Anda di halaman 1dari 30

This article was downloaded by: [Northwestern University]

On: 21 December 2014, At: 20:33


Publisher: Taylor & Francis
Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer
House, 37-41 Mortimer Street, London W1T 3JH, UK

Communications in Statistics - Simulation and


Computation
Publication details, including instructions for authors and subscription information:
http://www.tandfonline.com/loi/lssp20

Brocess capability with asymmetric tolerances


a
Russell A Boyles
a
Industrial Research and Development , The New Zealand Institute , PO Box
31-310, Lower Hutt, New Zealand
Published online: 27 Jun 2007.

To cite this article: Russell A Boyles (1994) Brocess capability with asymmetric tolerances, Communications in
Statistics - Simulation and Computation, 23:3, 615-635, DOI: 10.1080/03610919408813190

To link to this article: http://dx.doi.org/10.1080/03610919408813190

PLEASE SCROLL DOWN FOR ARTICLE

Taylor & Francis makes every effort to ensure the accuracy of all the information (the “Content”)
contained in the publications on our platform. However, Taylor & Francis, our agents, and our licensors
make no representations or warranties whatsoever as to the accuracy, completeness, or suitability
for any purpose of the Content. Any opinions and views expressed in this publication are the opinions
and views of the authors, and are not the views of or endorsed by Taylor & Francis. The accuracy of
the Content should not be relied upon and should be independently verified with primary sources of
information. Taylor and Francis shall not be liable for any losses, actions, claims, proceedings, demands,
costs, expenses, damages, and other liabilities whatsoever or howsoever caused arising directly or
indirectly in connection with, in relation to or arising out of the use of the Content.
This article may be used for research, teaching, and private study purposes. Any substantial or
systematic reproduction, redistribution, reselling, loan, sub-licensing, systematic supply, or distribution
in any form to anyone is expressly forbidden. Terms & Conditions of access and use can be found at
http://www.tandfonline.com/page/terms-and-conditions
COMMUN. STATIST.-SIMULA., 23(3), 615-643 (1994)

PROCESS CAPABILITY WITH ASYlVIMETRIC


TOLERANCES

Russell A. Boyles

The New Zealand Institute


for Industrial Research and Development
PO Box 31-310,Lower Hutt, New Zealand

Key Words and Phrases: capability index; process yield; process cen-
tering; target value; loss function.

ABSTRACT
Many widely-used statistical packages and recent authors address-
ing process capability apply ,C, in cases where tolerances are asymmet-
Downloaded by [Northwestern University] at 20:33 21 December 2014

ric. This can misrepresent process capability, and points to the need to
apply appropriate capability measures in the asymmetric case. We
compare and contrast six relatively unfamiliar capability indices ap-
propriate for asymmetric tolerances to give them greater visibility and
to help potential users choose among them. For some of these indices
methods of estimation from process data have not been published; work
on this is deferred to subsequent papers.

L INTRODUCTION
A manufacturing process is characterized from the customer's point
ofview by numerical measurements taken on process operations or prod-
ucts. A specification for a process or product characteristic X consists

615

Copyright 8 1994 by Marcel Dekker, Inc.


616 BOYLES

of lower and upper limits LSL and USL together with a target value
T somewhere between these limits. On the assumption that succes-
sive measurements XI,X2,X3, . . . are identically distributed according
to some fixed distribution P , an assessment of process capability for X
consists of relating in some way the supplier's process P to the cus-
tomer's specification (LSL, T , USL).
Existence of a marginal distribution for the X process is clearly
the minimum logical prerequisite for discussing process capability. The
usual stronger assumption that XI,X2,X 3 , .. . are independent and iden-
tically distributed (iid) excludes stable processes with naturally-occurring
serial correlation, i.e., stationary processes (Anderson (1971),Box and
Jenkins (1976)). Recent authors have recognized that in many applica-
tions the iid concept of statistical control needs to be expanded to include
stationary processes (Montgomery and Mastrangelo (1991), Box and
Kramer (1992)). Although questions of statistical control pertain to the
conditional distribution P(Xt+l I Xt, Xt-2,. . .), questions of process
capability pertain to the marginal distribution P(X), so we can discuss
Downloaded by [Northwestern University] at 20:33 21 December 2014

process capability at a conceptual level without restricting ourselves to


iid processes.
A specification is symmetric if T = M = (USL + LSL)/2. In this case
USL - T and T - LSL have a common value known as the tolerance which
we denote as D. Although the symmetric case is by far the more com-
mon, there are sufiiciently many instances where T is not the midpoint
of (LSL,USL) to warrant consideration of asymmetric specifications for
which we have two distinct tolerances ( D l ,D,) defined by

Dl = T -LSL, D, = USL - T .

