

under various conditions. Li et al. [6] presented the 

hidden Markov model-based fault diagnosis method in 

speed-up and speed-down process for rotating machinery. 

However, the vibration signal of the run-up or rundown 

process is more complex than that of the stationary 

process. Conventional signal processing methods, which 

were developed for constant speed machinery monitoring, 

are based on digital sampling carried out in equal time 

intervals. If the machine operates under varying speed or 

load, its dynamic and vibrations become non-stationary. 

The vibration signal sampled from the rotating machinery 

is a non-stationary signal, whose amplitudes and 

frequencies both vary with time. Fixed time sampling 

cannot cope with the varying rotational frequency of the 

machine, resulting in increasing leakage error and 

spectral smearing [1,2]. Therefore, most of the 

conventional methods for signal processing become 

inappropriate when monitoring the vibrations of varying 

speed machinery [1,2]. Some progress has been made in 

the theoretical analysis [7,8], the signal processing 

methodology [9,10], measurements and practical 

applications of varying speed machinery monitoring 

[11,12,13]. 

At present, two techniques are mainly used to process 

the non-stationary signal: time frequency analysis (such 

as the short time Fourier transform (STFT), wavelet 

transform (WT) [14], Wigner-Ville distribution (WVD) 

[15,16,17] and Hilbert-Huang transform [18,19,20]) and 

order tracking technique [11,12,13]. The time frequency 

analysis involves three-dimensional functions that allow 

for visualizing the frequency and amplitude variations of 

the spectral components [14]. However, when the


analyzed vibration signal is composed of many spectral 

components and with large changes of the machine speed 

during measurement, they become very difficult to 

analyze. Recently, order tracking has been become one of 

the important methods for fault diagnosis in rotating 

machinery [11,12,13]. Vibration signals produced from 

rotating machinery are speed dependent and hence orders 

as opposed to absolute frequencies are preferred as the 

frequency base. Orders represent the number of cycles 

per revolution and are thus ideal for representing speeddependent 

vibrations. Therefore, order tracking normally 

exploits a vibration or a noise signal supplemented with 

the information of shaft speed for fault diagnosis of 

rotating machinery. The order spectrum gives the 

amplitude of the signal as a function of harmonic order 

and shaft speed in rotating machinery [11]. 

In this work, the computed order tracking approach 

and bi-spectrum analysis are introduced and applied 

specifically to gearbox fault diagnosis during run-up. 

This method is based on the re-sampling technique and 

the bi-spectrum estimation of the re-sampling signal, 

which is a function of the angle of the input shaft of the 

gearbox. This re-sampling signal can be obtained by resampling 

of the vibration signal that has been previous 

sampled in the time domain. The order power spectrum 

and order bi-spectrum techniques are based on the signal 

processing of the angle domain signal, where the 

resample signal is in accordance with the shaft angle of 

the gearbox. The order power spectrum and order bispectrum 

are then evaluated for the vibration signal resampled 

constantly in angle at equidistant phases of the 

input shaft of the gearbox. In this case, the results of the 

order power spectrum or order bi-spectrum are expressed 

as results of order analysis where the frequency axes are 

changed to the axes of orders independent of the input 

shaft speed. The usefulness of this approach will be 

shown by experimental example in Section VI. 

To address the issues discussed above, this paper is 

organized as follows. Section I gives a brief introduction 

of the order tracking analysis technology. Section II 

briefly describes the bi-spectrum. Section III presents the 

principles and procedure of the computed order tracking. 

Section IV gives the method and procedure of the fault 

diagnosis based on computed order tracking and order bispectrum. 

Section V looks at the experimental set-up. 

Section VI gives the applications of the method based on 

computed order tracking and order bi-spectrum to faults 

diagnosis of bearing faults. Finally, the main conclusions 

of this paper are provided in Section VII. 

II. A BRIEF INTRODUCTION OF BI-SPECTRUM 

x be a real, discrete, zero-mean stationary 

process with third-order cumulant R xx ( τ1, 

τ 2 ) defined 

as [21] 

Let { (n) 

} 

Rxx τ , τ ) = E[ 

x( 

n) 

x( 

n + τ ) x( 

n + τ )] (1) 

( 1 2 

1 

2 

Then the bi-spectrum of { (n) 

} 

expression 


x is given by the 

+∞ +∞ 

− j( 

ω1 

τ1+ 

ω2τ 

2 ) 

B xx ( ω1 

, ω2 

) = ∑∑Rxx 

( τ1, 

τ 2 ) e 

(2) 

τ1= −∞ τ2= 

−∞ 

where ω1 ≤ π , ω2 ≤ π , ω1 + ω2 

≤ π . 

