Beruflich Dokumente
Kultur Dokumente
1. Introduction
A standard practice in statistical process control is to evaluate process performance
via process capability analysis after the process is announced under statistical
control. Process capability analysis can be conducted both in univariate and
multivariate environments. In univariate environment, a single quality characteristic
of a product is considered and process performance is evaluated with respect to
its corresponding specication limit(s). Many researchers including (Kane, 1986;
Marcucci Beazley, 1988; Chan et al., 1988; Choi and Owen, 1990; Spiring, 1991;
Koons, 1992; Wheeler and Chambers, 1992; Pearn et al., 1992; Bissel, 1994; Pearn
and Chen, 2002; Noorossana, 2002), and Chen and Chen (2004) contributed to the
Received July 15, 2008; Accepted September 13, 2010
Address correspondence to Mohammad R. Niavarani, Department of Mechanical and
Manufacturing Engineering, The University of Melbourne, Melbourne, Australia; E-mail:
m_r_niavarani@yahoo.com
341
342
Niavarani et al.
343
well as from engineering specications that can lead to very different denitions of
capability in the multivariate domain. The next section provides discussion on two
new multivariate indices.
1/p
(1)
Figure 1 illustrates the method for the case when the engineering specications
dene a rectangular tolerance region and bivariate normal process measurements
dene an elliptical probability contour referred to by Shahriari et al. (1995) as
the process region. Their method forms a modied process region by drawing
the smallest rectangle around the ellipse. They dene the edges of the rectangle as
the lower and upper process limits referred to as LPLi and UPLi, respectively. These
limits are dened as:
1
2
p det
i
(2)
UPLi = i +
1
det
1
2
p det
i
LPLi = i
(3)
1
det
2
is the upper 100% of the 2
where is the probability of Type I error, p
distribution with p degrees of freedom associated with the probability contour,
1
1
by deleting the ith
det1
i is the determinant of i , a matrix obtained from
row and column, and i = 1 2 3 p. In practice, estimates for large samples may
be used in place of and . The concept in the article published by Shahriari et al.
(1995) is to construct a modied process region with the same general geometric
shape as the engineering tolerance region; however, as shown in Fig. 1, the size
344
Niavarani et al.
and relative dimensions are determined by the probability contour. This concept is
illustrated by the following relationship:
1/p
(4)
(5)
pn 1
F
PV = P T
n
pnp
(6)
where is the probability of Type I error and Fpnp is the F distribution with
p and n p degrees of freedom in the numerator and denominator, respectively.
Given that PV never exceeds 1; values close to zero indicate that the center of
process is far from the engineering target value.
The third component of the vector compares the location of the modied
process region to the tolerance region. It indicates whether any part of the modied
process region falls outside the engineering specications. This component has a
value of 1 if the tolerance region contains the entire modied process region;
0 Otherwise
345
(7)
Based on the probability level chosen for the contour, this component can reect the
amount of nonconforming parts. In summary, the method proposed by Shahriari
et al. (1995) consists of a three-component vector dened as [CP M PV , and LI]
where CP M, PV , and LI represent a comparison of volumes of regions, locations of
centers, and locations of regions, respectively.
As shown in Fig. 2, in calculation of CP M, it is clear that if the modied process
region in rectangular form (two-dimensioned case) is compared to the tolerance
region, the area of the elliptical contour (process region) is overestimated by the
modied process region (shaded region) considered in the denominator of Eq. (4).
The same concept is applicable to higher dimensions as well. In this case, the
denominator of Eq. (4) is overestimated and consequently CP M value would be
underestimated. This estimation may cause an incorrect decision in the judgment
regarding the process performance especially when CP M value is close to 1. The
error can be considered as e1 calculated as:
e1 =
VR VE
VR
(8)
where VR and VE are denoted for volume of the modied process region in
rectangular shape and volume of the process region in ellipsoid shape, respectively.
The above viewpoint could be generalized to the ellipsoid case where three or
higher dimensions are involved.
According to the above discussion and also considering the emphasis made by
Wang et al. (2000) pointing out that the concept is to construct a modied process
region with the same general geometric shape as the engineering tolerance region,
we propose the use of a new CpM referred to hereafter as NCpM. This index divides
the tolerance region by the real area or volume of the process region instead of the
346
Niavarani et al.
modied region. This new index dened as following helps to eliminate e1 leading
to the overestimation of CpM:
1/p
pi=1 USLi LSLi
(9)
NCP M =
T p/2 pi=1 i1/2
where i are the characteristic roots (sometimes refereed to as latent roots or
eigenvalues) of the ith variant.