It is often assumed that asymmetric tolerances are necessarily re-


lated in some way to a skewed process distribution. For an example
of such a relationship, let P, denote the 100a percentile of the pro-
cess distribution. In some circumstances the customer may allow the
supplier to set a specification based on the process distribution. A nat-
ural way to do this would be (DL, T, USL) = (P,, %,,, PI-,) for some
PROCESS CAPABILITY 617

small positive a. For a skewed distribution we would clearly have


Dl = 'P0.5 - Pa # 'PI-, - %.5 = D,.
Asymmetric tolerances can also arise from transforming data to
achieve approximate Normality. Suppose X has a non-Normal dis-
tribution and a symmetric specification (LSL,T , USL). If Y = g ( X ) is
approximatelyNormal for some increasing nonlinear function g then it
may be convenient to work with Y instead of X, and Y will have the
asymmetric specification {g(LSL),g ( T ) ,g(USL)).
In general asymmetric tolerances are not related to the shape of
the supplier's process distribution; instead they simply reflect the cus-
tomer's view that deviations from target are less tolerable in one direc-
tion than in the other. When a symmetric specification is proposed for a
new product characteristic such a view may not become apparent until
a preliminary capability study shows the supplier's process unable to
satisfy the initial specification. The supplier may then find the customer
willing to expand the tolerance on one side of the target but not on the
other. This is a fairly common way for asymmetric tolerances to arise.
Downloaded by [Northwestern University] at 20:33 21 December 2014

For example, Figure 1shows a histogram of dimensional measure-


ments taken on 22 airfoil castings plotted against an asymmetric spec-
ification (39,42,51). This was an initial run of parts for a new product.
The specification for this particular characteristic had originally been
set at 42 f 3, but the initial data clearly showed that the supplier's pro-
cess could not hold the upper limit at 45. Rather than force the supplier
to perform some expensive retooling, the customer was willing to ex-
pand the upper specification limit to 51. Investigation of the apparent
bimodality of the distribution led to elimination of variability in some
operator-controllable set-up parameters, which in turn resulted in a
tighter distribution with a mean value closer to target. This gave an
acceptable process relative to the expanded tolerance.
For symmetric tolerances there are several well-known indices
widely used to measure the capability of a Normally-distributed pro-
cess characteristic. When the process mean p and standard deviation a
are known these capability indices are defined as follows: C, = (USL -
BOYLES

38 39 40 41 42 43 44 45 46 47 48 49 50 51 52

Section L-L Thickness

FIG 1. Histogram of dimensional measurements on 22 airfoil castings.


USL at 51 was originally at 45.
Downloaded by [Northwestern University] at 20:33 21 December 2014

LSL)/30, Cpi = ( p - LSL)/3a, Cpu= (USL - /~)/3a,Cpk= min{Cpl,Cpu},


and C,, = (USL - LSL)/~T where . T ~= o2 + ( p - T ) 2 .For a survey of re-
cent work on these indices, bibliographies of past work, and reminders
of limitations and possible misuses, readers should consult the October
1992 issue of the Journal of Quality !ikchnology (Vol. 24, No. 4) and the
recent book Kotz and Johnson (1993).
The indices C,, CpI,C,,, and Cpkform a system based on process
yield relative to the interval (LSL,USL) (Boyles(1991)),and are indepen-
dent of T, so there is no need to modify them when T # M. On the other
hand, the yield-based system, which fails to account for process cen-
tering with symmetric tolerances (Boyles (1991)),has an even greater
problem with asymmetric tolerances: process yield is maximized (for
fixed a) by /A = M but the customer has declared a target value T # M.
In this case yield and centering are conflicting criteria.
The first proposals specifically addressing asymmetric tolerances
consisted of replacing the true specification limits (T - Dl, T Du)with+
PROCESS CAPABILITY 619

the ersatz symmetric limits T fmin(Dr,D,), then applying the standard


definitions of C,, Cpk,and Cpm(Kane(1986);Chan, Cheng, and Spiring
(1988)).This approach yields "one-sided*indices C;, C& and C&,,which
can understate capability by restricting the process to a proper subset
of the actual specification range.
Another approach to the asymmetric case has been simply to ap-
ply the standard indices as written, including Cpm.This practice has
been adopted in many widely-used statistical packages (e.g., STAT-
GRAPHICS, STATISTICA, SPC-PC). Moreover, several recent authors
addressing process capability treat Cpmas though it were an appro-
priate measure for asymmetric tolerances (e.g., Kushler and Hurley
(1992),Franklin and Wasserman (1992)). This practice amounts to re-
placing the actual specification limits with the ersatz symmetric limits
T k (Dl + D,)/2. This can either over- or understate process capability,
depending on the position of p relative to T.
The basic problem with Cpmand C;,, when T # M is that process
shifts away from target are evaluated without respect to direction. For
Downloaded by [Northwestern University] at 20:33 21 December 2014

example, in Figure 2 Cpm= 1.2 and C;m = 0.6 for both processes A and
B. These processes are the same distance from target, but A has much
higher yield than B, so intuitively A should score higher than B. Others
have made similar observations (Kane (1986); Choi and Owen (1990);
Pearn, Johnson, and Kotz (1992);Kotz and Johnson (1993)).
An additional problem with Cpmwhen T # M is illustrated by
process C in Figure 2. For this process Cpm= 1, but the expected
percentage nonconforming is 6.7% rather than the 0.27% conventionally
associated with an index value c = 1. Cpmclearly overstates process
capability in this case.
These difficulties add to the body of evidence suggesting that capa-
bility indices are susceptible to misuse. On the other hand, capability
indicescan be very useful summary statistics in presentations of process
data, and undoubtedly will continue to play a role in customer-supplier
relations. On the assumption that the positive uses of capability indices
should continue, there is clearly a need for quality engineers, computer
BOYLES

