Professional Documents
Culture Documents
Abstract: Variation reduction of manufacturing processes is an essential objective of process quality improvement. It is highly
desirable to develop a methodology of variation source identication that helps quickly identify the variation sources, hence
leading to quality improvement and cost reduction in manufacturing systems. This paper presents a variation source identication
method based on the analysis of the covariance matrix of process quality measurements. The identication procedure utilizes the
fact that the eigenspace of the quality measurement covariance matrix can be decomposed into a subspace due to variation sources
and a subspace purely due to system noise. The former subspaces for different samples will be the same if the same variation
sources dominate. A testing procedure is presented, which can determine the closeness of the subspaces under sampling
uncertainty. A case study is conducted to illustrate the effectiveness of this methodology. 2006 Wiley Periodicals, Inc. Naval
Research Logistics 53: 383396, 2006.
Keywords: eigenspace; factor analysis; principal component analysis; variation reduction; variation source identication
1.
INTRODUCTION
384
(1)
2.
Problem Formulation
(2)
385
gf
f
f .
(3)
ff0
gf
f ff0
as A, then
y A f .
(4)
386
(5)
n
jp1
j, where 2 is the variance of the
measurement noise. Matrix R is an arbitrary p by p orthogonal matrix. Clearly, the estimation of A* is not unique. It
possesses rotational indeterminacy. Based on the result in
(6), we can get the corresponding estimation result for A,
2
A ML D1/2 Up p ML
I1/2 R,
(7)
(8)
(6)
2
2f 1 ML
,
2.3.
2
ML
and
387
(9)
In this variation source identication procedure, two important testing procedures are required: (1) the number of
signicant eigenvalues of a covariance matrix (i.e., the
number of faults in the system) and (2) determination of
current faults based on the selected fault geometry vectors.
The testing methods are discussed in Section 2.4.
REMARKS:
388
2.4.
The proposed procedure also requires process inspection before a fault geometry vector can be added
to the library. Process inspection is very important in
the proposed variation source identication procedure because the whole procedure is built up on
statistical testing results. Due to the inherent uncertainty of statistical tests, there will always be errors
in the conclusions of the statistical tests. We could
have a false alarm or miss a detection in the statistical tests. These errors are inherent: we can only try
to reduce them by increasing the sample size or
developing more efcient test statistics, but we can
never eliminate them. Thus, the process inspection is
a necessary step in the procedure to eliminate the
inuence of statistical testing error. For example, we
can depend on process inspection for decision making when a multiple match occurs; i.e., if multiple
matches are found, process inspection is needed to
make the nal decision on which faults have occurred. In practice, we also need process inspection
to identify the physical mechanism of the new fault
and then eliminate it. After all, it is meaningless to
store a fault geometry vector of an unknown fault
into the library. One point that needs to be emphasized is that the proposed procedure only needs process
inspection when a new fault (that has not occurred
before) occurs. Compared with traditional process control techniques such as SPC, in which process inspection is needed for every fault, the process inspection
efforts are signicantly reduced.
In this variation source identication procedure, we
do not put the fault geometry vectors from multiple
fault case in the library due to their rotational indeterminacy. According to Eq. (7), when multiple
faults occur, the estimated fault geometry vector
could be of an arbitrary rotation within the linear
space spanned by the corresponding multiple column vectors of A matrix. It cannot be utilized for
fault diagnosis.
Testing Procedures Used in Variation Source
Identication
2.4.1. Estimation of the Number of Faults
Different methods to solve this problem have been studied and compared through simulation by Apley and Shi [4].