If the purpose is to compare two regions having the same general geometric
shape, then in order to reduce the error of overestimation, the engineering tolerance
region can be compared to a rectangle with the closest area to the process region
area. As is shown in Fig. 3, the sides of the rectangle should be parallel to ellipsoid
diameters. The modied CpM or MCpM can be calculated as
1/p
pi=1 USLi LSLi
(10)
MCP M =
2p T p/2 pi=1 i1/2
In this proposed method, the error, e2 , can be calculated as:
e2 =
VR VE
VR
(11)
where VR is denoted volume of the closest area to the process region area
(the rectangle with parallel sides to ellipsoid diameters) and VE are denoted for
volume of the process region in ellipsoid shape.
Comparison of Figs. 2 and 3 reveals that the modied process region dened
in the denominator of MCpM is less than the modied process region used in
the denominator of CpM which leads to a closer estimation of the true process
capability. In the other word, e2 is always less that e1 . As the correlation between
the variables increases, the difference between these errors also increases. For the
case of uncorrelated variables with equal variance, the process region will become a
circle and as a result these two indices as well as errors would be equal. The second
and third components of the vector, i.e., PV and LI dened by Eqs. (6) and (7) are
347
essentially the same as before. The above viewpoint could be generalized to higher
dimensions as well.
VolR1
VolR2
(12)
CP
D
(13)
where
CP =
=
(14)
and
D = 1+
21
n
X T0 S 1
X T0
n1
(15)
348
Niavarani et al.
Lim e3 = Lim 1 p = 1
p+
p+
2
(16)
(17)
(18)
(19)
Equation (19) indicates that as the number of quality characteristics increases, the
estimation error increases and in limit approaches to one. This means that VE and
349
consequently MCPM tend to zero. This estimation can cause an incorrect decision in
judging the real capability of the process.
To eliminate the error in the estimation of the tolerance region, we propose a
new MCPM referred to as NMCPM dened as:
NMCPM =
NCP
D
(20)
where
NCP =
(21)
(22)
32909 838
838
62
350
Niavarani et al.
Table 1
Simulation results for two-dimension case and n = 50
Symmetry
of tolerance
Symmetric
tolerance
Correlation
coefcient
= 02
= 05
= 09
Asymmetric
tolerance
= 02
= 05
= 09
Indices
CP M = 07702
NCP M = 08831
MCP M = 07833
CP = 04845 MCPM = 04348
NCP = 07854, NMCPM = 10008
CP M = 07695
NCP M = 09466
MCP M = 08389
CP = 05547, MCPM = 04979
NCP = 08993, NMCPM = 07876
CP M = 07720
NCP M = 133
MCP M = 117
CP = 10972, MCPM = 09847
NCP = 17984, NMCPM = 13434
CP M = 07737
NCP M = 08887
MCP M = 07846
CP = 04909, MCPM = 04406
NCP = 07958, NMCPM = 07062
CP M = 07707
NCP M = 09421
MCP M = 08349
CP = 05503, MCPM = 04939
NCP = 08921, NMCPM = 06467
CP M = 07709
NCP M = 12993
MCP M = 115615
CP = 10469, MCPM = 11515
NCP = 16972, NMCPM = 12100
Errors
e1 = 02404
e2 = 02146
e3 = 02146
e1 = 03340
e2 = 02146
e3 = 02146
e1 = 06640
e2 = 02146
e3 = 02146
e1 = 02419
e2 = 02146
e3 = 02146
e1 = 03280
e2 = 02146
e3 = 02146
e1 = 06458
e2 = 02146
e3 = 02146
641216
256486
160304
256486
641216
160304
160304
160304
100190
In these tables, more attention has been drawn to those parts of indices that are
revised in this article as a base for calculating estimation errors. So, PV and LI have
been relinquished as the similar components in proposed and previous indices.