Therefore, in the same way that the power spectrum 

decomposes the power of a signal, the bi-spectrum 

decomposes the third-order cumulant. The bi-spectrum is 

a function of two frequency variables, ω1 and ω 2 , and 

whilst the power spectrum includes the contribution of 

each individual frequency component independently, the 

bi-spectrum analyses the frequency interaction between 

the frequency components at ω 1 , ω 2 and ω 1 + ω2 

[22- 

23]. 

III. THE PRINCIPALS OF COMPUTED ORDER TRACKING 

There are two popular techniques for producing 

synchronously sampled data: the traditional approach that 

uses special hardware to dynamically adapt the sample 

rate and a technique where the vibration signals and a 

tachometer signal are synchronously sampled, that is, 

they are sampled conventionally at equal time increments. 

From the synchronously sampled tachometer signal resample 

times required to produce synchronous sampled 

data are calculated. This process is referred to as 

computed order tracking and is particularly attractive, as 

it requires no special hardware. Also, this approach is 

more flexible than the traditional method, as for example 

different sample rates may be synthesized. The computed 

order tracking is considerably more flexible than the 

traditional approach. It may be organized to produce 

equally accurate or more accurate results than the 

traditional method. An added benefit is that computed 

order tracking requires no specialized hardware, which is 

an important factor in many conditions monitoring 

applications. Therefore, computed order tracking 

techniques are introduced and applied in this paper. 

The objective of computed order tracking (COT) [9] is 

a calculation of the vibration signal sampled constant in 

angle from sampled constant in time. From the 

mathematical point of view, this task could be solved by 

interpolation theory. 

To determine the resample times, it will be assumed 

that the shaft is undergoing constant angular acceleration. 

With this basis, the shaft angle θ (t) 

can be described by 

a quadratic equation of the following form [9]: 

2 

θ ( t ) = b0 

+ b1t 

+ b2t 

(3) 

where b 0,b 1 and b3 are unknown coefficients, which are 

found by fitting three successive key-phasor arrival times 

( 1 t , 1 t and t 3 ) which occur at known shaft angle 

increments ∆ φ . This can be obtained by the following 

conditions:


⎧ θ ( t1) 

= 0 

⎪ 

⎨ θ ( t2 

) = ∆φ 

⎪ 

⎩θ 

( t3 

) = 2∆φ 

Substituting these conditions into Eq. (3) and arranging 

in a matrix format gives, 

(4) 

2 

⎛ 0 ⎞ ⎡1 

t ⎤ 1 t1 

⎧b0 

⎫ 

⎜ ⎟ ⎢ 

2 ⎥⎪ 

⎪ 

⎜ ∆φ 

⎟ = ⎢1 

t2 

t2 

⎥⎨b1 

⎬ (5) 

⎜ ⎟ ⎢ 

2 ⎥⎪ 

⎪ 

⎝2∆φ 

⎠ ⎣1 

t3 

t3 

⎦⎩b2 

⎭ 

This set of equations is then solved for the unknown 

b components. Once these values are known, Eq.(3) 

{ } 

i 

may be solved for t , yielding 

2 

2 

[ 4b2 

( k∆ 

− b0 

) + b1 

b1 

] 

1 

t = θ − (6) 

2b 

where k is the interpolation coefficient which can be 

obtained as follow 

θ = k ∆θ 

(7) 

where θ is the shaft angle and ∆ θ is the desired 

angular spacing between re-samples. 

Once the resample times are calculated, the 

corresponding amplitudes of the signal are calculated by 

interpolating between the sampled data. After the 

amplitudes are determined, the re-sample data are 

transformed from the angle domain to the order domain 

by means of an FFT. 

The order spectrum and order bi-spectrum techniques 

are based on the signal processing of the angle domain 

signal, where the resample signal is in accordance with 

the shaft angle of the gearbox. The order spectrum and 

order bi-spectrum are then evaluated for the resample 

signal. The usefulness of this approach will be shown 

with an experimental example in Section VI. 

