Download as pdf or txt
Download as pdf or txt
You are on page 1of 12

QUALITY AND RELIABILITY ENGINEERING INTERNATIONAL

Qual. Reliab. Engng. Int. 2006; 22:225–236


Published online 16 December 2005 in Wiley InterScience (www.interscience.wiley.com). DOI: 10.1002/qre.712

Research Quality Evaluation of a


Manufactured Product with
Multiple Characteristics
F. K. Wang∗,†
Department of Industrial Management, National Taiwan University of Science and Technology, Taipei, Taiwan 106, R.O.C.

A majority of manufactured products have many quality characteristics that are


important to the customer. To assess or evaluate the capability of a manufactured
product with multiple characteristics, the quality characteristics need to be
determined in advance and some characteristics can be correlated with each other.
Also, the specification limits of the quality characteristics can be one-sided or two-
sided. In order to deal with such multivariate data, there is a need to develop a
new approach. In this study, the geometric distance variable is used to combine the
correlated or uncorrelated quality characteristics. Then, the Luceño capability index
is used to summarize the performance for each geometric distance variable. Finally, a
composite capability index, MCpc , composed of several univariate capability indices,
is proposed to analyze the capability of a manufactured product with multiple
characteristics. In addition, the probability of the product non-conforming is also
proposed. The application of the proposed methodology to a real-life case study is
presented. Copyright  c 2005 John Wiley & Sons, Ltd.

KEY WORDS: composite capability index; manufactured product; quality evaluation

1. INTRODUCTION

Q uantitative assessment of the quality of a manufactured product and using the assessment for
competitive benchmarking and diagnostics of manufactured product failure are very important for
continuous improvement in modern manufacturing industries. Capability study offers an excellent
opportunity to quantify how well a process can meet customer requirements. The capability indices, Cp , Cpk ,
and Cpm , are widely used to evaluate the process performance based on a single engineering specification
in the industry1. However, many process problems in which several related variables are of interest are
particularly important today. The automatic inspection procedure can make it relatively easy to measure
many parameters on each manufactured product. That is, capability analysis usually involves more than one
engineering specification. Kotz and Johnson2,3 reviewed the multivariate process capability indices thoroughly

∗ Correspondence to: Fu-Kwun Wang, Associate Professor, Department of Industrial Management, National Taiwan University of Science
and Technology, No. 43, Keelung Road, Sec. 4, Taipei, Taiwan 106, R.O.C.
† E-mail: fukwun@mail.ntust.edu.tw

Contract/grant sponsor: National Science Council of Taiwan; contract/grant number: NSC-92-2416-H-027-005

Received 16 March 2003


Copyright 
c 2005 John Wiley & Sons, Ltd. Revised 30 November 2004
226 F. K. WANG

Table I. Development of the multivariate capability index

Index Reference
 1/2
nr
Cpm = n Chan et al.4
i=1 (xir − T )T A−1 (xir − T )
 −1
k 1/v (µ − T )T A−1 (µ − T )
v Cp = = Cp ; v Cpm = v Cp 1 + Pearn et al.5
kv v
v
volume(XT A−1 X ≤ K 2 )
where Cp =
volume(XT A−1 X ≤ χv,0.9973
2 )
MCR = |SA−1 | + (µ − T )T A−1 (µ − T ) Kotz and Johnson2
E[L(X)] = tr(A−1 S) + (µ − T )T A−1 (µ − T )
where L(X) = (X − T )T A−1 (X − T )

MCpk = 13 −1 (p) Wierda6


 V (modified specification region)

 C = 0 1/2
Cp  p
 |S| (π k)1/2 ((v/2 + 1))−1
MCpm =  1/2 Taam et al.7
D
where
 n

 D= 1+ (x̄ − µ0 )T S −1 (x̄ − µ0 )
 n−1
r0
MCp = where r = min{C : Pr(h(X − µ0 ≥ C)) ≥ 1 − α} Chen8
r
[CpM , PV, LI] where Shahriari et al.9


V (USL − LSL ) 1/V
i=1 i i v(n − 1)
CpM = V , PV = P T 2 > F (v, n − v)
n−v
i=1 (UPLi − LPLi )

1 if process region within tolerance region
LI =
0 otherwise
volume of window of opportunity
Vrn = Veevers10
volume of specification region
Multivariate capability index using the geometric distance approach Wang and Hubele11

Multivariate capability index (normal and non-normal data) using principal component analysis Wang and Du12

Non-parametric multivariate capability index Yeh and Chen13


α
MCf = 0 where α = 1 − Pr(h(X − T ) ≤ r)
α
Quality evaluation of GT&T in form and location Wang and Hubele14

for assessing multivariate process data. Furthermore, the development of a multivariate capability index is shown
in Table I. According to the literature review2–14 , the multivariate process capability indices can be obtained
from:

(1) the ratio of a tolerance region to a process region2,3,5,6,10 ;


(2) the function of the multivariate probability distribution which is used to compute the probability of the
non-conforming product2,6,8,13 ;
(3) other approaches that consist of loss functions and vector representation2,5,9 ; and
(4) the multivariate capability index, which is based on the geometric distance approach and principal
component analysis11,12,14 .