FIG 2. Processes A and B both have C,, = 1.2 and C&, = 0.6. Process
C has C
,, = 1 but expected percent nonconforming of 6.7%.
Downloaded by [Northwestern University] at 20:33 21 December 2014

programmers, and statistical researchers to choose appropriate indices


when tolerances are asymmetric. There has been relatively little pub-
lished literature dealing specifically with the asymmetriccase, andlack-
ing sufficientguidance workers have by default simply continued to use
C,, and the yield-based system.
There have been several indices proposed for asymmetric toler-
ances which overcome the basic problem of C,, and C;, i.e., which score
A higher than B in Figure 2, but they have not been widely recognized.
Our intent here is to compare and contrast these relatively unfamiliar
indices to give them greater visibility and help potential users choose
among them. We use a combination of graphical and analytical methods
to gain operational understanding of the indices as functions of process
parameters. This study complements and substantially extends that of
Choi and Owen (1990).
The assumption of a Normally distributed process characteristic
does not cover all cases of practical interest, but processes for which one
PROCESS CAPABILITY 62 1

can sensibly discuss capability most often have characteristics which


are approximatelyNormal or which can be transformed to approximate
Normality via Y = g ( X ) for some g. With this in mind we compare and
contrast the indices using the Normal process model.
Methods for estimating capability indices from process data are
necessary and sometimes controversial. For example, the usual Normal-
theory methods for estimating indices designed for symmetric toler-
ances have been shown to lack robustness against non-Normality (e.g.,
Franklin and Wasserman (1992)). These difficulties can only get worse
when we move to indices designed for asymmetric tolerances. Nonnal-
theory methods have been developed for estimating some of the indices
studied here, while for others no methods have yet been published. Es-
timation methods are not addressed here; research on these problems
will be reported elsewhere.

2. THE YIELD-BASED SYSTEM


In this section we set the stage for later comparisons by discussing
Downloaded by [Northwestern University] at 20:33 21 December 2014

the yield-based system of process capability which is independent of T.


This system as usually applied consists of the indices C,, Cpl,C,,, and
Cpk.Let
a* = P r ( X < LSL), a , = P r ( X > USL)
so that a = a1 + a , is the expected fraction nonconforming. When
-
X N ( p ,2 )it is easy to derive the relations

and when p = M
cp= W1(1- a / 2 ) / 3
where @ is the standard Normal cumulative distribution function. These
three indices are transformations converting the exponential scale of
fraction nonconforming to an arithmetic scale standardized so that the
value 1 corresponds to "30 qualityn,413 corresponds to "40 quality", and
622 BOYLES

so on. These relations could also be used to define yield-basedindices for


a process with arbitrary cumulative distribution function F by taking

Similar proposals are mentioned in Kotz and Johnson (1993).


Unlike its brethren Cpk is only approximately related to process
yield (Boyles (1991)). A smooth version Spkwhich is a one-to-one trans-
form of fraction nonconforming (for Normal X ) can be obtained as fol-
lows. It is easy to show that

By plugging this into the Cp formula above we obtain

where
S ( X , Y ) = Q-' [{@(XI + @(Y)) I21 13.
Downloaded by [Northwestern University] at 20:33 21 December 2014

S satisfies S ( x ,x) = 213, so Spk = Cpk= Cp when p = M. For x # y we


have
min(x, y)/3 < S ( x ,y ) < S(min(x,y), +co)
which implies for p # M

It can be shown that the difference between Cpkand the right-hand


member above goes to 0 as Cpk increases, so Cpkcan (and should) be
viewed as an approximation to Spk. Figure 3 plots contours of Spk and
Cpkfor index values in the set {l/3,2/3,l,4/3,5/3,2). We will refer to
these as the standard values.
Spkinvolves the smooth function S ( x ,y) rather than min(x, y); Cpkis
easier to calculate by hand, and in many ways easier to work with ana-
lytically. We nevertheless use Spkas a baseline for comparison through-
out the paper because it represents actual rather than approximate pro-
PROCESS CAPABILITY 623

FIG 3. Contours of Cpk(dashed) and the smooth modification Spk(solid)


for the standard values: 113, 213, 1,413, 513, 2 (top to bottom in plot).
Downloaded by [Northwestern University] at 20:33 21 December 2014

cess yield, and because the difference between Spk and Cpkhas greater
ramifications when T $: M than when T = M.