IV. PROPOSED ORDER BI-SPECTRUM METHOD FOR 

FAULTS DETECTION OF BEARING 

The procedure of proposed order bi-spectrum method 

is given as follows: 

1) Non-stationary vibration signal under run-up 

condition is sampled using a constant time increment; 

2) Non-stationary vibration signal is re-sampled at a 

constant angle increment. Then the non-stationary 

vibration signal in time domain is transformed into 

stationary one in angle domain; 

3) To demodulate the constant angle increment signal 

using Hilbert transform; 

4) The order bi-spectrum is calculated according to Eq. 

(2); 

5) The diagnostic conclusions are drawn according to 

the order bi-spectrum. 


V. EXPERIMENTAL SET-UP 

The test apparatus used in this study is shown in Fig.1 

[24,15]. The experimental set-up consists of a singlestage 

gearbox, driven by a 4.5 kW AC governor motor. 

The driving gear has 30 teeth and the driven gear has 50 

teeth. Therefore, the transmission ratio is 50/30, which 

means that an decrease in rotation speed is achieved. The 

module of the gear is 2.5 mm. The monitoring and 

diagnostic system is composed of three accelerometers, 

amplifiers, a speed and torque transducer, B&K 3560 

spectrum analyzer and a computer. The sampling span is 

3.2 kHz, the sampling frequency is 8192 Hz and the 

sampling time is 2 seconds. This time included one speed 

up of the gearbox from idle speed up to steady. After 

sampling, the measured vibration signals were loaded 

into MATLAB from data-files. Then, the vibration 

signals were re-samples. For their re-sampling, the 

algorithm described in the previous section was used. As 

a result of experiment, the vibration signals generated by 

the tested gearbox were obtained sampled constant in 

time as well as sampled constant in angle. 

Figure 1. Experimental set-up 

VI. BEARING FAULTS DIAGNOSIS BASED ON ORDER BI- 

SPECTRUM 

In this section, the order power spectrum and order bispectrum 

will be applied to vibration signal analysis 

measured from a gearbox during speed-up process. 

Ball bearings are installed in many kinds of machinery. 

Many problem of those machines may be caused by 

defects of the ball bearing. Generally, localized defects 

may occur on inner race, outer race or rollers of bearing. 

A local fault may produce periodic impacts, the size and 

the repetition period which are determined by the shaft 

rotation speed, the type of fault and the geometry of the 

bearing. The successive impacts produce a series of 

impulse response, which maybe amplitude modulated 

because of the passage of fault through the load zone. The 

spectrum of such a signal would consists of a harmonics 

series of frequency components spaced at the component 

fault frequency with the highest amplitude around the 

resonance frequency. These frequency components are 

flanked by sidebands if there is an amplitude modulation 

due to the load zone. According to the period of the 

impulse, we can judge the location of the defect using 

characteristic frequency formulae.


The tested bearing was used to study only one kind of 

surface failure: the bearing was damaged on the inner 

race or outer race. The ball bearing tested has a groove on 

the inner race or outer race. Localized defect was seed on 

the inner race or outer race by an electric-discharge 

machine to keep their size and depth under control. The 

size of the artificial defect was 1mm in depth and the 

width of the groove was 1.5 mm. The type of the ball 

bearing is 206. There are 9 balls (z=9) in a bearing and 

the contact angle α = 0° 

, ball diameter d=9.5mm, 

bearing pitch diameter D=41.75mm. Then the 

characteristic frequency of the inner race or outer race 

defect can be calculated according to the Eq.(8), Eq.(9), 

respectively. 

z ⎛ d ⎞ 

= + α f 

finner ⎜1 

cos ⎟ 

2 ⎝ D ⎠ 

z ⎛ d ⎞ 

= − α f 

fouter ⎜1 

cos ⎟ 

2 ⎝ D ⎠ 

where f r is the rotating frequency of the input shaft. 

Therefore, according to Eq.(8) and Eq.(9), the 

characteristic frequency of the inner race and outer race 

defect are given as follows: 

inner 

r 

r 

r 

(8) 

(9) 

f = 5. 

42 f 

(10) 

f = 3. 

58 f 

outer 

r 

(11) 

Then the characteristic order of the inner race and 

outer race are obtained as follows: 

O = 5. 