Here, we may conclude that four constraints exist in the conventional methods. First, the normality
assumption on multivariate data is usually required. Second, the confidence intervals of the multivariate
capability indices are difficult to derive (this is because the quality measurements have enormous amounts of

Copyright 
c 2005 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2006; 22:225–236
QUALITY EVALUATION OF A MANUFACTURED PRODUCT 227

correlated data and the lack of theoretical distribution of the multivariate capability indices). Third, the estimated
probability of non-conforming for some conventional methods is not available. Finally, the computation for
high dimensions (such as more than three quality variables) is not obtainable except by the geometric distance
approach and the principal component analysis method in Wang and Hubele11 and Wang and Du12. From the
above discussion, it is clear that application of the conventional methods is limited. Furthermore, a manufactured
product may have multiple quality characteristics, some of which may be correlated and the rest may be
uncorrelated. Also, the specification limits of the quality characteristics can be one-sided or two-sided. In order
to deal with such multivariate data, there is a need to develop a new approach. In this study, the assumption of
normality for quality measurement data is not sustained. The geometric distance variable is used to combine
the correlated or uncorrelated quality characteristics. Then, the Luceño capability index15 is used to summarize
the performance for each geometric distance variable. Finally, a composite capability index, MCpc , composed
of several univariate capability indices, is proposed to analyze the capability of a manufactured product with
multiple characteristics. In addition, the probability of the product non-conforming is also presented. Section 2
presents the analysis procedure with a performance metric for the correlated quality characteristics and a
composite capability index. The case study is used to demonstrate the application of the proposed methodology
in Section 3. Conclusions are then made along with suggestions for further research in the final section.

2. THE ANALYSIS PROCEDURE


In this section, the analysis procedure of a manufactured product with multiple characteristics is presented in
Figure 1. The first step is to determine the correlated and uncorrelated quality characteristics. The following
subsections present the procedure of obtaining a performance metric for the correlated quality characteristics
and a composite capability index. Examples will be provided to demonstrate the proposed methodology.

2.1. A performance metric for the correlated quality characteristics


When these quality characteristics are related, the analysis should be based on a multivariate statistical
technique. While numerous authors have recently proposed alternative definitions for multivariate capability
indices, those methods may not be practical in some cases11 . Wang and Hubele11 proposed a new process
variable, geometric distance (GD), which reduces dimensionality of the measurement data and uses any type
of multivariate process data. The geometric distance metric used was the Euclidean distance (or L2 norm) and
only for two-dimensional (2D) and 3D data. For example, consider the position of a hole. Using the L2 norm,
let (x, y, z) be an inspection point, where (x, y) represents the measured points of the center of a hole and z is
the measured diameter of a hole. Also, let (x ∗ , y ∗ , z∗ ) be the nominal target location, where (x ∗ , y ∗ ) represents
the nominal point of the center of a hole and z∗ is the nominal diameter of a hole. Then the geometric distance
variable can be defined by

G= (x − x ∗ )2 + (y − y ∗ )2 + (z − z∗ )2 (1)

In order to justify the underlying distribution of this geometric distance metric, a goodness of fit technique
is implemented to determine distribution with the best fit. Rodriguiz16 suggested fitting data to actual families
of distributions such as the gamma, lognormal, or Weibull. Parameter estimation is fairly straightforward with
these standard distributions and goodness-of-fit tests are available. A goodness-of-fit technique can be used to
ascertain the underlying distribution in different situations. The procedure is based on a link between a general
composite hypothesis H0 and the problem of testing normality when both mean and variance are unknown.
Let Fn (x) denote the empirical distribution function of the sample, that is,


n
Fn (x) = n−1 I [xi ≤ x]
i=1

Copyright 
c 2005 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2006; 22:225–236
228 F. K. WANG

C o llect
1 .M u ltip le q u ality ch aracteristics and
to leran ce reg io n
2 . M easu rem en t d ata

1 . D eterm in e th e co rrelated and F eed b ack


u n co rrelated q u ality ch aracteristics 1 . P ro cess is in cap ab le
2 . C o m p u te: (a) G eo m etric d istan ce 2 . A ctio n to im p ro v e th e p ro cess
v ariab les, ( b ) M R D (m ax im u m rad ial p erfo rm an ce
d istan ce) 3 . C o m p u te th e in d ex u sin g t h e g eo m etric
d istan ce ap p ro ach as a referen ce v alu e
to see th e cu rren t p ro cess p erfo rm an ce

A re all m easu rem en t p o in ts No


w ith in to leran ce reg io n ?

Y es

U sin g B est-F it so ftw are


to fin d th e b est-fitting d istrib u tio n
fo r each g eo m etric d istan ce v ariab le

C o m p u te
1 . P o in t estim atio n a n d co n fid en ce in terv al
for each geo m etric d istan ce v ariab le
2 . T h e p ro b ab ility o f t h e p r o d u c t n o n - co n fo rm in g
fo r each g eo m etric d istan ce v ariab le
F in ally , co m p u te th e co m p o site cap ab ility
in d ex w ith th e p ro b ab ility

Figure 1. A flowchart of the analysis procedure

where if I [a ≤ b] = 1 if a ≤ b, and I [a ≤ b] = 0 if a > b. Any statistic that measures the difference between
Fn and F is referred to as an empirical distribution function (EDF) statistic.
Best-Fit software17 provides 26 distribution types for a goodness-of-fit test based on three critical values
of the chi-square test, the Kolmogorov–Smirnov test and the Anderson–Darling test. Using this software, the
best-fitting distribution of the geometric distance variable can be easily determined.
The intent of proposing the geometric distance metric is to reduce dimensionality of the measurement data.
While these data are now univariate, when they do not conform to a normal distribution, then the traditional