3. SIX INDICES FOR ASYMMETRIC TOLERANCES


In this section we describe six capability indices for asymmetric
tolerances which take direction into account when evaluating process
shifts away from target. Four of these have appeared in published or
unpublished reports in recent years, two are new. We use contour plots
to visualize the indices in relation to the yield-based index Spk,and we
consider process centering in the form of bounds placed on lp - TI by
each of the indices. General discussion and conclusions are presented
in Section 4.

3.1 C,, and a Smooth Modification


Aside from the "one-sidednindices discussed in Section 1,one of the
earliest proposals for asymmetric tolerances is the Cpnindex defined by
BOY LES

Choi and Owen (1990) as


p - L S L USL-p
Cpn= min -
( 3r ' 3r
with T defined as in Section 1. Earlier references to a similar index are
cited by Rodriguez (1992). This index is also discussed in the symmetric
case by Pearn, Kotz, and Johnson (1992), who refer to it as Cpmk(see
also Kotz and Johnson (1993)).
Figure 4 plots contours of C p , (dashed) and Spk(solid)for the stan-
dard values in a symmetric case (a) and an asymmetric case (b). In the
symmetric case C p , I Spk with equality only for p = T. In the asymmet-
ric case C p , = CPk < Spk when p = T, and Cpn < C p k5 Spk when p # T,
so C,, < Spk for all values of p. Thus when T # M a process with Cp, = c
has fraction nonconforming strictly less than the value a = 2{1- @ ( 3 c ) )
corresponding to a process with Spk = C .
With respect to process centering, the condition Cpn2 c implies
Downloaded by [Northwestern University] at 20:33 21 December 2014

We assume c > 113. Separate consideration of the cases p < T and


p > T then yields

For example, with c = 1 this gives 38 < p < 46 for the symmetric case in
Figure 4(a) and 46 < p < 52 for the asymmetric case in Figure 4(b).
By analogy with C p kand Spkwe define a smooth version of C p , by

Using properties of S presented in Section 2 we have Spn = Cpnwhen


p = M,andforp # M

Figure 5 plots contours of Sp, (dashed) and Spk(solid) for the standard
values in a symmetric case (a)and an asymmetric case (b). In both cases
PROCESS CAPABILITY
Downloaded by [Northwestern University] at 20:33 21 December 2014

16 LSL T

14
12
10
m
k
i?
8
6
4
2
0

FIG 4. Contours of C,, (dashed) and S p k (solid)for the standard values


in a symmetric case (a) and an asymmetric case (b).
BOY LES
Downloaded by [Northwestern University] at 20:33 21 December 2014

FIG 5. Contours of S,, (dashed) and Spk (solid) for the standard values
in a symmetric case (a) and an asymmetric case (b).
PROCESS CAPABILITY 627

Sp, < Spkwith equality only for p = T . The contours of Spn are tangent
to those of Spk at p = T, SO regardless of the value of T processes with
Sp, = c will have cu = 2{1 - Q ( 3 c ) ) when p = T and cu x 2 { 1 - @ ( 3 c ) )for
p in a neighborhood of T . For large c this appealing property also holds
approximately for C .,
With respect to process centering, we note that Spn 2 c implies
Cpn> C' where
c' = @-'{2@(3c) - 1 ) / 3 < c.
For example, c' = 0.927 when c = 1. Thus S,, 2 c implies

T - m i n ----
D" <p<T+min
(3:;1'3d- 1
) (- "' ,A)
3c'-1 3 c ' S l'

For example, with c = 1 this gives 37.8 < p < 46.2 for the symmetric case
in Figure 5(a)and 45.5 < p < 52.1 for the asymmetric case in Figure 5(b).

3.2 C,, Generalizations


A different approach to asymmetrictolerances arising in the Taguchi
school of quality engineering (Taguchi,Elsayed, and Hsiang (1989))in-
Downloaded by [Northwestern University] at 20:33 21 December 2014

volves the use of a piecewise quadratic loss function

A I ( X- T ) ' / D ? if x 5 T
C ( x )=
A,(x - T ) ' / D ; ifx > T

where A l and A , represent monetary losses associated with outcomes at


the lower and upper specification limits, respectively, Taguchi, Elsayed,
and Hsiang (1989)represent process capabilityby the expectedloss E ( C )
and specifically reject the use of unitless indices. The idea of expressing
process capability in monetary terms is very appealing, but in some
cases it may be extremely difficult to assign meaningful values to A1 and
A,, while in other cases it may be desirable to keep such information
confidential. As previously noted by Johnson (1992),in such situations
there may be an interest in unitless measures derived from C .
With symmetric tolerances ,C, can be expressed as 1 / 3 6 where
A0 = E { C O ( X ) )and C O ( x ) = ( x - T ) ' / D 2 . This is a special case of
the relationship developed by Johnson (1992). A class of generalized
628 BOYLES

C,, indices is obtained by considering quantities of the form 1 / 3 6


where X = E { L ( X ) )for some choice of the constants (A,,A,). Prior to
discussing particular choices for ( A [A,), , we derive the general form of
X when X N ( P ,a2).
N