42 

(12) 

inner 

O = 3. 

58 

(13) 

outer 

A. Application of Order Bi-spectrum to Fault Diagnosis 

of Inner Race 

The rotating speed signal of the input shaft for the 

tested gearbox is displayed in Fig.2. Fig.2 (a) represents 

the sampling pluses of the input shaft from the optical 

encoder (60 pulses per rotational period). The encoder 

signals consist of 16384 points and have a total duration 

of 2 seconds. To obtain approximate values of rotational 

speed for every data point, polynomial curve fitting was 

used. It was found that linear approximate was sufficient 

for this research. polynomial coefficients were 

determined for each data and analytical descriptions of 

the rotational speed were obtained. Fig.2 (b) is the 

calculated instantaneous rotating speed using 

interpolating method. Fig.2 (b) clearly shows that the 

rotating speed of the input shaft runs up from idle to 

steady speed about 700 rpm. 

The original vibration signal with inner race fault is 

displayed in Fig.3 (a). Fig.3 (a) shows that the vibration 

signals are non-stationary which the amplitude of the 


vibration is increasing during the input shaft speed up. 

The result of applying conventional spectral analysis 

(FFT) to the specified non-stationary signal is given in 

Fig.3 (b). Fig.3 (b) displays the FFT of the vibration 

signals with inner race fault. It is very clear that the 

resulting spectrum is significantly obscured by spectral 

smearing. Besides, traditional spectral averaging cannot 

be applied to the non-stationary signal during the input 

shaft run-up process. Fig.3 (b) clearly shows that spectral 

smearing substantially affects the result of conventional 

analysis based on time sampling. Therefore, classical 

Fourier analysis has some limitation such as being unable 

to process non-stationary signals. 

Figure 2. Rotating speed of the input shaft 

Figure 3. Time-domain vibration signal with inner race fault and FFT 

Figure 4. Angular resample signal of Fig.3 (a)


The angular re-sampling technique is applied to the 

vibration signal of Fig.3 (a). Fig.4 displays the re-sample 

vibration signal with uniform angular increment of 

0.008722 rad. It is clear that there are periodic impacts in 

the angle domain vibration signal. There are significant 

fluctuations in the peak amplitude of the signal. However, 

it is hardly possible to evaluate the bearing fault condition 

only through such angle domain vibration signal. Fig.5 

shows the order power spectrum of the re-sample 

vibration signal. The order power spectrum, as shown in 

Fig.5, is dominated by the repetition order of the gear 

mesh order and its harmonics. It can be seen from Fig.5, 

that the order power spectrum represents the complicated 

quantities. This complexity of the order power spectrum 

follows from the frequency smearing and modulation 

effects. Therefore, the conventional order power 

spectrum was not capable of revealing the characteristic 

order of inner race fault that was corrupted by the 

modulation and noise. 

Figure 5. Order spectrum of inner race fault 

Figure 6. Order bi-spectrum of inner race fault (contour) 

The order bi-spectrum was evaluated according to the 

conventional direct method [2] after the re-sample signal 

has been demodulated by Hilbert transform. The order bispectrum 

is depicted in Fig.6 (contour plot) and Fig.7 

(mesh plot). From Fig.6 we can see that the graphs of the 

order quantities are much simple than that of the order 

power spectrum of Fig.5. In case of the order bi-spectrum, 

it can be identified that the characteristic order ( O ) of 

inner 

inner race fault and its harmonics are represented clearly 

in the order bi-spectrum. The simplicity of the order 


quantity representation can be put down to the ability of 

the order signal processing method to eliminate 

undesirable spectral smearing and modulation effects. 

Fig.6 and Fig.7 demonstrate the advantage of the order 

quantity application for the analysis vibration signals 

generated by gearbox under running up condition. 

Especially, the order bi-spectrum better identifies the 

order components and consequently leads to a better 

understanding of the transient vibration characteristics 

than that of the order power spectrum. 

Figure 7. Order bi-spectrum of inner race fault (mesh) 

B. Application of Order Bi-spectrum to Fault Diagnosis 

of Outer Race 

Figure 8. Time-domain vibration signal with outer race fault and FFT 

Figure 9. Angular resample signal of Fig.8 (a)


Figure 8(a) shows the original vibration signal with 

outer race fault during the input shaft speed-up. Fig.8 (b) 

displays the FFT of the vibration signal with outer race 

fault. It is clear that the resulting spectrum is the same as 

the inner race fault that is significantly obscured by 

spectral smearing. 