This problem can be formulized as a hypothesis testing
problem. Denote Sy as the sample covariance matrix of the
n-dimensional quality measurements. Then, to determine
whether p faults exist in the system, we can test the hypothesis that 1 2 . . . p 2 p1 . . . n ,
where i , i 1 . . . n, are the eigenvalues of y. Several
Naval Research Logistics DOI 10.1002/nav
(10)
and
MDLl Nn llogal /gl l2n llogN/2,
(11)
389
390
2 since C
12
12 n 1 2
VR
Type I
error
50
10
50
100
150
200
50
100
150
200
50
100
150
200
0.1016
0.0503
0.0207
0.0132
0.1419
0.0635
0.0244
0.0102
0.188
0.0658
0.0246
0.0089
20
30
VR
Type I
error
100
10
50
100
150
200
50
100
150
200
50
100
150
200
0.1163
0.0552
0.0221
0.0127
0.1804
0.0673
0.0275
0.008
0.2489
0.0783
0.0285
0.0064
20
30
391
3.1.
3. CASE STUDY
Introduction to the Process
Figure 3. Illustration of a machining process. [Color gure can be viewed in the online issue, which is available at www.interscience.wiley.com.]
Naval Research Logistics DOI 10.1002/nav
392
0.1968
0.3933
0.2154
0.2352
0.2432
0.3947
0.0971
0.0894
0.3946
0.0249
0.0463
0.3945
0.4775
0.0985
0.0256
0.1879
0.2277
0.3943
0.3708
0.0427
0.0464
0.3117
0.3655
0.3942
0.2329
0.1866
0.0374
T.
0.3607
0.2864
0.2699
0.1240
0.3079
0.0373
3.2.
0.4322
0.1715
0.0891
0.0193
0.4243
0.0372
0.0299
0.3325
0.0991
0.1011 T
0.2178 .
0.2795
1
2
3
393
10
11
12
13
14
15
1
2
3
4
5
6
7
8
0.2423
5.5265
12.4105
6.2765
12.6120
23.6044
7.2109
11.4467
0.1995
0.2165
0.1758
0.2100
11.4840
8.8845
0.1933
6.6512
0.1729
0.1824
0.1713
0.1791
0.2248
0.1866
0.1524
0.2322
0.1534
0.1464
0.1419
0.1547
0.2026
0.1742
0.1452
0.1802
0.1266
0.1406
0.1242
0.1462
0.1502
0.1485
0.1299
0.1515
0.1181
0.1289
0.1204
0.1093
0.1358
0.1231
0.1200
0.1370
0.0959
0.1007
0.1102
0.1012
0.1230
0.1199
0.1018
0.1158
0.0905
0.0945
0.1006
0.0936
0.1056
0.0881
0.0967
0.0918
0.0824
0.0845
0.0901
0.0862
0.1000
0.0803
0.0794
0.0765
0.0663
0.0679
0.0762
0.0695
0.0789
0.0706
0.0693
0.0718
0.0595
0.0538
0.0694
0.0620
0.0695
0.0628
0.0663
0.0666
0.0543
0.0460
0.0544
0.0521
0.0563
0.0549
0.0515
0.0517
0.0472
0.0439
0.0445
0.0395
0.0503
0.0500
0.0481
0.0464
0.0448
0.0343
0.0388
0.0347
0.0444
0.0353
0.0411
0.0408
0.0302
0.0262
0.0327
0.0298
0.0278
0.0219
0.0227
0.0274
394
Cases
MDL(0)
MDL(1)
MDL(2)
MDL(3)
MDL(4)
MDL(5)
No. of faults
123.3794
163.9793
200.8788
236.4687
270.5744
302.0241
0
1025.1
176.7
209.1
238.1
267.9
299.2
1
1344.5
151.1
190.9
226.7
260.1
291.4
1
1291.4
185.8
219.9
251.5
281.3
308.2
1
2006.5
1630.6
197.3
230.4
263.1
290.9
2
1941.4
1231.7
235.7
262.3
296.5
325.1
2
1354
172.4
205.7
240
272.2
302.3
1
1515.9
986.4
199.8
231.9
261.5
287.2
2
Important eigenvectors
0.1923
0.2424 0.1008 0.0389
0.4096
0.2611 0.0949 0.0589
0.4048
0.2335 0.1038
0.0065
0.4365
0.3169 0.1545
0.0305
0.1581
0.3659 0.3792
0.3976
0.1507 0.0081 0.0111 0.0155
0.3651
0.2780 0.1109 0.0327
0.2236
0.3940 0.3915
0.4014
0.2272
0.3952 0.4202
0.3863
0.1939
0.2781 0.1137 0.0139
0.1768
0.2116
0.1841
0.1333
0.4022
0.0243
0.2477
0.3958
0.4069
0.1642
0.2985
0.3814
0.4091
0.2986
0.4358
0.0028
0.4050
0.3729
0.3828
0.3140
0.3708
0.2802
0.2927
0.2391
0.3053
0.0726
0.3701
0.2814
0.2165
0.3823
0.4325
0.1823
0.1517
0.1490
0.1077
0.2199
0.3508
0.0843
0.0681
0.4242
0.4872
0.1108
0.0926
0.1195
0.0224
0.3912
0.3116
0.0308
0.0647
0.4587
* The eigenvectors of Case 2, 3, and 7 are denoted v1 , v2, and v3, respectively.