Let Z = (X - p ) / a , ( = ( T - p ) / u , andlet [event] denote the indicator


function of event. Then X can be written as

where

-
When X N(p,a 2 )we can express each of the parameters and T: as
u2(Iz- 2(I1+ C210)where Ik is an integral of the form
Downloaded by [Northwestern University] at 20:33 21 December 2014

d) being the ~ V ( 0 , l density


) function. Using the property # ( z ) = -zd)(z)
and integration b y parts we obtain

We now evaluate I2 - 2C11 + C210,taking (a,b) = (-m, C) for T: and


(a, b) = (C,SCO)for T;, to obtain

Finally our expression for X is

+
Note that h(-C) = 1 C2 - h(C).
PROCESS CAPABILITY 629

The most natural choice for the constants ( A l ,A , ) is A1 = A , = 1 .


We denote the resulting index by Cim.The superscript * is used to dis-
tinguish this index from c,,, to which it reduces in the symmetric case,
from the "one-sided"index C;, (Chan, Cheng, and Spiring (1988)),and
from C&, discussed below. Kotz and Johnson (1993) also discuss C;,
citing an unpublished report of the author's subsumed in the present
paper.
Figure 6 plots contours of C& (dashed) and Spk(solid)for the stan-
dard values in a symmetric case (a) and an asymmetric case (b). In the
symmetric case C,; = Cpm.The contours of Cp, are tangent to those of
Spkat p = T, so Cp, shares with Sp, the appealing property that pro-
cesses with C, = c will have a M 2{1 - Q ( 3 c ) ) for p in a neighborhood
of T. Unfortunately this property does not extend to the asymmetric
case. Figure 6(b) shows that processes with C& 2 c may have Spk< C ,
i.e., CY > 2 { 1 - Q ( 3 c ) ) . With respect to process centering, recall that in
the symmetric case Cpm2 c implies Ip - TI < D / 3 c (Boyles (1991)). We
show in the following paragraph that
Downloaded by [Northwestern University] at 20:33 21 December 2014

whenever C,; > c. For example, with c = 1this gives 36.7 < p < 47.3 for
the symmetric case in Figure 6(a)and 42.0 < p < 52.7 for the asymmetric
case in Figure 6(b).
We first prove the following for fixed p: (i) X is an increasing function
of u, and
Al(p - T ) 2 / D ; 2 if p < T
(ii) limloX = ifp=T
A,(p -T)2/D: if p > T.
+
To demonstrate (i), it is easily verified that hl(C) = 2{C@(C) d ( C ) ) , and
using this it is straightforward to calculate

'Ib prove (ii) note that


BOYLES

m
k
iij
Downloaded by [Northwestern University] at 20:33 21 December 2014

FIG 6. Contours of Spk (solid) with (a) C,, (dashed) in a symmetric


case, (b) C* (dashed) in an asymmetric case, and (c) c,+,(dashed) in an
rrn
asymmetric case.
PROCESS CAPABILITY

22 26 30 34 38 42 46 50 54 58 62

Mu

FIG 6 . Continued
Downloaded by [Northwestern University] at 20:33 21 December 2014

from which follows

( p - T)' i f p < T
lim a2h(C)=
a10 0 i f p '~ ~

with a similar argument showing that

( p - T ) 2 if p >T
lim a 2 h ( - ( ) =
010 0 ifpsT

Suppose now that 1/3& 1 c. Then (i) and (ii) imply

hence
BOYLES

Taking Al = A , = 1 now yields the desired result for C&.


Since Cimlacks calibration with process yield (Spk)one might seek
another index of the form 1 / 3 6 . By analogy with S,, it would make
sense to seek constants ( A l ,A,) for which the index would agree with
SVkwhen p = T, but it turns out this is not possible. It is, however,
possible to choose A , = A , = A so that the index agrees with Cpkwhen
p = T. The resulting equation is

which simplifies to

where r = Dl/D,. The index corresponding to Al = A, = A i r ) is denoted


C& and satisfies C,i, = c&/ JA(r). Since A ( r ) 2 1 with equality only
when r = 1, Clm 5 C& with equality only in the symmetric case where
Downloaded by [Northwestern University] at 20:33 21 December 2014

they both reduce to C,,. Choi and Owen (1990) discuss C&, citing an
unpublished preliminary draft of Boyles (1991).
Figure 6(c) plots contours of C,f, (dashed) and Svk(solid) for the
standard values in an asymmetric case. For c < 1 the index is con-
servative with respect to process yield in that processes with C,+,= c
have cu < 2{1 - Q ( 3 c ) ) . For c > 1 processes with C,+, = c can still have
cu > 2{1 - @ ( 3 c ) }although
, not by as much as with.,C ; With respect to
process centering, the analysis of X carried out above implies that
T -D l / 3 c m< p <T + D,/~c&
whenever C,+, 2 c. In Figure 6(c) we have A ( r ) = A ( 3 ) = 1.8 so for c = 1
we obtain 44.0 < p < 52.0.
We have investigated several indices of the form 1 / 3 6 where A1 #
A,, but have found nothing worth reporting.