Figure 10. Order spectrum of outer race fault 

Figure 11. Order bi-spectrum of outer race fault (contour) 

Figure 12. Order bi-spectrum of outer race fault (mesh) 

Figure 9 displays the re-sample vibration signal with 

uniform angular increment. Fig.10 is the order power 

spectrum of the re-sample vibration signal. The 

conventional order power spectrum was not capable of 


revealing the characteristic order of outer race fault in the 

same way. The order bi-spectrum is depicted in Fig.11 

(contour plot) and Fig.12 (mesh plot), respectively. It can 

be seen clearly from Fig.11 and Fig.12 that there are the 

characteristic order ( O outer ) of outer race fault and its 

harmonics. Therefore, the outer race fault can be easily 

detected by using order bi-spectrum. Fig.11 and Fig.12 

demonstrate the advantage of the order bi-spectrum for 

the analysis vibration signals generated by gearbox 

during run-up process. 

VII. CONCLUSIONS 

A method for fault diagnosis of bearing under run-up 

condition was presented based on a newly developed 

signal processing technique termed as computed order 

tracking and order bi-spectrum. Using computed ordertracking 

technique, the non-stationary vibration signals of 

bearing faults in time domain can be transformed into 

stationary ones in the angle domain. The definition of the 

order bi-spectrum for analysis of vibration signals 

generated by rotating machinery was introduced. This 

method is based on the bi-spectrum estimation from the 

vibration signal sampled constant in angle with respect to 

the shaft speed of the gearbox. The order bi-spectrum 

method assists in the elimination of spectral smearing and 

modulation effects caused by the variation in shaft speed. 

The experimental results show that order bi-spectrum can 

be effectively used as a diagnostic feature for bearing 

faults. 


The authors are grateful to the National Natural 

Science Foundation of China (No.50975185), Zhejiang 

Provincial Natural Science Foundation (No.Y1080040). 

The authors are also grateful to the editors and reviewers 

for their constructive comments. 

REFERENCES 

[1] E.Lopatinskaia, J.Zhu, J.Mathew, “Monitoring varying 

speed machinery vibration- I. The use of non-stationary 

recursive filters”, Mechanical Systems and Signal 

Processing, vol.9, no.6, pp.635-645, 1995. 

[2] E.Lopatinskaia,J.Zhu, J.Mathew, “Monitoring varying 

speed machinery vibration -II. Recursive filters and angle 

domain,” Mechanical Systems and Signal Processing, 

vol.9, no.6, pp.647-655, 1995. 

[3] G.Meltzer,Y.Y.Ivanov, “Fault detection in gear drives with 

non-stationary rotational speed- part I: the time-frequency 

approach,” Mechanical Systems and Signal Processing, 

vol.17, no.5, pp.1033-1047, 2003. 

[4] G.Meltzer,Y.Y.Ivanov, “Fault detection in gear drives with 

non-stationary rotational speed- part II: the time-quefrency 

approach,” Mechanical Systems and Signal Processing, 

vol.1, no.2, pp.273-283, 2003. 

[5] JianDa Wu, ChinWei Huang, Rongwen Huang, “An 

application of a recursive kalman filtering algorithm in 

rotating machinery fault diagnosis,” NDT&E International, 

vol.37, no.3, pp.411-419, 2004. 

[6] Zhinong Li, Zhaotong Wu, Yongyong He, Chu Fulei, 

“Hidden Markov model-based fault diagnostics method in


speed-up and speed-down process for rotating machinery,” 

Mechanical Systems and Signal Processing, vol.19 no.2, 

pp.329-339, 2005. 

[7] R.Potter, M.Gribler, “Computed Order Tracking Obsoletes 

Older Methods,” Proceedings of the SAE Noise and 

Vibration conference, pp. 63-67, 1989. 

[8] R.Potter, “A New Order Tracking Method for Rotating 

Machinery,” Sound and Vibration, vol.7, pp.30-34, 1990. 

[9] K.R.Fyfe, E.D.S.Munck, “Analysis of computed order 

tracking,” Mechanical Systems and Signal Processing, 

vol.11, no.2, pp.187-205, 1997. 

