Statistical Quality Control Charts to Monitor Processes Under ...

Statistical Quality Control Charts to Monitor Processes Under ... Statistical Quality Control Charts to Monitor Processes Under ...

01.08.2014 Views

Intro EWMA CUSUM Comparison Example Future Work References Statistical Quality Control Charts to Monitor Processes Under Competing Risks Lin Fang Department of Mathematics and Statistics McMaster University ( Supervised by Prof. Román Viveros-Aguilera ) 1/23

Intro EWMA CUSUM Comparison Example Future Work References<br />

<strong>Statistical</strong> <strong>Quality</strong> <strong>Control</strong> <strong>Charts</strong> <strong>to</strong><br />

Moni<strong>to</strong>r <strong>Processes</strong> <strong>Under</strong> Competing<br />

Risks<br />

Lin Fang<br />

Department of Mathematics and Statistics<br />

McMaster University<br />

( Supervised by Prof. Román Viveros-Aguilera )<br />

1/23


Intro EWMA CUSUM Comparison Example Future Work References<br />

Outline<br />

1 Introduction<br />

2 Exponentially Weighted Moving Average (EWMA)<br />

<strong>Control</strong> Chart<br />

3 Cumulative Sum (CUSUM) <strong>Control</strong> Chart<br />

4 Comparing the CUSUM <strong>Control</strong> <strong>Charts</strong> with the EWMA<br />

<strong>Control</strong> <strong>Charts</strong><br />

5 Example: A Manufacturing Process<br />

6 Future Work<br />

7 References<br />

2/23


Intro EWMA CUSUM Comparison Example Future Work References<br />

1. Introduction<br />

<strong>Processes</strong> under Competing Risks<br />

Assumptions<br />

System fails due <strong>to</strong> an observed cause.<br />

Failure times are statistically independent with continuous<br />

distributions.<br />

Notations<br />

n (≥ 1) systems sampled, each system is subjected <strong>to</strong><br />

p (≥ 2) competing risks.<br />

T ij is the time <strong>to</strong> failure associated with the j th risk<br />

component of system i ( i = 1, . . . , n , j = 1, . . . , p).<br />

T ij has a pdf. f j (t | θ j ) and a survival function ¯F j (t | θ j ).<br />

3/23


Intro EWMA CUSUM Comparison Example Future Work References<br />

Notations (continued)<br />

Observed data: {(Y 1 , C 1 ), (Y 2 , C 2 ), . . . , (Y n , C n )}.<br />

Y i = min{T 1i , . . . , T pi } and C i = j.<br />

WLOG, the 1 st competing risk component is assumed <strong>to</strong><br />

be the one of interest.<br />

Terminology<br />

In-control and out-of-control parameter values: θ (0)<br />

1<br />

and θ (1)<br />

1<br />

Run length (LR) and average run length (ARL)<br />

In-control and out-of-control ARLs: ARL 0 and ARL 1<br />

4/23


Intro EWMA CUSUM Comparison Example Future Work References<br />

2. EWMA <strong>Control</strong> Chart<br />

Work of Steiner and MacKay (2001)<br />

Conditional Expected Value (CEV) weight<br />

W i = E(T 1i | µ (0)<br />

1 , σ 1, Y i , C i ) =<br />

{<br />

Yi if C i = 1,<br />

R ∞<br />

Y i<br />

if C i ≠ 1 .<br />

t 1i f 1 (t 1i |θ 1 )dt 1i<br />

¯F 1 (Y i |θ 1 )<br />

(1)<br />

EWMA chart statistic is defined as<br />

U l = λ ¯W l + (1 − λ)U l−1 , l = 1, 2, . . . , (2)<br />

where U 0 = µ (0)<br />

1<br />

and smoothing parameter 0 < λ ≤ 1.<br />

5/23


Intro EWMA CUSUM Comparison Example Future Work References<br />

<strong>Control</strong> limits: LCL and UCL.<br />

Performance of EWMA control charts.<br />

Table: The out-of-control ARLs of some selected EWMA charts under exponential<br />