Naval Research Logistics DOI 10.1002/nav
0.3383
0.0460
0.0264
0.0452
0.0686
0.3876
0.1678
0.0517
0.0773
0.3660
0.2712
0.1665
0.1704
0.1918
0.0659
0.4001
0.0166
0.0531
0.0331
0.2268
0.1198
0.3108
0.2854
0.3065
0.0705
0.4578
0.1239
0.0328
0.0263
0.1086
0.0041
0.4063
0.4331
0.4000
0.0763
0.4023
0.2548
0.0306
0.0263
0.0149
0.0598
0.3231
0.3223
0.3508
0.0534
0.2871
0.2181
0.0898
0.1022
0.0327
0.0653
0.2032
0.2228
0.2606
0.2441
0.0734
0.1929
0.2861
0.2824
0.1070
395
APPENDIX
VR
50
10
50
100
150
200
50
100
150
200
50
100
150
200
50
100
150
200
15
20
30
Table A1. Critical angles (degree) and summary statistics ( 2 0.01 2 , p 1).
90%
95%
99%
N
n
VR
90%
4.01
3.14
2.83
2.61
5.02
3.97
3.55
3.32
5.90
4.63
4.15
3.88
7.31
5.77
5.14
4.80
VR
50
10
50
75%
50%
33%
75%
50%
33%
75%
50%
33%
75%
50%
33%
75%
50%
33%
75%
50%
33%
75%
50%
33%
75%
50%
33%
75%
50%
33%
75%
50%
33%
75%
50%
33%
75%
50%
33%
100
150
15
50
100
150
20
50
100
150
30
50
100
150
0.28
0.33
0.36
0.38
1.01
0.74
0.62
0.54
0.29
0.36
0.41
0.42
0.31
0.40
0.44
0.46
5.29
4.08
3.62
3.30
6.34
4.93
4.34
4.01
7.26
5.60
4.96
4.59
8.74
6.79
5.98
5.53
5.71
4.38
3.86
3.51
6.78
5.23
4.59
4.23
7.70
5.92
5.22
4.81
9.21
7.12
6.25
5.76
6.53
4.95
4.34
3.93
7.63
5.84
5.08
4.66
8.59
6.54
5.72
5.24
10.15
7.77
6.78
6.22
100
10
15
20
30
50
100
150
200
50
100
150
200
50
100
150
200
50
100
150
200
2.79
2.20
1.97
1.83
3.51
2.76
2.47
2.32
4.11
3.25
2.90
2.71
5.10
4.03
3.58
3.37
0.17
0.22
0.24
0.26
0.68
0.50
0.42
0.37
0.19
0.23
0.27
0.29
0.19
0.24
0.26
0.30
3.65
2.84
2.51
2.30
4.39
3.41
3.01
2.79
5.01
3.90
3.45
3.19
6.03
4.70
4.14
3.85
Table A2. Critical angles (degree) and summary statistics ( 2 0.01 2 , p 2).