[10] K. M. Bossley and R. J. Mckendrick, “Hybrid computed 

order tracking,” Mechanical Systems and Signal 

Processing, vol.13, no.4, pp.627-641, 1999. 

[11] J.R.Blough, “Development and analysis of time variant 

discrete Fourier transform order tracking,” Mechanical 

Systems and Signal Processing, vol.17, no.6, pp.1185-1199, 

2003. 

[12] Hui Li, Yuping Zhang, Haiqi Zheng, “Angle Order 

Analysis Technique for Processing Non-stationary 

Vibrations,” Proceedings of 7th International Symposium 

on Test and Measurement, vol.5, pp.4000-4003, 2007. 

[13] Hui Li, Yuping Zhang, “Order Tracking and AR Spectrum 

Based Bearing Fault Detection Under Run-up Condition,” 

Proceedings of the First International Congress on Image 

and Signal Processing, vol.5, pp.286-290, 2008. 

[14] J.Lin and L. Qu, “Feature extraction based on Morlet 

wavelet and its application for mechanical fault diagnosis,” 

Journal of Sound and Vibration, vol.234, no.1, pp135-148, 

2001. 

[15] Y.S.Shin and J.J.Jeon, “Pseudo Wigner-Ville timefrequency 

distribution and its application to machinery 

condition monitoring,” Journal of Shock and Vibration, 

vol.1, no.4, pp. 65-76, 1993. 

[16] Hui Li, Haiqi Zheng, Liwei Tang, “Wigner-Ville 

Distribution Based on EMD for Faults Diagnosis of 

Bearing,” Lecture Notes in Computer Science, vol.4223 

pp.803-812, 2006. 

[17] W.J.Staszewski, K. Worden and G.R.Tomlinson, “Thefrequency 

analysis in gearbox fault detection using the 

Wigner-Ville distribution and pattern recognition,” 

Mechanical Systems and Signal Processing, vol. 11, no.5, 

pp. 673-692, 1997. 

[18] Hui Li, Yuping Zhang, Haiqi Zheng, “Hilbert-Huang 

transform and marginal spectrum for detection and 

diagnosis of localized defects in roller bearings,” Journal 

of Mechanical Science and Technology, vol.23, no.2, 

pp.291-301, 2009. 

[19] H.Li, H.Q.Zheng, L.W.Tang, “Faults Monitoring and 

Diagnosis of Ball Bearing Based on Hilbert-Huang 

Transformation,” Key Engineering Material, vol.291, 

pp.649-654, 2005. 

[20] Hui Li, Yuping Zhang, Haiqi Zheng, “Wear Detection in 

Gear System Using Hilbert-Huang Transform,” Journal of 


Mechanical Science and Technology, vol.20, no.11, 

pp.1781-1789, 2006. 

[21] W.B.Collis, P.R.White, J.K.Hammond, “Higher-order 

spectra: the bispectrum and trispectrum,” Mechanical 

Systems and Signal Processing, vol.12, no.3, pp.375-394, 

1998. 

[22] J.W.A.Fackrell, P.R.White, J.K.Hammond, R.J.Pinnington, 

“The interpretation of the bispectra of vibration signals- I. 

Theory,” Mechanical Systems and Signal Processing, vol.9, 

no.3, pp. 257-266, 1995. 

[23] J.W.A.Fackrell, P.R.White, J.K.Hammond, R.J.Pinnington, 

“The interpretation of the bispectra of vibration signals- II. 

Experimental results and application,” Mechanical Systems 

and Signal Processing, vol.9, no.3, pp.267-274, 1995. 

[24] Hui Li, Haiqi Zheng, Liwei Tang, “Bearing Fault 

Detection and Diagnosis Based on Order Tracking and 

Teager-Huang Transform,” Journal of Mechanical Science 

and Technology, vol.24, no.3, pp.811-822, 2010. 

[25] Hui Li, Haiqi Zheng, Liwei Tang, “Bearing Fault 

Detection and Diagnosis Based on Teager-Huang 

Transform,” International Journal of Wavelets, 

Multiresolution and Information Processing, vol.7, no.5, 

pp.643-663, 2009. 