Copyright 
c 2005 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2006; 22:225–236
QUALITY EVALUATION OF A MANUFACTURED PRODUCT 229

indices (e.g. Cp and Cpk ) are not suitable. In such cases a suitable capability index such as the reliable capability
index, Cpc , as proposed by Luceño15, may be useful. Here, Cpc is given by

 USL − LSL
 6√π/2E|X − m|
 for two-sided specifications
Cpc = (2)

 USL − m
 √ for one-sided specification
3 π/2E{max(X − m, 0)}

where m = (USL + LSL)/2 for the two-sided specification case and m is close to, or equal to, zero for the
one-sided case. The index takes into account the process location and variability, and its confidence interval,
based on the central limit theorem, does not require the assumption of normality of the underlying data. A
100 × (1 − α)% confidence interval for Cpc is given by

Ĉpc
√ (3)
1 ± tα/2, n−1 sc /(c̄ n)

where

n  
n 
c̄ = |Xi − m| n and sc2 = |Xi − m|2 − nc̄2 n−1
i=1 i=1

The Luceño index for one-sided specifications is proposed here as the performance metric when the distance
data cannot be fitted to a normal distribution. In this case, m = 0 and USL = MRD, where MRD is the maximum
radial distance from the target to a perimeter of the tolerance region, as defined by the Euclidean distance.
Furthermore, the upper bound on the estimated proportion non-conforming is given by
 MRD
1 − F (MRD) = 1 − f (x) dx (4)
0

where f (x) is the density function of the best-fitting distribution. In order to demonstrate the proposed
methodology, two examples are used to illustrate the implementation of the proposed methodology.
Example 1. A geometric dimensioning and tolerancing (GD&T) drawing specifies a target value for the hole
diameter corresponding to the midpoint of allowable hole sizes and the allowable perpendicularity of the hole
depending on its size. The specifications call for a hole diameter between 0.39 and 0.43 and for the center line of
the hole to be within a cylinder of diameter 0.04 at maximum material condition (i.e. maximum hole diameter),
increasing to diameter 0.08 at least material condition (i.e. minimum hole diameter). Therefore, the tolerance
of perpendicularity depends on the hole diameter: for a hole with a diameter of 0.39, the allowable tolerance
zone is a 0.04 diameter circle, whereas for a hole diameter of 0.43, the allowable tolerance is a 0.08 diameter
circle. A hole meeting this specification will fit a gage with a 0.35 diameter pin. These observations are given in
Table II and GD&T call-outs result in a 3D specification region as illustrated in Figure 2.
Note that some process points are not within the tolerance region. The center of the specification is
µT0 = [0, 41, 0, 0] and the measurement mean vector is X̄T = [0.411 63, 0.000 00, 0.010 00]. With respect to
Mardia’s multivarite skewness testing, the P -value is 0.01 and Mardia’s Kurtosis statistic is 0.03. It can thus
be concluded that this data set is not multivariate normally distributed at the 95% confidence level (IMSL-
subroutine18).
 Using the geometric distance approach, the new univariate variable is given by Gi =
(Diameteri − 0.41)2 + (xi − 0)2 + (yi − 0)2 , i = 1, 2, . . . , 20. Thus, the trivariate process data
became univariate process data. The goodness-of-fit technique is used to test the underlying distribution
of these geometric distance data. It can be seen that the lognormal distribution provides the best
fit among the 26 competing models, where the lognormal distribution (µ̂ = 0.0211, σ̂ = 0.0131) of
the K-S test statistic is 0.106 567 with P -value > 0.15 and the Anderson–Darling test statistic is

Copyright 
c 2005 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2006; 22:225–236
230 F. K. WANG

Tolerance region

0.44
0.43
0.42

Diameter
0.41
0.40
0.39
4
0.0 4
02 0.0
2
0.0
0.

00
0. 0.0
0
02 .02 xis
y

⫺0 xa
0.
ax

4 4

0
0. 0.0
is

⫺ ⫺

Figure 2. Tolerance region and process points of Example 1

Tolerance region

0.44
0.43
0.42
Diameter

0.41
0.40
0.39

04
0. 4
0.0
02

2
0.0
0.

00 .00
0.
02 0 xis
2

.
⫺0 xa
y

0
0.
ax

04 0.04

is

0 .
⫺ ⫺

Figure 3. Tolerance region and process points of Example 1 after process correction


0.399 507 with P -value ∧ ∧ ∧
√ > 0.15. Here, MRD = 0.02 2 + 0.04 2 + 0.04 2 = 0.06. Therefore, we have
Ĉpc = (0.06 − 0)/(3 π/2 × 0.021 063) = 0.758 and the 95% confidence interval for Cpc is [0.568, 1.136].
Example 2. From the above example, the capability index has shown that this process is not capable. Thus, some
corrections are made and the process performance should be improved. Again, 20 observations are given in
Table II and GD&T call-outs result in a 3D specification region as illustrated in Figure 3.
 Using the geometric distance approach, the new univariate variable is given by Gi =
(Diameteri − 0.41)2 + (xi − 0)2 + (yi − 0)2 , i = 1, 2, . . . , 20. Thus, the trivariate process data become

Copyright 
c 2005 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2006; 22:225–236
QUALITY EVALUATION OF A MANUFACTURED PRODUCT 231