3.3 Cikp and a Smooth Modification


Another recent proposal is the Johnson-Kotz-Pearn index Cik, (Kotz
and Johnson (1993))which can be written as
PROCESS CAPABILITY 633

where the parameters T: = a 2 h ( ( ) and r,2 = a 2 h ( - ( ) are defined as in


Section 3.2. Kotz and Johnson (1993)discuss Cjkpin the context of non-
normal process distribution~and do not emphasize its applicability to
asymmetric tolerances.
Figure 7 plots contours of C j k p(dashed) and Spk(solid)for the stan-
dard values in a symmetric case (a) and an asymmetric case (b). As
with the C,, generalizations in Section 3.2, in the asymmetric case it is
possible to have a > 2 { 1 - @ ( 3 c ) )when Cjkp2 c. With respect to process
centering, it is straightforward to infer from the analysis of X in Section
3.2 that
T -D ~ / ~ J< Z pC < T + 0 , / 3 J Z c
whenever Cjkp2 c. For example, with c = 1 this gives 38.2 < p < 45.8 for
the symmetric case in Figure 7(a)and 44.3 < p < 51.9 for the asymmetric
case in Figure 7(b).
We can define a smooth version of Cjkpby
Downloaded by [Northwestern University] at 20:33 21 December 2014

Using properties of S presented in Section 2 we have

with Sjkp = Cjkponly if 01/71 = D u / r U . In the symmetric case the


condition for equality reduces to p = T; in general the point of equality
corresponds to the unique solution of

all quantities being defined as in Section 3.2.


Figure 8 plots contours of Sjkp (dashed) and Spk (solid)for the stan-
dard values in a symmetric case (a) and an asymmetric case (b). The
strange behavior of the Sjkpcontours for c 5 113 was unexpected, and we
have little to offer by way of explanation. For purposes of comparison
we have retained the common scaling of Figures 3-7, even though this
causes the contours for c = 113 to run outside the plot frame.
BOYLES

m
.-E,
(I)
6
Downloaded by [Northwestern University] at 20:33 21 December 2014

FIG 7. Contours of Cikp(dashed)and Spk(solid)for the standard values


in a symmetric case (a) and an asymmetric case (b).
PROCESS CAPABILITY
Downloaded by [Northwestern University] at 20:33 21 December 2014

LSL

14
----- - - - - _ _ _ _
12

FIG 8. Contours of Sjkp(dashed) and Spk(solid) for the standard values


in a symmetric case (a) and an asymmetric case (b).
636 BOYLES

As the contours show, Sjkp= Spkwhen ,u = T regardless of the value


of T. However, the contours are tangent at ,u = T only in the symmetric
case, and in the asymmetric case it is possible to have a > 2{l - Q(3c))
when Sjkp2 c. With respect to process centering, Sjkp 2 c implies
Cjkp> cl, with c' < c defined as in Section 3.1. Thus Sjkp> c implies

For example, with c = 1 this gives 37.9 < ,u < 46.1 for the symmetric case
in Figure 8(a) and 43.9 < p < 52.0 for the asymmetric case in Figure
Wb).

4. GENERAL DISCUSSION
The capability indices described in Section 3 were illustrated in a
symmetric case ( r = D,/D, = 1) and an asymmetric case with r = 3. In
this section we provide global comparisons of the indices for specifica-
tions spanning the range of r values likely to occur in practice. As before
we focus first on relationships to the yield-based index Spk,and second
Downloaded by [Northwestern University] at 20:33 21 December 2014

on process centering in the form of bounds placed on ( p - T I . Lastly we


present our conclusions and recommendations.

4.1 Process Yield


Inspection of the contour plots in Section 3 reveals that with asym-
metric tolerances the indices are maximized (for fixed u) not by p = T,
but by a value p* between T and M. The location of p* relative to T and
M reflects the compromise established by each index between process
centering and process yield. We equivalently observe that as we move
along a contour, a increases from 0 to a maximum value a*then back to
0 again, where a* corresponds to p* between T and M. Let S;, denote
the Spkvalue of the (p*, a * ) process. For each index S,*,is a function of
r and the index value c. At a given level c we may compare indices in
terms of process yield over a continuum of asymmetric specifications by
plotting the corresponding S,', as functions of r. This is done in Figure 9,
following preparatory analyses in the following paragraphs.
PROCESS CAPABILITY 637

FIG 9. Process yield in terms of Spkfor a continuum of processes with


constant values c = 1 and c = 513 for the four indices shown. The curves
are symmetric about r = D,/D, = 1 on a logarithmic scale.
Downloaded by [Northwestern University] at 20:33 21 December 2014

Since Cpn5 Spk,for Cpnwe always have S;, 2 c. What remains is to


determine how far S,', may vary above c. The condition Cpn= c implies

o2 = ( ( p - ,!3L)2/9~2
- (p- T)2 if p < M
(USL - LSL)2/36c2- ( M - T ) 2 if Cl = M
2 (p- T)2
(USL - p ) 2 / 9 ~- if p > M.