Hui Li was born in Hebei province, 

China, on August 23,1968. He received 

his B.S. degree in Mechanical 

Engineering from the Hebei Polytechnic 

University, Hebei, China, in 1991. He 

received his M.S. degree in Mechanical 

Engineering from the Harbin University 

of Science and Technology, Heilongjiang, 

China, in 1994. He received his PhD 

from the School of Mechanical 

Engineering of Tianjin University, Tianjin, China, in 2003. He 

was a postdoctoral researcher in Shijiazhuang Mechanical 

Engineering College from August 2003 to September 2005, and 

in Beijing Jiaotong University from March 2006 to December 

2008. 

He is currently a professor in Mechanical Engineering at 

Shijiazhuang Institute of Railway Technology, China. His 

research and teaching interests include hybrid driven 

mechanism, kinematics and dynamics of machinery, 

mechatronics, CAD/CAPP, signal processing for machine 

health monitoring, diagnosis and prognosis. He has written 

more than 170 papers and conference proceedings. 

Dr. Li is currently a senior member of the Chinese Society of 

Mechanical Engineering.

Aims and Scope. 

Call for Papers and Special Issues 

Journal of Computers (JCP, ISSN 1796-203X) is a scholarly peer-reviewed international scientific journal published monthly for researchers, 

developers, technical managers, and educators in the computer field. It provide a high profile, leading edge forum for academic researchers, industrial 

professionals, engineers, consultants, managers, educators and policy makers working in the field to contribute and disseminate innovative new work 

on all the areas of computers. 

JCP invites original, previously unpublished, research, survey and tutorial papers, plus case studies and short research notes, on both applied and 

theoretical aspects of computers. These areas include, but are not limited to, the following: 

• Computer Organizations and Architectures 

• Operating Systems, Software Systems, and Communication Protocols 

• Real-time Systems, Embedded Systems, and Distributed Systems 

• Digital Devices, Computer Components, and Interconnection Networks 

• Specification, Design, Prototyping, and Testing Methods and Tools 

• Artificial Intelligence, Algorithms, Computational Science 

• Performance, Fault Tolerance, Reliability, Security, and Testability 

• Case Studies and Experimental and Theoretical Evaluations 

• New and Important Applications and Trends 

Special Issue Guidelines 

Special issues feature specifically aimed and targeted topics of interest contributed by authors responding to a particular Call for Papers or by 

invitation, edited by guest editor(s). We encourage you to submit proposals for creating special issues in areas that are of interest to the Journal. 

Preference will be given to proposals that cover some unique aspect of the technology and ones that include subjects that are timely and useful to the 

readers of the Journal. A Special Issue is typically made of 10 to 15 papers, with each paper 8 to 12 pages of length. 

The following information should be included as part of the proposal: 

• Proposed title for the Special Issue 

• Description of the topic area to be focused upon and justification 

• Review process for the selection and rejection of papers. 

• Name, contact, position, affiliation, and biography of the Guest Editor(s) 

• List of potential reviewers 

• Potential authors to the issue 

• Tentative time-table for the call for papers and reviews 

If a proposal is accepted, the guest editor will be responsible for: 

• Preparing the “Call for Papers” to be included on the Journal’s Web site. 

• Distribution of the Call for Papers broadly to various mailing lists and sites. 

• Getting submissions, arranging review process, making decisions, and carrying out all correspondence with the authors. Authors should be 

informed the Instructions for Authors. 

• Providing us the completed and approved final versions of the papers formatted in the Journal’s style, together with all authors’ contact 

information. 

• Writing a one- or two-page introductory editorial to be published in the Special Issue. 

Special Issue for a Conference/Workshop 

A special issue for a Conference/Workshop is usually released in association with the committee members of the Conference/Workshop like 

general chairs and/or program chairs who are appointed as the Guest Editors of the Special Issue. Special Issue for a Conference/Workshop is 

typically made of 10 to 15 papers, with each paper 8 to 12 pages of length. 

Guest Editors are involved in the following steps in guest-editing a Special Issue based on a Conference/Workshop: 

• Selecting a Title for the Special Issue, e.g. “Special Issue: Selected Best Papers of XYZ Conference”. 

• Sending us a formal “Letter of Intent” for the Special Issue. 

• Creating a “Call for Papers” for the Special Issue, posting it on the conference web site, and publicizing it to the conference attendees. 

Information about the Journal and Academy Publisher can be included in the Call for Papers. 