Table II. The measurement data for Examples 1–2

Example 2
Example 1 (after process correction)
Parts # Diameter x y Diameter x y
1 0.411 02 0.000 46 0.020 58 0.406 07 −0.009 95 −0.013 55
2 0.410 77 0.029 74 0.008 07 0.408 63 −0.007 99 −0.000 60
3 0.414 72 0.002 41 0.009 78 0.409 29 0.002 29 0.001 82
4 0.414 23 −0.000 34 0.021 66 0.416 14 −0.006 38 −0.003 69
5 0.417 42 −0.004 79 0.012 60 0.406 12 −0.005 17 −0.004 29
6 0.411 78 0.016 42 −0.002 25 0.407 63 0.008 95 0.013 44
7 0.410 11 −0.037 19 0.007 35 0.391 29 −0.002 04 −0.001 99
8 0.414 88 −0.001 25 0.012 31 0.414 57 −0.005 11 −0.010 28
9 0.410 31 0.011 29 0.008 07 0.415 09 −0.004 51 −0.001 21
10 0.411 44 0.006 26 0.020 89 0.410 64 −0.002 89 0.004 33
11 0.412 80 0.001 30 0.028 11 0.415 74 −0.004 64 −0.007 05
12 0.414 35 −0.062 88 0.009 12 0.413 04 −0.005 06 0.006 06
13 0.414 76 −0.022 82 0.006 44 0.413 71 −0.009 92 0.006 37
14 0.405 05 0.011 80 0.012 67 0.403 35 −0.006 18 −0.005 94
15 0.410 99 0.003 28 −0.000 76 0.410 00 0.009 93 0.000 33
16 0.408 65 0.008 57 0.011 82 0.410 17 0.008 21 −0.005 72
17 0.412 58 0.017 50 0.007 72 0.405 91 0.008 89 −0.002 94
18 0.409 46 0.012 62 0.016 03 0.403 03 0.001 68 −0.004 72
19 0.403 68 0.004 29 −0.006 22 0.411 46 −0.000 07 −0.002 33
20 0.413 51 0.003 42 0.010 72 0.419 93 0.004 68 −0.003 50

univariate process data. The goodness-of-fit technique is used to test the underlying distribution of these
geometric distance data. It can be seen that the logistic distribution provides the best fit among the 26 competing
models, where the logistic distribution (α̂ = 0.01, β̂ = 0.002 29) of the K-S test statistic is 0.128 453 with
P -value > 0.15 and the
√ Anderson–Darling test statistic is 0.347 876 with P -value > 0.15. Therefore, we have
Ĉpc = (0.06 − 0)/(3 π/2 × 0.010 019) = 1.593 and the 95% confidence interval for Cpc is [1.298, 2.062],
and the probability of the product non-conforming is 3 × 10−9 .

2.2. A composite capability index


Based on the analysis procedure, suppose that a manufactured product with multiple quality characteristics can
have k geometric distance variables. Then, the estimated composite capability index can be simply determined
by the following equation:
 k 1/ k
M Ĉpc = Ĉpc(i) (5)
i=1

where
MRDi
Ĉpc(i) = √
3 π/2E[Gi ]
represents the estimated capability index of the ith geometric distance variable. An approximate 100 ×
(1 − α)% confidence interval on MCpc is obtained from
 k 1/ k k 1/ k
Ĉpc(i) Ĉpc(i)
√ ≤ MCpc ≤ √ (6)
i=1
1 + tα,n−1 Sci /(c̄i n) i=1
1 − tα,n−1 Sci /(c̄i n)

where c̄i and Sc2i represent the sample mean and standard variance of the ith geometric distance variable,
respectively, and t is the student t distribution.

Copyright 
c 2005 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2006; 22:225–236
232 F. K. WANG

INTEL P4
CPU

INTEL P4 Desktop
SKT Connect

Figure 4. The layout of a SKT connecter

Furthermore, the estimated proportion of a manufactured product with multiple characteristics non-
conforming is given by
k 
 MRDi
P (out of tolerance) = 1 − f (x) dx (7)
i=1 0

where f (x) is the density function of the best-fitting distribution for each geometric distance variable.

3. CASE STUDY
In this section, we present the capability analysis of a real-life data set called the connecter (Figure 4)
obtained from a manufacturer in the computer industry in Taiwan. This data set contains a sample of 100
parts (Table III) that were tested on seven quality characteristics of interest to the manufacturer. These seven
quality characteristics are X1 (contact gap X), X2 (contact loop Tp), X3 (LLCR), X4 (contact xTp), X5
(contact loop diameter), X6 (LTGAPY) and X7 (RTGAPY), respectively. The specification limits for these seven
quality characteristics can be two-sided or one-sided, and they are 0.10 ± 0.04 mm, 0 + 0.50 mm, 11 ± 5 m,
0 + 0.2 mm, 0.55 ± 0.06 mm, 0.07 ± 0.05 mm and 0.07 ± 0.05 mm, respectively. Based on the quality
characteristics and the manufacturing processes, we found that variables (X1 , X2 , X3 ) are correlated; variables
X4 and X5 are uncorrelated with other variables, and variables (X6 ,X7 ) are correlated. Using the geometric
 variables are given by G1 = (X1 − 0.1) + X2 + (X3 − 11) , G2 =
distance approach, four new univariate 2 2 2