We assume c > 113. It is then straightforward to show that a 2 is max-


imized for p 5 M by p = min{M,T + D1/(9c2 - 1 ) ) and for p > M by
p = max{M, T - D,/(9c2 - 1 ) ) . Separate considerationof the cases T < M
and T > M yields

T + D1/(9c2 - 1) if r 5 (9c2- 1)/(9c2+ 1)


T - D,/(9c2 - 1) if r > (9c2+ 1)/(9c2- I )
M otherwise
638 BOYLES

from which we obtain

D:/(9c2 - 1) if r I (9c2- 1)/(9c2+ 1)


= D t / ( 9 c 2- 1 ) if r > (9c2+ 1)/(9c2- 1 )
{(Dl + 0,)' - 9c2(Di- D , ) 2 ) / 3 6 ~ 2otherwise.
It is now straightforward to calculate (p* - LSL)/a* and (USL - p*)/u*
as functions of ( r ,c) from which S;, is obtained via

s;, = s[ ( p a- L S L ) / u S ,(USL - p*)/o*].


The ( p * , a * )process is more difficult to characterize for the smooth
version S,, than for C,,. Since C ,, 5 S,, < Spk,it can be shown that for
any level c the S;, curve for S,, is bounded between c and the S,', curve
for C,,. Accordingly, we represent both indices by the results for C,,.
For the C,, generalizations, the condition 1 / 3 6 = c implies
Downloaded by [Northwestern University] at 20:33 21 December 2014

Using hl(C)= 2 { ( @ ( C ) + 4 ( ( )it} is straightforward to show d2 ( u - ~/ d) p 2 >


0 so that we find the unique minimum by solving 8 / d p = 0. After
some rearranging, and assuming Al = A,, this yields

Let (' = ( ' ( r ) denote the unique solution of this equation, which can be
calculated by simple iteration (k = g(&-1) or other numerical methods.
It is straightforward to verify that

Once C* and o* are found the corresponding process mean is given by


p* = T - ('a*. We then have
PROCESS CAPABILITY

where
+
Di/a* = 3 c \ I ~ l @ ( C * )r2Au@(-C*)

+
D,/a* = 3 ~ 4 r - ~ ~ i @ ( (A*, @
) (-(*).
We take Ai = A, = 1 for C,*, and A1 = A, = A ( r ) for CA, where A ( r ) is
defined as in Section 3.2.
Turning now to CJkp,it is easy to show that a is maximized for a
given value c when D I / = ~ D U / r u . After rearranging this yields the
equation ( = f (() where f is defined in Section 3.3. The unique solution
(* = ( ' ( r ) can be calculated by simple iteration Ck = f((k-l) or other
numerical methods. Then

a* = ~ i / 3J2h(o
c = D , / ~ J2h(-o
c

and
S;, = s (-C* +3 4 5 , C* + 3cJh(-C.)) .
The ( p * ,a * )process is more difficult to characterize for the smooth ver-
sion Sjkpthan for Cjkp. As suggested by comparison of Figure 8(b) with
Downloaded by [Northwestern University] at 20:33 21 December 2014

Figure 7(b),it can be shown that ( p * , a * )for Sjkpis closely approximated


by ( p * ,a * )for Cjc. Accordingly, we represent both indices by the results
for Cjkp.
Plots of S,', us. r are shown in Figure 9 for C,,, C,",, C&, and Cjkp
at index levels c = 1 and c = 513. The curves are shown only for r > 1
since they are symmetric about r = 1 on a logarithmic scale.
Only C,, (and S,,) guarantee process yield at or above nominal
index levels for all values of r . The curves for C& and Cjkprapidly
fall below reference index values as r departs from 1. The curves for
C,+,start out roughly matching those for C,, as r departs from 1, then
fall off,eventually dropping below reference index values. The range
of specifications for which C,+, guarantees process yield at or above a
nominal index level c depends on c. For example, C,+,= 1 guarantees
Spk 2 1 for 115 < r < 5, but C&, = 513 implies Spk2 513 only for an
interval slightly larger than 112 5 r 5 2.
640 BOYLES

Table 1: Constants for bounds on Ip - TI implied by various indices.

1 + 1
Spn m i n { l / ( 3 c 1 I ) , 1/r(3c1- I)} min{r/(3c1- I ) , 1/(3c1+ 1 ) } 1

4.2 Process Centering


For each of the indices discussed in Section 3.3 we derived bounds
on ] p - TI as necessary conditions that the index be no less than a given
level c. These bounds can be expressed in the form
Downloaded by [Northwestern University] at 20:33 21 December 2014

for unitless functions ( k ~k,), of ( r ,c) given in Table 1. In the table c1 is


defined as in Section 3.1 and A ( r ) is defined as in Section 3.2.
We plot (-kl, k,) us. r in Figure 10 for c = 413. The value 0 on
the vertical axis represents T. Positive values on this axis represent
fractions of D, and negative values represent fractions of Dl. The curves
for Spnand Sjkp are omitted to simplify the plot; the ( k ~k,), values for
these indices are very close to but slightly larger than those for Cp, and
Cjkp,respectively
The limits for Cjkpare tighter that those for C,+,, with the difference
becoming negligible for large or small r. The limits for C&, are tighter
than those for C& except for r = 1 when the indices coincide. For
all three of these indices (and Sjkp) the limits are symmetric about 0,
indicative of an appealing proportionality to tolerance on both sides
of the specification. Cpn (and Spn)are not symmetric about 0. They
produce limits which are proportional to tolerance on the "short" side of
PROCESS CAPABILITY 64 1