351
Table 2
Simulation results for two-dimension case and n = 100
Symmetry
of tolerance
Symmetric
tolerance
Correlation
coefcient
= 02
= 05
= 09
Asymmetric
tolerance
= 02
= 05
= 09
Indices
CP M = 07645
NCP M = 08744
MCP M = 07749
CP = 04790, MCPM = 04469
NCP = 07666, NMCPM = 07140
CP M = 07756
NCP M = 09370
MCP M = 08304
CP = 05428, MCPM = 05130
NCP = 08800, NMCPM = 08479
CP M = 07660
NCP M = 13207
MCP M = 11704
CP = 10758, MCPM = 10193
NCP = 17494, NMCPM = 17842
CP M = 07706
NCP M = 08835
MCP M = 07829
CP = 04825, MCPM = 04560
NCP = 07822, NMCPM = 07223
CP M = 07679
NCP M = 09345
MCP M = 08282
CP = 05400, MCPM = 05104
NCP = 08755, NMCPM = 08092
CP M = 07649
NCP M = 12873
MCP M = 11408
CP = 10246, MCPM = 19684
NCP = 16610, NMCPM = 14984
Errors
e1 = 02355
e2 = 02146
e3 = 02146
e1 = 03318
e2 = 02146
e3 = 02146
e1 = 06629
e2 = 02146
e3 = 02146
e1 = 02395
e2 = 02146
e3 = 02146
e1 = 03277
e2 = 02146
e3 = 02146
e1 = 06463
e2 = 02146
e3 = 02146
For better understanding, all indices and parameters included in tables are
briey explained as follows:
CP M calculated based on original engineering tolerance region and modied
process region
NC P M calculated based on original engineering tolerance region and original
process region
MC P M calculated based on original engineering tolerance region and modied
process region (a rectangle with parallel sides to ellipsoid diameters)
352
Niavarani et al.
Table 3
Simulation results for two-dimension case and n = 1000
Symmetry
of tolerance
Symmetric
tolerance
Correlation
coefcient
= 02
= 05
= 09
Asymmetric
tolerance
= 02
= 05
= 09
Indices
CP M = 07600
NCP M = 08665
MCP M = 07679
CP = 14632, MCPM = 14605
NCP = 07510, NMCPM = 07438
CP M = 07601
NCP M = 09297
MCP M = 08240
CP = 05334, MCPM = 05302
NCP = 08646,
NMCPM = 08479
CP M = 07609
NCP M = 13112
MCP M = 11620
CP = 10607, MCPM = 10545
NCP = 17195, NMCPM = 17338
CP M = 07582
NCP M = 08674
MCP M = 07688
CP = 04643, MCPM = 04616
NCP = 07527, NMCPM = 07204
CP M = 07603
NCP M = 09262
MCP M = 08208
CP = 05293, MCPM = 05262
NCP = 08580, NMCPM = 08978
CP M = 07614
NCP M = 12766
MCP M = 11313
CP = 10055, MCPM = 09996
NCP = 16300, NMCPM = 14685
Errors
e1 = 02307
e2 = 02146
e3 = 02146
e1 = 03315
e2 = 02146
e3 = 02146
e1 = 06632
e2 = 02146
e3 = 02146
e1 = 02359
e2 = 02146
e3 = 02146
e1 = 03259
e2 = 02146
e3 = 02146
e1 = 06440
e2 = 02146
e3 = 02146
353
Table 4
Simulation results for three-dimension case and n = 50
Symmetry of tolerance
Symmetric tolerance
Asymmetric tolerance
Indices
CP M = 05728
NCP M = 10816
MCP M = 07920
CP = 01971, MCPM = 01769
NCP = 12781, NMCPM = 10599
CP M = 05665
NCP M = 10763
MCP M = 07881
CP = 01946, MCPM = 01746
NCP = 12616, NMCPM = 09843
Errors
e1 = 08506
e2 = 06073
e3 = 06073
e1 = 08521
e2 = 06073
e3 = 06073
Table 5
Simulation results for three-dimension case and n = 100
Symmetry of tolerance
Symmetric tolerance
Asymmetric tolerance
Indices
CP M = 05601
NCP M = 10678
MCP M = 07819
CP = 01885, MCPM = 01782
NCP = 12225, NMCPM = 11735
CP M = 05649
NCP M = 10692
MCP M = 07830
CP = 01895, MCPM = 01791
NCP = 12290, NMCPM = 10470
Errors
e1 = 08548
e2 = 06073
e3 = 06073
e1 = 08520
e2 = 06073
e3 = 06073
Table 6
Simulation results for three-dimension case and n = 1000
Symmetry of tolerance
Symmetric tolerance
Asymmetric tolerance
Indices
CP M = 05604
NCP M = 10513
MCP M = 07698
CP = 01793, MCPM = 01782
NCP = 116245, NMCPM = 11171
CP M = 05592
NCP M = 10502
MCP M = 07891
CP = 01787, MCPM = 01777
NCP = 11589, NMCPM = 11888
Errors
e1 = 08485
e2 = 06073
e3 = 06073
e1 = 08490
e2 = 06073
e3 = 06073
354
Niavarani et al.