X4 , G3 = (X5 − 0.55) and G4 = (X6 − 0.07) + (X7 − 0.07) , respectively, and the results are given in
2 2 2
Table III. The four maximum radial distance (MRDs) are MRD1 = 5.025, MRD2 = 0.2, MRD3 = 0.06 and
MRD4 = 0.0707, respectively. Also, from the correlation matrix of these four geometric distance variables,
we found that these variables do not have significant correlation with each other.
The goodness-of-fit technique is used to test the underlying distribution of all geometric
distance variables. For the geometric distance variable (G1 ), the gamma distribution provides
the best fit where the gamma distribution (α̂ = 1.68, β̂ = 0.59) of the K-S test statistic
is 0.1055 with P -value > 0.15 and the Anderson–Darling test statistic is 1.3053 with
P -value > 0.15. With respect to the geometric distance variable (G2 ), the Weibull distribution provides
the best fit, where the Weibull distribution (θ̂ = 1.18, β̂ = 0.0304) of the K-S test statistic is 0.0438 with
P -value > 0.10 and the Anderson–Darling test statistic is 0.1678 with P -value > 0.25. For the geometric
distance variable (G3 ), the normal distribution provides the best fit, where the normal distribution (µ̂ = 0.0177,

Copyright 
c 2005 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2006; 22:225–236
QUALITY EVALUATION OF A MANUFACTURED PRODUCT 233

Table III. The measurement data for the case study and the geometric distance variables

Original variables Geometric distance variables


X1 X2 X3 X4 X5 X6 X7 G1 G2 G3 G4
0.1165 0.0614 10.7824 0.0090 0.5536 0.0642 0.0585 0.2267 0.0090 0.0036 0.0129
0.1259 0.0277 10.8395 0.0091 0.5290 0.0994 0.0889 0.1649 0.0091 0.0210 0.0350
0.1265 0.0762 10.9538 0.0153 0.5444 0.0980 0.0974 0.0930 0.0153 0.0056 0.0392
0.1185 0.0957 10.5824 0.0216 0.5747 0.0856 0.0837 0.4289 0.0216 0.0247 0.0208
0.1414 0.1319 10.7395 0.0088 0.5621 0.0803 0.0870 0.2949 0.0088 0.0121 0.0199
0.0920 0.0476 10.7109 0.0226 0.5196 0.0804 0.0743 0.2931 0.0226 0.0304 0.0113
0.0804 0.0440 10.4824 0.0104 0.5175 0.0616 0.0534 0.5199 0.0104 0.0325 0.0186
0.1103 0.0901 10.7966 0.0064 0.5653 0.0573 0.0620 0.2227 0.0064 0.0153 0.0150
0.1022 0.0918 10.8681 0.0782 0.5245 0.0650 0.0523 0.1607 0.0782 0.0255 0.0184
0.1103 0.0823 10.7824 0.0310 0.5242 0.0992 0.0804 0.2329 0.0310 0.0258 0.0310
0.1069 0.0943 10.6681 0.0265 0.5321 0.0564 0.0718 0.3451 0.0265 0.0179 0.0137
0.1075 0.0950 10.6109 0.0564 0.5214 0.1035 0.0920 0.4006 0.0564 0.0286 0.0401
0.1170 0.1177 10.9538 0.0069 0.5451 0.0965 0.0867 0.1276 0.0069 0.0049 0.0313
0.1276 0.1378 10.7681 0.1162 0.5302 0.0997 0.0797 0.2712 0.1162 0.0198 0.0312
0.1039 0.0645 10.6109 0.0341 0.5326 0.0785 0.0785 0.3944 0.0341 0.0174 0.0120
0.1251 0.0988 10.6395 0.0671 0.5310 0.0945 0.0770 0.3746 0.0671 0.0190 0.0255
0.1153 0.1553 11.2966 0.0208 0.5622 0.0724 0.0626 0.3352 0.0208 0.0122 0.0078