90%
95%
99%
N
n
VR
C
7.04
7.41
7.76
6.87
7.03
7.22
6.75
6.86
6.99
7.68
8.13
8.73
7.24
7.40
7.64
7.15
7.33
7.41
8.24
8.62
9.28
7.59
7.92
8.36
7.42
7.61
7.92
9.12
9.55
10.28
8.24
8.60
8.99
7.91
8.17
8.58
2.02
2.04
1.86
2.21
2.31
2.22
2.33
2.24
2.20
1.87
1.89
1.91
2.09
2.11
2.04
2.18
2.14
2.08
1.77
1.75
1.58
2.01
2.03
1.90
2.15
2.11
1.95
1.58
1.58
1.40
1.85
1.82
1.65
1.93
1.88
1.75
7.67
8.14
8.76
7.25
7.50
7.86
7.04
7.22
7.47
8.43
8.96
9.82
7.70
7.96
8.39
7.50
7.76
7.98
9.08
9.53
10.45
8.11
8.52
9.15
7.82
8.08
8.54
10.13
10.58
11.54
8.86
9.28
9.89
8.38
8.71
9.29
7.91
8.41
9.15
7.40
7.67
8.10
7.15
7.35
7.65
8.71
9.26
10.23
7.87
8.16
8.66
7.63
7.91
8.19
9.39
9.86
10.87
8.29
8.73
9.43
7.97
8.25
8.77
10.49
10.95
11.98
9.08
9.52
10.21
8.54
8.90
9.54
8.42
8.98
9.94
7.70
8.04
8.60
7.37
7.63
8.03
9.29
9.89
11.06
8.21
8.56
9.22
7.88
8.22
8.62
10.03
10.54
11.75
8.68
9.16
10.02
8.26
8.58
9.23
11.23
11.70
12.90
9.53
10.00
10.86
8.88
9.28
10.05
100
10
50
100
150
15
50
100
150
20
50
100
150
30
50
100
150
75%
50%
33%
75%
50%
33%
75%
50%
33%
75%
50%
33%
75%
50%
33%
75%
50%
33%
75%
50%
33%
75%
50%
33%
75%
50%
33%
75%
50%
33%
75%
50%
33%
75%
50%
33%
5.82
6.00
6.10
5.49
5.66
5.70
5.46
5.56
5.66
6.08
6.18
6.58
5.73
5.82
6.12
5.64
5.75
5.93
6.49
6.50
7.09
6.02
6.14
6.38
5.92
5.94
6.19
7.00
7.15
7.72
6.53
6.54
6.88
6.17
6.34
6.60
1.88
1.82
1.74
1.96
1.93
1.87
1.95
1.96
1.90
1.80
1.74
1.68
1.88
1.78
1.72
1.89
1.90
1.84
1.72
1.57
1.52
1.77
1.72
1.63
1.84
1.76
1.76
1.45
1.39
1.36
1.71
1.58
1.55
1.72
1.68
1.62
95%
99%
3.93
3.04
2.67
2.44
4.67
3.61
3.18
2.94
5.29
4.10
3.61
3.33
6.31
4.91
4.31
4.00
4.45
3.41
2.98
2.71
5.20
3.99
3.49
3.21
5.82
4.49
3.93
3.61
6.87
5.30
4.63
4.28
90%
95%
99%
6.18
6.43
6.71
5.72
5.94
6.09
5.63
5.78
5.95
6.52
6.68
7.25
6.01
6.15
6.55
5.86
6.01
6.27
6.98
7.05
7.78
6.32
6.51
6.86
6.16
6.23
6.56
7.57
7.77
8.45
6.88
6.95
7.41
6.45
6.67
7.02
6.32
6.59
6.93
5.81
6.04
6.23
5.70
5.86
6.05
6.68
6.86
7.48
6.10
6.26
6.70
5.93
6.10
6.39
7.15
7.24
8.02
6.43
6.63
7.02
6.24
6.32
6.69
7.77
7.98
8.70
7.00
7.09
7.59
6.54
6.78
7.16
6.60
6.91
7.38
5.98
6.24
6.52
5.82
6.01
6.27
7.00
7.22
7.95
6.30
6.49
7.02
6.08
6.28
6.63
7.50
7.62
8.50
6.64
6.87
7.35
6.39
6.51
6.95
8.16
8.38
9.19
7.24
7.36
7.94
6.72
6.99
7.44
396
ACKNOWLEDGMENT
The nancial support of this work is provided by NSF
Award DMI-0322147. The authors appreciate the editors
and referees valuable comments and suggestions.