competing risks when n = 1, µ (0)<br />

1 = 1 3 and µ 2 = 1<br />

µ (1)<br />

1 /µ(0) 1<br />

EWMA 0.02 EWMA 0.05 EWMA 0.08 EWMA 0.10 EWMA 0.15 EWMA 0.20<br />

0.25 48.21 50.57 46.94 52.14 91.23 404.90<br />

0.50 481.15 1520.83 1935.13 2967.16 7789.23 34822.51<br />

0.75 129907.48 50874.29 22229.49 16106.06 9704.83 7190.29<br />

1.00 500.46 500.47 500.57 499.47 499.58 500.25<br />

1.25 93.35 92.60 95.35 99.73 110.70 121.01<br />

1.50 49.73 43.97 42.43 43.37 47.10 51.66<br />

1.75 34.64 28.83 26.58 26.53 27.79 29.99<br />

Problem arising due <strong>to</strong> sample size and/or underlying<br />

distribution of competing risks.<br />

6/23


Intro EWMA CUSUM Comparison Example Future Work References<br />

3. CUSUM <strong>Control</strong> Chart<br />

Form of the CUSUM control charts<br />

Upper-sided CUSUM chart<br />

Z + 0<br />

= 0<br />

Z +<br />

l<br />

= max(0, Z +<br />

l−1 + X l + k), l = 1, 2, . . . , (3)<br />

Lower-sided CUSUM chart<br />

Z − 0<br />

= 0<br />

Z −<br />

l<br />

= min(0, Z −<br />

l−1 + X l + k), l = 1, 2, . . . . (4)<br />

Reference value: k (Hawkins and Olwell (1998))<br />

<strong>Control</strong> limit: h + and h − 7/23


Intro EWMA CUSUM Comparison Example Future Work References<br />

Two candidates for the CUSUM score X l<br />

Likelihood function<br />

L(Θ) =<br />

nY<br />

i=1<br />

»<br />

f ci (y i | θ ci )<br />

pY<br />

–<br />

¯Fk (y i | θ k ) . (5)<br />

k≠c i<br />

Score function<br />

∂<br />

8> <<br />

log f<br />

s i (θ (0)<br />

∂θ 1 (y i | θ 1 ) ˛ if c i = 1,<br />

1 ˛θ1<br />

1 ) = =θ (0)<br />

1<br />

∂<br />

> : log ¯F (6)<br />

∂θ 1 (y i | θ 1 ) ˛ if c i ≠ 1 .<br />

1 ˛θ1 =θ (0)<br />

1<br />

Log-likelihood ratio (LLR)<br />

j (1) L(θ 1<br />

LLR = log<br />

) ff<br />

L(θ (0)<br />

1 )<br />

nX<br />

„<br />

«<br />

=<br />

log f 1 (y i | θ (1)<br />

1 ) − log f 1(y i | θ (0)<br />

1 )<br />

{i:c i =1}<br />

nX<br />

„<br />

+ log ¯F 1 (y i | θ (1)<br />

1 ) − log ¯F 1 (y i | θ (0)<br />

1<br />

«. ) (7)<br />

{i:c i ≠1}<br />

8/23


Intro EWMA CUSUM Comparison Example Future Work References<br />

Table: <strong>Control</strong> parameters (ARL 0 = 500) and out-of-control ARLs for score and LLR<br />

CUSUM charts <strong>to</strong> moni<strong>to</strong>r processes under normal competing risks with µ (0)<br />

1 = 0 and<br />

σ 1 = µ 2 = σ 2 = 1.<br />

Score CUSUM<br />

LLR CUSUM<br />

shift in µ 1 k h ARL1 k h ARL1<br />

n = 1<br />

-2.5σ 1 -1.247 -1.84 2.17 -0.036 -1.78 2.17<br />

-2.0σ 1 -0.994 -2.34 3.09 -0.029 -2.26 3.08<br />

-1.5σ 1 -0.739 -3.10 4.87 -0.021 -3.00 4.87<br />

-1.0σ 1 -0.486 -4.40 9.30 -0.013 -4.28 9.28<br />

-0.5σ 1 -0.237 -7.23 26.74 -0.004 -7.12 26.74<br />

0.5σ 1 0.214 7.16 29.31 -0.006 7.30 29.29<br />

1.0σ 1 0.398 4.46 11.46 -0.029 4.67 11.34<br />

1.5σ 1 0.542 3.34 6.87 -0.073 3.58 6.68<br />

2.0σ 1 0.646 2.78 5.08 -0.141 3.08 4.86<br />

2.5σ 1 0.715 2.48 4.27 -0.231 2.88 4.03<br />

9/23


Intro EWMA CUSUM Comparison Example Future Work References<br />

Table: <strong>Control</strong> parameters (ARL 0 = 500) and out-of-control ARLs for score and LLR<br />

CUSUM charts <strong>to</strong> moni<strong>to</strong>r processes under exponential competing risks with µ (0)<br />