0.0961 0.0142 10.9657 0.0019 0.5131 0.0810 0.0577 0.3373 0.0019 0.0369 0.0165
0.1170 0.0808 11.0657 0.0343 0.5281 0.0782 0.0785 0.1055 0.0343 0.0219 0.0118
0.1158 0.0780 10.2490 0.0631 0.5532 0.0854 0.0821 0.7552 0.0631 0.0032 0.0196
0.1134 0.1511 10.8324 0.0750 0.5460 0.0672 0.0621 0.2261 0.0750 0.0040 0.0084
0.1356 0.0407 11.0824 0.0138 0.5225 0.0377 0.0601 0.0985 0.0138 0.0275 0.0338
0.1222 0.0654 10.8157 0.0021 0.5472 0.1091 0.0921 0.1968 0.0021 0.0028 0.0449
0.1124 0.1516 7.3990 0.0406 0.5788 0.0473 0.0552 3.6042 0.0406 0.0288 0.0271
0.1155 0.0695 6.7824 0.0111 0.5322 0.0760 0.0639 4.2182 0.0111 0.0178 0.0086
0.1223 0.0853 7.2538 0.0180 0.5333 0.0967 0.0992 3.7473 0.0180 0.0167 0.0396
0.1130 0.0869 10.5681 0.0195 0.5268 0.0945 0.0932 0.4408 0.0195 0.0232 0.0337
0.1137 0.0652 6.5109 0.0073 0.4953 0.0463 0.1148 4.4896 0.0073 0.0547 0.0507
0.1112 0.1357 6.8824 0.0293 0.5301 0.0757 0.0919 4.1199 0.0293 0.0199 0.0226
0.1114 0.0909 10.7990 0.0034 0.5269 0.0879 0.0727 0.2209 0.0034 0.0231 0.0181
0.0978 0.0970 10.4324 0.0408 0.5198 0.0935 0.0832 0.5759 0.0408 0.0302 0.0270
0.1209 0.0939 10.6657 0.0211 0.5107 0.0971 0.1126 0.3479 0.0211 0.0393 0.0505
0.1092 0.0419 11.4324 0.0451 0.5260 0.0679 0.0651 0.4345 0.0451 0.0240 0.0053
0.1161 0.1088 11.1990 0.0032 0.5353 0.0856 0.0874 0.2274 0.0032 0.0147 0.0234
0.1131 0.0698 10.3824 0.0137 0.5331 0.0748 0.0812 0.6217 0.0137 0.0169 0.0122
0.1193 0.1445 11.9490 0.0745 0.5393 0.1064 0.1008 0.9602 0.0745 0.0107 0.0477
0.1233 0.0900 12.2681 0.0253 0.5294 0.0688 0.0850 1.2715 0.0253 0.0206 0.0150
0.0890 0.1102 12.3824 0.0674 0.5253 0.0763 0.0825 1.3868 0.0674 0.0247 0.0140
0.1074 0.0507 12.3681 0.0104 0.5245 0.0879 0.0822 1.3690 0.0104 0.0255 0.0217
0.0947 0.0845 12.4538 0.0024 0.5212 0.0733 0.0307 1.4562 0.0024 0.0288 0.0394
0.1048 0.0661 12.2824 0.0003 0.5305 0.1004 0.0980 1.2841 0.0003 0.0195 0.0413
0.1192 0.0448 11.8395 0.0476 0.5334 0.1066 0.1053 0.8409 0.0476 0.0166 0.0508
0.1143 0.1009 12.2252 0.0215 0.5203 0.0968 0.1012 1.2294 0.0215 0.0297 0.0411
0.1102 0.0815 12.5109 0.0309 0.5224 0.1038 0.1113 1.5132 0.0309 0.0276 0.0534
0.1084 0.1617 12.0681 0.0883 0.5304 0.0855 0.0891 1.0803 0.0883 0.0196 0.0246
0.1071 0.0423 12.0252 0.0129 0.5262 0.0861 0.0837 1.0261 0.0129 0.0238 0.0211
0.1106 0.0585 11.8252 0.0382 0.5352 0.0860 0.1060 0.8273 0.0382 0.0148 0.0394
0.1051 0.1151 12.3824 0.0349 0.5089 0.0963 0.1085 1.3871 0.0349 0.0411 0.0466
0.1167 0.0827 12.3395 0.0166 0.5439 0.0938 0.0897 1.3421 0.0166 0.0061 0.0309
0.1062 0.0566 12.2538 0.0300 0.5130 0.0875 0.0923 1.2551 0.0300 0.0370 0.0283
0.1091 0.0221 12.2824 0.0077 0.5208 0.0923 0.0965 1.2826 0.0077 0.0292 0.0346
0.1138 0.1190 12.4395 0.0093 0.5402 0.0900 0.0939 1.4445 0.0093 0.0098 0.0312
0.1309 0.0699 12.6324 0.0219 0.5271 0.0839 0.0908 1.6341 0.0219 0.0229 0.0250