REFERENCES
[1] R.D. Anderson and H. Rubin, Statistical inference in factor
analysis, in Proceedings of the Third Berkeley Symposium
of Mathematical Statistics and Probability, Vol. 5, 1956, pp.
111150.
[2] T.W. Anderson, Asymptotic theory for principal component
analysis, Ann Math Stat 34 (1963), 122148.
[3] D.W. Apley and J. Shi, Diagnosis of multiple xture faults in
panel assembly, ASME J Manufact Sci Eng 120 (1998),
793 801.
[4] D.W. Apley and J. Shi, A factor-analysis methods for diagnosing variability in multivariate manufacturing processes,
Technometrics 43 (2001), 84 95.
[5] D.W. Apley and H.Y. Lee, Identifying spatial variation patterns in multivariate manufacturing processes: A blind separation approach, Technometrics 45(3) (2003), 220 234.
[6] R.R. Barton and D.R. Gonzalez-Barreto, Process-oriented
basis representations for multivariate process diagnostics,
Qual Eng 9(1) (1996), 107118.
[7] R.J. Boik, Spectral models for covariance matrices, Biometrika 89 (2002), 159 182.
[8] A.J. Camelio, S.J. Hu, and D.J. Ceglarek, Modeling variation propagation of multi-station assembly systems with
compliant parts, in Proceedings of the 2001 ASME Design
Engineering Technical Conference, Pittsburgh, PA, September 9 12, 2001.
[9] D. Ceglarek and J. Shi, Dimensional variation reduction for
automotive body assembly, J Manufact Rev 8 (1995), 139
154.
[10] D. Ceglarek and J. Shi, Fixture failure diagnosis for autobody
assembly using pattern recognition, ASME J Eng Ind 188
(1996), 55 65.
[11] M. Chang and D.C. Gossard, Computational method for
diagnosis of variation-related assembly problem, Int J Prod
Res 36 (1998), 29852995.
[12] L.H. Chiang, E.L. Russell, and R.D. Braatz, Fault detection
and diagnosis in industrial systems, Springer, London, 2001.
[13] K. Choochaow, Dimension reduction in PCA: Likelihoodbased methods. Ph.D. dissertation, Montana State University,
Bozeman, MT, 2002.
[14] Y. Ding, D. Ceglarek, and J. Shi, Modeling and diagnosis of
multistage manufacturing processes: Part I state space
model, in Proceedings of the 2000 Japan/USA Symposium
on Flexible Automation, July 2326, 2000, Ann Arbor, MI,
2000JUSFA-13146.
[15] Y. Ding, D. Ceglarek, and J. Shi, Fault diagnosis of multistage manufacturing processes by using state space approach, in ASME Transactions, Journal of Manufacturing
Science and Engineering, 2002, pp. 313322.
[16] Y. Ding, J. Shi, and D. Ceglarek, Diagnosability analysis of
multistage manufacturing processes, ASME J Dynam Syst
Measure Contr 124 (2002), 113.
[17] Y. Ding, S. Zhou, and Y. Chen, A comparison of process
variation estimators for in-process dimensional measurements and control, ASME Trans J Dynam Syst Measure
Contr 127 (2005), 69 79.
Naval Research Logistics DOI 10.1002/nav