1 = 1 3<br />

and µ 2 = 1.<br />

n = 1<br />

Score CUSUM<br />

LLR CUSUM<br />

µ (1)<br />

1 /µ(0) 1<br />

k h ARL1 k h ARL1<br />

0.5 -0.6429 -15.26 22.83 0.0108 -1.55 22.60<br />

0.6 -0.5000 -19.28 34.30 0.0061 -2.00 34.17<br />

0.7 -0.3649 -23.44 53.62 0.0031 -2.53 53.58<br />

0.8 -0.2368 -30.38 92.92 0.0012 -3.29 92.72<br />

0.9 -0.1154 -45.31 187.89 0.0003 -4.96 188.14<br />

1.1 0.1098 41.31 187.16 0.0002 4.63 186.36<br />

1.2 0.2143 35.10 106.58 0.0009 3.96 106.33<br />

1.3 0.3140 30.47 70.34 0.0018 3.48 70.12<br />

1.4 0.4091 27.64 51.69 0.0031 3.17 51.49<br />

1.5 0.5000 25.05 40.41 0.0045 2.91 40.22<br />

10/23


Intro EWMA CUSUM Comparison Example Future Work References<br />

Table: <strong>Control</strong> parameters (ARL 0 = 500) and out-of-control ARLs for score and LLR<br />

CUSUM charts <strong>to</strong> moni<strong>to</strong>r processes under Weibull competing risks with µ (0)<br />

1 = 4.43,<br />

µ 2 = 5.32 and β 1 = β 2 = 2.<br />

n = 1<br />

Score CUSUM<br />

LLR CUSUM<br />

µ (1)<br />

1 /µ(0) 1<br />

k h ARL1 k h ARL1<br />

0.5 -0.144 -1.36 9.77 1.339 -43.12 9.33<br />

0.6 -0.116 -1.72 14.54 0.774 -58.48 14.10<br />

0.7 -0.086 -2.23 23.83 0.389 -82.26 23.48<br />

0.8 -0.056 -3.02 44.54 0.162 -120.50 44.32<br />

0.9 -0.027 -4.54 107.78 0.036 -189.26 107.79<br />

1.1 0.026 4.71 119.01 0.000 212.65 118.94<br />

1.2 0.050 3.75 58.01 0.111 172.15 57.65<br />

1.3 0.072 3.12 36.74 0.196 150.30 36.34<br />

1.4 0.092 2.75 26.68 0.274 136.65 26.08<br />

1.5 0.110 2.51 21.11 0.356 127.31 20.51<br />

11/23


Intro EWMA CUSUM Comparison Example Future Work References<br />

4. Comparison<br />

n = 1<br />

ARL1<br />

0 100 200 300 400 500<br />

LLR CUSUM (−1 sd)<br />

EWMA r = 0.05<br />

EWMA r = 0.10<br />

EWMA r = 0.20<br />

−2.5 −2.0 −1.5 −1.0 −0.5 0.0<br />

µ1<br />

Figure: ARL curves for detecting mean decreases under normal competing risks.<br />

12/23


Intro EWMA CUSUM Comparison Example Future Work References<br />

n = 1<br />

ARL1<br />

0 10 20 30 40<br />

LLR CUSUM (−1 sd)<br />

EWMA r = 0.05<br />

EWMA r = 0.10<br />

EWMA r = 0.20<br />

−1.4 −1.2 −1.0 −0.8 −0.6<br />

µ1<br />

Figure: (Continued) ARL curves for detecting mean decreases under normal<br />

competing risks.<br />

13/23


Intro EWMA CUSUM Comparison Example Future Work References<br />

n = 5<br />

ARL1<br />

0 200 400 600 800 1000<br />

LLR CUSUM (0.5)<br />

EWMA r = 0.05<br />

EWMA r = 0.10<br />

EWMA r = 0.20<br />

0.4 0.6 0.8 1.0<br />

mean ratio<br />

Figure: ARL curves for detecting mean decreases under exponential competing<br />

risks.<br />

14/23


Intro EWMA CUSUM Comparison Example Future Work References<br />

n = 5<br />

ARL1<br />

0 10 20 30 40<br />

LLR CUSUM (0.5)<br />

EWMA r = 0.05<br />

EWMA r = 0.10<br />

EWMA r = 0.20<br />

0.3 0.4 0.5 0.6 0.7<br />

mean ratio<br />

Figure: (Continued) ARL curves for detecting mean decreases under exponential<br />

competing risks.<br />

15/23


Intro EWMA CUSUM Comparison Example Future Work References<br />

n = 1<br />

ARL1<br />

0 200 400 600 800 1000<br />

LLR CUSUM (0.7)<br />

EWMA r = 0.05<br />

EWMA r = 0.10<br />

EWMA r = 0.20<br />

0.4 0.6 0.8 1.0<br />

mean ratio<br />

Figure: ARL curves for detecting mean decreases under Weibull competing risks.<br />