Copyright 
c 2005 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2006; 22:225–236
234 F. K. WANG

Table III. Continued


Original variables Geometric distance variables
X1 X2 X3 X4 X5 X6 X7 G1 G2 G3 G4
0.1075 0.0417 11.4824 0.0081 0.5178 0.0683 0.0743 0.4842 0.0081 0.0322 0.0046
0.1157 0.0377 11.9157 0.0056 0.5173 0.0718 0.0681 0.9166 0.0056 0.0327 0.0026
0.1193 0.0752 12.4157 0.0351 0.5527 0.0714 0.0655 1.4178 0.0351 0.0027 0.0047
0.1239 0.0849 11.8324 0.0553 0.5278 0.0553 0.0410 0.8370 0.0553 0.0222 0.0325
0.1133 0.0418 11.8157 0.0048 0.5204 0.0858 0.0981 0.8169 0.0048 0.0296 0.0322
0.1206 0.0818 9.9824 0.0287 0.5361 0.0868 0.0964 1.0211 0.0287 0.0139 0.0313
0.1220 0.0733 10.3538 0.0024 0.5361 0.0947 0.0882 0.6507 0.0024 0.0139 0.0307
0.1161 0.1170 10.1109 0.0622 0.5565 0.0663 0.0617 0.8969 0.0622 0.0065 0.0091
0.1186 0.1044 11.7824 0.0495 0.5499 0.0956 0.0846 0.7895 0.0495 0.0001 0.0295
0.1020 0.0529 9.2966 0.0217 0.5247 0.0948 0.1156 1.7042 0.0217 0.0253 0.0519
0.1116 0.0563 9.4657 0.0321 0.5281 0.0850 0.0863 1.5354 0.0321 0.0219 0.0222
0.1111 0.0352 11.8824 0.0268 0.5166 0.1040 0.1013 0.8831 0.0268 0.0334 0.0462
0.0986 0.0437 12.1824 0.0160 0.5417 0.0926 0.1137 1.1832 0.0160 0.0083 0.0492
0.1166 0.0651 11.9157 0.0259 0.5192 0.0767 0.0891 0.9181 0.0259 0.0308 0.0202
0.1187 0.0535 12.3490 0.0281 0.5433 0.0890 0.0985 1.3502 0.0281 0.0067 0.0343
0.1206 0.1073 12.1157 0.0551 0.5514 0.0925 0.1016 1.1210 0.0551 0.0014 0.0388
0.1139 0.0956 11.5490 0.0173 0.5441 0.1027 0.1062 0.5575 0.0173 0.0059 0.0488
0.1103 0.0332 12.0324 0.0055 0.5189 0.0687 0.0445 1.0329 0.0055 0.0311 0.0255
0.1089 0.0843 12.1966 0.0097 0.5395 0.0925 0.0750 1.1996 0.0097 0.0105 0.0230
0.1043 0.0558 12.0252 0.0372 0.5469 0.0711 0.0882 1.0267 0.0372 0.0031 0.0182
0.1056 0.0651 11.9395 0.0205 0.5438 0.0873 0.0840 0.9418 0.0205 0.0062 0.0223
0.1145 0.1210 11.9538 0.0258 0.5823 0.0910 0.0907 0.9615 0.0258 0.0323 0.0295
0.1277 0.0945 11.9252 0.0383 0.5384 0.0949 0.0944 0.9304 0.0383 0.0116 0.0349
0.1120 0.1285 12.0538 0.0172 0.5410 0.0721 0.0794 1.0617 0.0172 0.0090 0.0096
0.1134 0.0854 12.0109 0.0117 0.5636 0.1001 0.0927 1.0146 0.0117 0.0136 0.0377
0.1237 0.0752 12.3681 0.0021 0.5483 0.0661 0.0669 1.3703 0.0021 0.0017 0.0050
0.1270 0.0933 12.2109 0.0055 0.5493 0.0780 0.0917 1.2148 0.0055 0.0007 0.0231
0.1154 0.1247 12.2395 0.0728 0.5581 0.0733 0.0786 1.2458 0.0728 0.0081 0.0092
0.0967 0.0685 11.3252 0.0149 0.5417 0.0891 0.0899 0.3324 0.0149 0.0083 0.0276
0.1056 0.0575 11.6681 0.0003 0.5653 0.0961 0.0852 0.6706 0.0003 0.0153 0.0302
0.1116 0.1168 11.3395 0.0116 0.5664 0.0900 0.0886 0.3592 0.0116 0.0164 0.0273
0.1113 0.0839 11.7538 0.0102 0.5581 0.0824 0.0891 0.7585 0.0102 0.0081 0.0228
0.1149 0.1019 11.9966 0.0592 0.5486 0.0803 0.0861 1.0019 0.0592 0.0014 0.0191
0.1091 0.0170 11.2252 0.0056 0.5341 0.1045 0.1097 0.2260 0.0056 0.0159 0.0526
0.1120 0.0734 11.7157 0.0228 0.5401 0.0826 0.0799 0.7195 0.0228 0.0099 0.0160
0.1074 0.0685 11.7324 0.0110 0.5300 0.0781 0.1002 0.7356 0.0110 0.0200 0.0313
0.1112 0.0807 11.8324 0.0403 0.5239 0.0982 0.1118 0.8363 0.0403 0.0261 0.0504
0.1097 0.0993 12.6657 0.0444 0.5407 0.0782 0.0798 1.6687 0.0444 0.0093 0.0128
0.1085 0.0882 12.2490 0.0429 0.5473 0.1172 0.1031 1.2522 0.0429 0.0027 0.0576
0.1009 0.1252 12.1490 0.0778 0.5281 0.0721 0.0942 1.1558 0.0778 0.0219 0.0243
0.1071 0.0686 9.3157 0.0069 0.5632 0.1030 0.0958 1.6857 0.0069 0.0132 0.0419
0.1157 0.0518 10.2824 0.0247 0.5346 0.0865 0.0746 0.7197 0.0247 0.0154 0.0171
0.1127 0.1203 10.1824 0.0431 0.5545 0.1016 0.1140 0.8265 0.0431 0.0045 0.0542
0.1124 0.1029 11.7824 0.0574 0.5394 0.0671 0.1070 0.7892 0.0574 0.0106 0.0371
0.1017 0.0753 9.3681 0.0516 0.5634 0.0873 0.1069 1.6337 0.0516 0.0134 0.0408
0.1210 0.0299 9.3824 0.0255 0.5554 0.0647 0.0999 1.6181 0.0255 0.0054 0.0304
0.1107 0.1671 12.5490 0.0845 0.5398 0.0656 0.0669 1.5580 0.0845 0.0102 0.0054

σ̂ = 0.0109) of the K-S test statistic is 0.0703 with P -value > 0.15 and the Anderson–Darling test statistic
is 0.6208 with P -value > 0.10. With respect to the geometric distance variable (G4 ), the Weibull distribution
provides the best fit, where the Weibull distribution (θ̂ = 2.12, β̂ = 0.0312) of the K-S test statistic is 0.0567
with P -value > 0.10 and the Anderson–Darling test statistic is 0.4194 with P -value > 0.25. The capability
indices of all geometric distance variables with the respective probabilities of the product non-conforming

Copyright 
c 2005 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2006; 22:225–236
QUALITY EVALUATION OF A MANUFACTURED PRODUCT 235

Table IV. Summary of the case study

95% confidence interval


Geometric distance Probability of the
variable Ĉpc Lower limit Upper limit product non-conforming
G1 1.342 1.126 1.662 0.001 021 6
G2 1.849 1.556 2.278 0.000 097 5
G3 0.902 0.791 1.049 0.000 052 0
G4 0.703 0.629 0.796 0.003 460 6
 pc = 1.120
MC 0.966 1.333 0.004 627 7

are summarized in Table IV. Such results provide a reasonable summary of this product with seven quality
characteristics for quality improvement. We found that the overall MCpc is greater than 1.00 and some Cpc
values are less than 1.00. The results show that this manufactured product does not satisfy the requirements
so, quality improvement must be executed. Since G3 and G4 are significantly less than 1.00, the quality
characteristics X5 , X6 and X7 are candidates for improvement first.