FIG 10. Bounds placed on 1p - TI for c = 413 by the four indices shown.
Positive values represent fractions of D,, negative values represent
fractions of Dl,
and 0 represents T.
Downloaded by [Northwestern University] at 20:33 21 December 2014

the specification, but on the "long" side the bounds become increasing
sharp relative to tolerance as r increases or decreases. This reflects a
tendency to understate capability for processes with p relatively close
to T on the "longnside of the specification.

4.3 Conclusions and Recommendations


Cpnand Spnare clearly superior to the other indices studied here in
terms of process yield, while the other four indices have more appealing
properties with respect to process centering. Of these four, only C,+,
should remain under consideration, since C;m, C j k pand , Sjkpprovide no
protection at all with respect to process yield.
The differences between Cpn and Spn become negligible as index
value c increases, Cpn is easier to calculate by hand, and in many
ways easier to work with analytically. On the other hand, S,, is closely
related to actual process yield ( S p dfor all c, while Cpnis closely related
642 BOYLES

to approximate yield (Cpd Thus we view Cpnas an approximation to


S,,. This leaves us with S,, and C& to choose between.
Spnguarantees levels of process yield conventionally associated
with given index levels c across all r values, while C,f, provides such
guarantees for ranges of (r, c ) likely to include many practical appli-
cations. C&, places bounds on p proportional to tolerance, while the
bounds given by Sp, are disproportionately sharp on the "long" side of
the specification. Of course, it can be argued that it is precisely this
property which keeps S,, well-calibrated with respect to process yield.
In using an index which places some sort of bounds on ) p-TI, practition-
ers are likely to be more concerned with failures to guarantee expected
levels of process yield than with relatively subtle issues involving the
precise nature of these bounds. For these reasons, among all the indices
reviewed here we find Spnmost appropriate for general use.

BIBLIOGRAPHY
Downloaded by [Northwestern University] at 20:33 21 December 2014

Anderson, T. W. (1971). The Statistical Analysis of !lime Series. John


Wiley & Sons, Inc., New York, NY.
Box, G. E. P. and Jenkins, G . M. (1976). lime Series Analysis, 2nd ed.
Holden-Day, San Francisco, CA.
Box, G. E. P. and Kramer T. (1992). "Statistical Process Monitoring
and Feedback Adjustmenb-A Discussion". Dchnometrics 34, pp.
25 1-267.
Boyles, R. A. (1991). "The Taguchi Capability Index". Journal of
Quality Dchnology 23, pp. 17-26. Corrigenda pp. 382-3.
Chan, L. K.; Cheng, S. W.; and Spiring, F. A. (1988). "A New Measure
of Process Capability: Cpm".Journal of Quality l%chnology20, pp,
162-175.
Choi, B. C. and Owen, D. B. (1990). "A Study of a New Process Capa-
bility Index". Communications in Statistics-Theory and Methods
19, pp. 1231-1245.
PROCESS CAPABILITY 643

Franklin, L. A. and Wasserman, G. S. (1992). "Bootstrap Lower Confi-


dence Limits for Capability Indices". Journal of Quality 12?chnol-
ogy 24, pp.196-210.
Johnson, T. (1992). "The Relationship of C,, to Squared Error Loss".
Journal of Quality lkchnology 24, pp. 211-215.
Kane, V. E. (1986). "Process Capability Indices". Journal of Quality
7bchnology 18, pp. 41-52. Corrigenda p. 265.
Kushler, R. H. and Hurley, P. (1992). "ConfidenceBounds for Capability
Indices". Journal of Quality lkchnology 24, pp. 188-195.
Kotz, S. and Johnson, N. L. (1993). Process Capability Indices. Chap-
man & Hall, London.
Montgomery, D. C. and Mastrangelo, C. M. (1991). "Some Statistical
Process Control Methods for Autocorrelated Data". Journal of
Quality Tbchnology 23,179-193.
Pearn, W. L.; Kotz, S.; and Johnson, N. L. (1992). "Distributional and
Inferential Properties of Process Capability Indices". Journal of
Quality l??chnology24, pp. 216-231.
Downloaded by [Northwestern University] at 20:33 21 December 2014

Taguchi, G.; Elsayad, E. A.; and Hsiang, T. C. (1989). Quality Engi-


neering i n Production Systems. McGraw-Hill, New York, NY.

Received December, 1993; Revised January, 1994.

Anda mungkin juga menyukai