16/23


Intro EWMA CUSUM Comparison Example Future Work References<br />

n = 1<br />

ARL1<br />

0 50 100 150 200<br />

LLR CUSUM (0.7)<br />

EWMA r = 0.05<br />

EWMA r = 0.10<br />

EWMA r = 0.20<br />

0.55 0.60 0.65 0.70 0.75 0.80 0.85<br />

mean ratio<br />

Figure: (Continued) ARL curves for detecting mean decreases under Weibull<br />

competing risks.<br />

17/23


Intro EWMA CUSUM Comparison Example Future Work References<br />

5. Example<br />

Problem from Nelson (1982): Production process s<strong>to</strong>ps<br />

because of shave die or other causes.<br />

Objective: moni<strong>to</strong>r whether the shave dies exhibit shorter<br />

lifetimes.<br />

Failure times observed: 11 (shave die) vs. 26 (other<br />

causes)<br />

Weibull distribution is fitted for both groups.<br />

Estimated in-control MLEs<br />

β die = 2.76 and η die = 78.35 (i.e. µ die = 69.73)<br />

β other = 2.44 and η other = 57.21 (i.e. µ other = 50.73)<br />

18/23


Intro EWMA CUSUM Comparison Example Future Work References<br />

Survival Probability s(t)<br />

0.0 0.2 0.4 0.6 0.8 1.0<br />

Shave Die<br />

Other Causes<br />

0 20 40 60<br />

Time (t)<br />

Figure: Kaplan-Meier curves and estimated survival functions by using Weibull<br />

MLEs.<br />

19/23


Intro EWMA CUSUM Comparison Example Future Work References<br />

CUSUM Chart for Shave Die<br />

CUSUM Chart for Other Causes<br />

LLR CUSUM<br />

−200000 −100000 0<br />

●<br />

● ●<br />

● ● ● ● ● ● ● ●<br />

●<br />

●<br />

●<br />

●<br />

●<br />

●<br />

●<br />

●<br />

●<br />

●<br />

●<br />

●<br />

●<br />

●<br />

●<br />

0 5 10 15 20 25<br />

LLR CUSUM<br />

−150000 −100000 −50000 0<br />

● ● ● ● ● ●<br />

● ●<br />

● ● ● ● ● ● ●<br />

● ● ●<br />

●<br />

●<br />

●<br />

●<br />

●<br />

●<br />

●<br />

●<br />

0 5 10 15 20 25<br />

Sample<br />

Sample<br />

(a)<br />

(b)<br />

Figure: An example that mean operating hours of shave die decreases after sample<br />

20.<br />

20/23


Intro EWMA CUSUM Comparison Example Future Work References<br />

CUSUM Chart for Shave Die<br />

CUSUM Chart for Other Causes<br />

LLR CUSUM<br />

−120000 −80000 −40000 0<br />

● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●<br />

●<br />

●<br />

●<br />

●<br />

●<br />

●<br />

●<br />

●<br />

●<br />

●<br />

●<br />

15 0 5 10 20 25 30 35 40 45<br />

LLR CUSUM<br />

−140000 −100000 −60000 −20000<br />

● ● ● ● ● ● ● ● ● ● ● ●<br />

● ● ● ● ● ● ● ● ●<br />

●<br />

●<br />

●<br />

●<br />

●<br />

●<br />

●<br />

●<br />

●<br />

●<br />

●<br />

●<br />

●<br />

●<br />

●<br />

●<br />

●<br />

●<br />

●<br />

●<br />

● ●<br />

● ● ●<br />

0 5 10 15 20 25 30 35 40 45<br />

Sample<br />

Sample<br />

(a)<br />

(b)<br />

Figure: An example that mean operating hours of other causes decreases after<br />

sample 20.<br />

21/23


Intro EWMA CUSUM Comparison Example Future Work References<br />

Future Work<br />

Masked failure problem: the cause(s) of failure could not<br />

be fully identified.<br />

Dependent risks: If the assumption of independence is no<br />

valid, how <strong>to</strong> extend the methods <strong>to</strong> take in<strong>to</strong> account<br />

correlation among the competing risks?<br />

22/23


Intro EWMA CUSUM Comparison Example Future Work References<br />

References<br />

Hawkins, D. M. and Olwell, D. H. (1998), Cumulative Sum<br />

<strong>Charts</strong> and Charting for <strong>Quality</strong> Improvement, New York:<br />

Springer-Verlag.<br />

Nelson, W. B. (1982), Applied Life Data Analysis, New<br />

Jersey: John Wiley & Sons.<br />

Stainer, S. H. and MacKay, R. J. (2001), Moni<strong>to</strong>ring<br />

<strong>Processes</strong> with Data Censored Owing <strong>to</strong> Competing Risks<br />

by Using Exponentially Weighted Moving Average <strong>Control</strong><br />

<strong>Charts</strong>. Applied Statistics, 50, part 3, pp. 293-302.<br />

23/23

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!