4. CONCLUSION
Although the univariate capability index of individual quality characteristics has been widely used in
the industry, an index of multiple quality characteristics is often needed. Such an index should quantify
the manufacturer’s ability to make a product that simultaneously satisfies all customer requirements.
A manufactured product may have multiple quality characteristics, some of which may be correlated and the rest
of which may be uncorrelated. Also, the specification limits of the quality characteristics can be one-sided or
two-sided. In order to deal with such multivariate data, the analysis of the procedure is presented. The geometric
distance variable is used to combine the correlated or uncorrelated quality characteristics. Then, the Luceño
capability index is used to summarize the performance for each geometric distance variable. Finally, a composite
capability index, MCpc , composed of several univariate capability indices, is proposed to analyze the capability
of a manufactured product with multiple characteristics. In addition, the theoretical distribution of the geometric
distance variable under any type of multivariate data can be complicated. Here, the goodness-of-fit technique can
be used to find the best-fitting distribution. Then, a computable probability of the product non-conforming for
each geometric distance variable can be obtained. Therefore, the probability of the product non-conforming
can also be obtained. Quality engineers can use this technique to solve the difficulties encountered for a
manufactured product with multiple characteristics. It would be interesting to see whether other methods of
estimating the capability index from more than three correlated variables, and thus the composite capability
index, can improve.

Acknowledgements
The author wishes to gratefully acknowledge the referee of this paper who helped to clarify and improve the
presentation. Also, the author is grateful for partial support from the National Science Council of Taiwan under
the grant NSC-92-2416-H-027-005.

REFERENCES
1. Kane VE. Process capability indices. Journal of Quality Technology 1986; 18:41–52.
2. Kotz S, Johnson NL. Process Capability Indices. Chapman and Hall: London, 1993.
3. Kotz S, Johnson NL. Process capability indices—a review, 1992–2000. Journal of Quality Technology 2002; 34:1–19.

Copyright 
c 2005 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2006; 22:225–236
236 F. K. WANG

4. Chan LK, Cheng SW, Spiring FA. A multivariate measure of process capability. Journal of Modeling and Simulation
1991; 11:1–6.
5. Pearn WL, Kotz S, Johnson NL. Distributional and inferential properties of process capability indices. Journal of
Quality Technology 1992; 24:216–231.
6. Wierda SJ. A multivariate process capability index. ASQC Quality Congress Transactions, Boston, MA, 1993.
American Society for Quality Control: Milwaukee, WI, 1993; 342–348.
7. Team W, Subbaiah P, Liddy JW. A note on multivariate capability indices. Journal of Applied Statistics 1993; 20:339–
351.
8. Chen H. A multivariate process capability index over a rectangular solid tolerance zone. Statistica Sinica 1994; 4:749–
758.
9. Shahriari H, Hubele NF, Lawrence FP. A multivariate process capability vector. Proceedings of the 4th Industrial
Engineering Research Conference. Institute of Industrial Engineers: Nashville, TN, 1995; 304–309.
10. Veevers A. Viability and capability indices for multi-response process. Journal of Applied Statistics 1998; 25:545–558.
11. Wang FK, Hubele NF. Quality evaluation using geometric distance approach. International Journal of Reliability,
Quality and Safety Engineering 1999; 6:139–153.
12. Wang FK, Du TC. Using principal component analysis in process performance for multivariate data. Omega 2000;
28:185–194.
13. Yeh AB, Chen H. A nonparametric multivariate process capability index. International Journal of Modeling and
Simulation 2001; 21:218–223.
14. Wang FK, Hubele NF. Quality evaluation of geometric tolerance regions in form and location. Quality Engineering
2001–2002; 14:203–209.
15. Luceño A. A process capability index with reliable confidence intervals. Communication in Statistics: Simulation and
Computation 1996; 25:235–245.
16. Rodriguiz RN. Recent developments in process capability analysis. Journal of Quality Technology 1992; 24:176–186.
17. BestFit Software 2.0. Palisade Corporation: New York, 1998.
18. Microsoft Fortran Power Station. IMSL Stat Library, Version 4.0, 1995.

Author’s biography
Fu-Kwun Wang received his PhD in Industrial Engineering from Arizona State University, Tempe, Arizona.
Currently, he is an Associate Professor at the Department of Industrial Management, National Taiwan
University of Science and Technology, Taiwan. His teaching, publications and research interests are reliability,
statistical quality control and applied statistics. He has published articles in Journal of Quality Technology,
Quality Engineering, Quality & Reliability Engineering International, Total Quality Management & Business
Excellence, Microelectronics & Reliability, Reliability Engineering and System Safety, International Journal
of Reliability, Quality and Safety Engineering, Omega, International Journal of Industrial Engineering, IEEE
Transactions on Semiconductor Manufacturing, International Journal of Production Research, IJCIM, Robotics
& CIM, and Production Planning & Control.

Copyright 
c 2005 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2006; 22:225–236

You might also like