• Establishing criteria for paper selection/rejections. The papers can be nominated based on multiple criteria, e.g. rank in review process plus 

the evaluation from the Session Chairs and the feedback from the Conference attendees. 

• Selecting and inviting submissions, arranging review process, making decisions, and carrying out all correspondence with the authors. 

Authors should be informed the Author Instructions. Usually, the Proceedings manuscripts should be expanded and enhanced. 

• Providing us the completed and approved final versions of the papers formatted in the Journal’s style, together with all authors’ contact 

information. 

• Writing a one- or two-page introductory editorial to be published in the Special Issue. 

More information is available on the web site at http://www.academypublisher.com/jcp/.

A Modified Technique for Analysis of Synchronous Counters Constructed with Flip-flops 

Dangui Yan, Ruijun Tong, Chengchang Zhang, and Changyong Li 

A New Method of Detecting Multi-component LFM Signals Based on Blind Signal Processing 

Qiang Guo, Yajun Li, and Changhong Wang 

Research on Self-built Digital Resource Backup Systems 

Li-zhen Shen 

Configuration Scheme for Small Scale Multi-FPGA Systems 

Chengchang Zhang, Lisheng Yang, Dangui Yan, and Changyong Li 

Order Bi-spectrum For Bearing Fault Monitoring and Diagnosis Under Run-up Condition 

Hui Li 

1971 

1976 

1983 

1988 

1994

(Contents Continued from Back Cover) 

The Analysis of China New Energy Vehicle Industry Alliance Status based on UCINET Software 

Xiongfei Guo and Yingqi Liu 

Efficiency Evaluation Information System Based on Data Envelopment Analysis 

Jing Han and Malin Song 

An Optimal Inventory Control Model for a Supply Chain with Shortage Constraints 

Yinkuan Gu and Hongxia Zhang 

Variable Selection for Credit Risk Model Using Data Mining Technique 

Kuangnan Fang and Hong Huang 

Corporate-, Product-, and User-Image Dimensions and Purchase Intentions —The Mediating Role of 

Cognitive and Affective Attitudes 

Xian Guo Li, Xia Wang, and Yu Juan Cai 

A Microcomputer-Based Predictive Digital Current Programmed Control System for Three-phase 

PWM Rectifier 

Zhongjiu Zheng, Guofeng Li, and Ninghui Wang 

Supply Chain Coordination under Return Policy with Asymmetric Information about Cost of Reverse 

Logistics Operations 


Economic Development and Financial Support for Coal Resource Cities — A Panel Data Analysis 

Zuhuai Yuan, Li Yang, Jing Han, and Keliang Wang 

REGULAR PAPERS 

Solving the Sparsity Problem in Recommender Systems Using Association Retrieval 

YiBo Chen, ChanLe Wu, Ming Xie and Xiaojun Guo 

Integrated Structure and Control Design for Servo System Based on Genetic Algorithm and Matlab 

Dingzhen Li and Ruimin Jin 

A Model to Select System Core and Its Application 

Chongming Li and Yue Ding 

De-noise Comprehensive Research On Airplane Cockpit Signals Recorded by CVR 

Dao-Lai Cheng, Chui-JieYi, and Hong-Yu Yao 

Fuzzy Support Vector Machines Control for 6-DOF Parallel Robot 

Dequan Zhu, Tao Mei, and Lei Sun 

Parameters Optimization of Least Squares Support Vector Machines and Its Application 

Chunli Xi, Cheng Shao, and Dandan Zhao 

The Expected Value Model of Multiobjective Programming and its Solution Method Based on 


Mingfa Zheng , Bingjie Li, and Guangxing Kou 

A Method for Building Partially Connected Neural Network 

Gang Li, Xingsan Qian, Chunming Ye, and Lin Zhao 

A Cooperative Co-evolution PSO for Flow Shop Scheduling Problem with Uncertainty 

Bin Jiao, Qunxian Chen, and Shaobin Yan 

A Double Margin Based Fuzzy Support Vector Machine Algorithm 

Kai Li and Xiaoxia Lu 

1852 

1857 

1862 

1868 

1875 

1880 

1886 

1891 

1896 

1903 

1913 

1920 

1926 

1935 

1942 

1949 

1955 

1962

Journal of Computers - Academy Publisher

You also want an ePaper? Increase the reach of your titles

Delete template?

Save as template?