Proposed indices are shown with bold format in the above denition as well as in
the tables below.
As it is indicated in Tables 13, in two-dimensioned case, the indices values
are quite sensitive to correlation coefcient. With the same problem parameters,
the higher correlation between quality characteristics, the more process capability
indices. Of CP M NCP M, and MCP M, the indices values do not rise in equal steps
and the rst one is not as sensitive as two others to correlation coefcient. Compare
to our two rst proposed indices, NCP M, and MCP M, the indices proposed by
Shahriari et al. (1995), CP M almost hovers around a constant value That is why
e1 rises exponentially by increasing (e.g., tripled from = 02 to = 09). In the
same circumstances from the correlation coefcient and sample size viewpoint, this
error is even worse in asymmetric tolerance than symmetric one. On the other hand,
the value of our two proposed indices naturally surge by increasing . Needless to
say, there is no error in applying NCPM in which two original areas are considered
in numerator and denominator of Eq. (9). Even if we want to use two similar shapes
in numerator and denominator as estimation, it is strongly recommend utilizing
MCP M. Because its value is closer to NCP M than CP M (provides a better estimate
of the process capability) and the error, e2 , remains constant in different correlation
coefcient. Similar to above sensitivity analysis, calculating CP and MCPM , as an
estimation instead of NCP and NMCPM , as the real value leads estimation error, e3 .
Remaining constant of e3 is promising in comparison with changeable value of e1 .
It is evident that e2 and e3 are the same in both symmetric and asymmetric cases.
The aforementioned ndings about the errors are also applicable in the threedimensioned case. In other words, unlike e2 and e3 those are not sensitive to
symmetry of tolerance, e1 increases in asymmetric tolerance.
From sample size point of view, as it is clear in tables, the more sample size, the
less estimation error.
In terms of dimension of problem, regarding the concept of previous and
proposed process capability indices as well as Eqs. (8), (11), (18), and (19), it
is obvious that increasing dimensions (no. of characteristics) fuels a surge of
exacerbating the errors. As it is obvious in Tables 46, by increasing the number of
characteristic from 2 to 3, the errors have been considerably enhanced.
6. Conclusion
In this article, we demonstrate that applying the modied areas or volumes instead
of the original ones leads an estimation error and an underestimation of the process
capability value. Through a simulation and sensitivity analysis, it was shown that
inappropriate decision regarding process performance is inevitable once the process
capability is underestimated. We also show that this issue is aggravated when
number of characteristics and/or correlation coefcient value increase.
According to the emphasis made by Wang et al. (2000) for CP M pointing out
that the concept is to construct a modied process region with the same general
geometric shape as the engineering tolerance region and Taam et al. (1993) for CP
and MCPM pointing out that the modifying tolerance region to possess the same
general geometric shape instead of process region, we propose NCP M instead of
CP M and also NCP and NMCPM rather than CP and MCPM in order to get rid of
estimation errors e1 and e3 . We also prove that, even for the case in which applying
the same shapes is a priority for user, utilizing MCP M instead of CP M fuels a decline
of estimation error from e1 to e2 .
355
Acknowledgment
The authors gratefully acknowledge the helpful comments of the anonymous
reviewers, which improved the content of the article.
References
Abu, Z. Md., Sultana, A. (2008). Assessment and comparison of multivariate process
capability indices in ceramic industry. J. Mech. Eng. 39(1):1825.
Bissel, D. (1994). Statistical Methods for SPC and TQM. New York: Chapman & Hall, p. 240.
Chan, L. K., Cheng, S. W., Spiring, F. A. (1991). A multivariate measure of process
capability. Int. J. Model. Simul. 11:1.
Chan, L. K., Cheng, S. W., Spiring, F. A. (1988). The robustness of the process capability
index to departures from normality. Statist. Theor. Data anal, Proc. Second Pacic Area
Statist. Conf. Tokyo: pp. 223239.
Chan, L. K., Cheng, S. W., Spiring, F. A. (1991). A multivariate measure of process
capability. Int. J. Model. Simul. 11:16.
Chen, J. P., Chen, K. S. (2004). Comparison of two process capabilities by using
indices CPM: an application to a color STN display. Int. J. Qual. Reliab. Manage.
21(1):90101.
Choi, B. C., Owen, D. B. (1990). A study of a new process capability index. Commun. Statist.
Theor. Meth. 19(4):12311245.
Chou, C. Y., Chen, C. H., Liu, H. R. (2004). Effect of non-normality on the economic design
of warning limit X-bar charts. Qual. Eng. 16(4):56775.
Clements, J. A. (1989). Process capability calculations for non-normal distributions. Quality
Progress 22:95100.
Hubble, N. F., Shahriari, H., Cheng, G. S. (1991). A Bivariate Process Capability Vector in
Statistics and Design in Process Control. New York: Marcel Dekker, pp. 299310.
Chen, J.-P., Ding, C. G. (2001). A new process capability index for non-normal distributions.
Int. J. Qual. Reliab. Manage. 18:762770.
Kane, V. E. (1986). Process capability indices. J. Qual. Technol. 18(1):4152.
Koons, G. F. (1992). Indices of Capability: Classical and Six Sigma Tools. Reading, MA:
Addison Wesley Publishing Co.
Kotz, S., Lovelace, C. R. (1998). Process Capability Indices in Theory and Practice. London:
Arnold.
Kotz, S., Johnson, N. L. (2002). Process capability indices a review 19922000. J. Qual.
Technol. 34:29.
Marcucci, M. O., Beazley, C. C. (1988). Capability indices: process performance measure.
42nd ASQC Ann. Qual. Congr. Trans. Dallas, TX: pp. 516523.
Munechika, M. (1986). Evaluation of process capability for skew distribution. Proc. 30th
EOQC Conf. Stockholm, pp. 383390.
Nickerson, D. M. (1994). Construction of a conservative condence region from projections
of an exact condence region in multiple linear regression. Amer. Statistician
48:120124.
Niverthi, M., Dey, D. K. (1995). Multivariate process capability: a bayesian perspective.
Technical Report, Department of Statistics, University of Connecticut, Storrs, CT.
Noorossana, R. (2002). Process capability analysis in the presence of autocorrelation. J. Qual.
Reliab. Eng. Int. 18:7577.
Pearn, W. L., Kotz, S., Johnson, N. L. (1992). Distributional and inferential properties of
PCI . J. Qual. Technol. 24(4):216231.
Pearn, W. L., Ko, C. H., Wang, K. H. (2002). A multi-process performance analysis chart
based on the incapability index Cpp: An application to the chip resistors. Microelectr.
Reliab. 42:11211125.
356
Niavarani et al.
Pearn, W. L., Chen, K. S., (2002). One-sided capability indices CPU and CPL: decision
making with sample information. Int. J. Qual. Reliab. Manage. 19(3):221245.
Shahriari, H., Huebele, N. F., Lawrence, F. P. (1995). A multivariate process capability
vector. Proc. 4th Indust. Eng. Res. Conf. Institute of Industrial Engineers, pp. 304309.
Somerville, S., Montgomery, D. C. (1996). Process capability indices and non-normal
distributions. Qual. Eng. 9(2):305316.
Spiring, F. A. (1991). Comments on J. Mirabellas letter. Qual. Progr. p. 24.
Spiring, F., Leung, B., Cheng, S. W., Yeung, A. (2003). A bibliography of process capability
papers. Quality and Reliability. Eng. Int. 19:445460.
Taam, W., Subbaiah, P., Lidd, J. W. (1993). A note on multivariate capability indices.
J. Appl. Statist. 20:339351.
Wang, F. K., Du, T. C. T. (2000). Using principal component analysis in process
performance for multivariate data. Omega. Int. J. Manage. Sci. 28:185194.
Wang, F. K., Hubele, N. F. (2001). Quality evaluation of geometric tolerance regions in
form and location. Qual. Eng. 14:203209.
Wang, F. K., Hubele, N. F., Lawrence, P., Miskulin, J. D., Shahriari, H. (2000). Comparison
of three multivariate process capability indices. J. Qual. Technol. 32(3):263275.
Wheeler, D. J., Chambers, D. S. (1992). Understanding Statistical Process Control. 2nd ed.
Knoxville, TN: SPC Press.
Wright, P. A. (1995). A process capability index sensitive to skewness. J. Statist. Comput.
Simul. 52:195203.
Copyright of Communications in Statistics: Theory & Methods is the property of Taylor & Francis Ltd and its
content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's
express written permission. However, users may print, download, or email articles for individual use.