100% found this document useful (1 vote)
368 views12 pages1 - Process Capability Indices - 1986
1_Process Capability Indices_1986
Uploaded by
Yo GoldCopyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF or read online on Scribd
100% found this document useful (1 vote)
368 views12 pages1 - Process Capability Indices - 1986
1_Process Capability Indices_1986
Uploaded by
Yo GoldCopyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF or read online on Scribd
/ 12
Process Capability Indices
VICTOR E. KANE
Ford Motor Company, Transmission and Chassis Division, P.O. Box 2097 (7) Livonia, MI 48150
The copobility indices C,, CPU, CPL, k and Cy. are presented and related to process parometers.
These indices ore shown to form a complementary system of measures of process performance, ond
can be used with bilateral and unilateral tolerances, with or without target values. A number of Japanese
indus
es currently use the five indices ond the U.S. cutomotive industry has storted using the:
in number of areas. Various opplications of the indices are discussed along with statistical sampling
considerations.
Introduction
se quantfcation of pres location and variation
“Tis conat to undersanding the quality of un
oreduced rom a manuactring proce. Conair the
Stuation where the process mean,» and standard
Sevttion rare unknown and estimated by # and
reapectvely roma pracheal siewpoint and. sre
pot unless and sometimes ar no conveniant sm
macy satishes when hundreds of Characters a
plant or supply boo ae considered, In many sta
tons, capability nas can be used to relate the pro
tess parameters wand o engineering speciatons
that may include unilateral or bilateral tolerances
with or without target dimension Coominal value)
The eaultng indices are unis and provide a com
mon easiyunderston language for quantifying the
perfomance of proves
‘The automotive industry is currently involved in
an extensive effort to implement statistical process
control (SPC) in their plants and supply base. Capa
bility indices derived from SPC have received in
‘creased usage not only in process assessments, but in
evaluation of purchasing decisions. Of particular in-
terest have been the C, and Cy indices used in Japan
(Sullivan (1984, 1985)) and recently in the U.S. auto-
motive industry (e.g., Ford Motor Company (1984).
‘These indices relate the natural tolerance (62) used
in the US. quality control literature (e.g, Juran and
Gryna[1980 p. 299)) to engineering specifications. The
purpose of this paper is to examine uses of capability
indices, along with their sampling properties and es-
timation procedures. Throughout the discussion it is
Dr. Kane is Manager of the Applied Systems Analysis De
partment within the Transmission and Chassis Division of Ford
Motor Company. He i a CQE and a Member of ASQC
Vol. 18, No. 1, January 1986
4
assumed that the process output is approximately
normally distributed and in a state of statistical con-
trol. Also, a "hat" (*) will denote an estimated quan.
tity
G, Index
Manufacturing processes undergo several stages of
development prior to actual production. The stages of,
“prove out" testing at the machine supplier, initial
testing at the manufacturing facility, and pre-pro:
duction testing all seek to determine whether ma:
chinery can produce on an on-going basis production
units that meet the required engineering specifica
tions. A typical baseline is the assessment of whether
the natural tolerance (82) of a process is within the
specification limits. An alternate formulation used in
Japan (Sullivan [1984)) is to evaluate the capability
index C,. The above quantities are merely different
ways to relate the allowable process spread (part tol
erance) to the actual process spread (natural toler-
ance) as shown in Figure 1 where USL = upper spec:
ification limit, LSL = lower specification limit, and NT
= natural tolerance. The process potential index is,
allowable process spread
‘actual process spread
_ USL ~ bse
a
ust ~ Ls
a
NT
AG, of 1.0 indicates that 2 process is judged to be
“capable” as indicated in Figure 2. tis generally nec
essary to estimate the process standard deviation in
(1) s0 an estimate G, of the process capability is ob
tained. Due to sampling variation and machine testing
Journal of Quality Technology a VICTOR E. KANE
FIGURE 1. Relotionship of C, Parameters.
limitations discussed below, G, = 1.0 is generally not
used as a minimally acceptable value
A capable process with an underlying stable normal
distribution will in theory result in 0.27% of parts be-
yond the specification limits. The benchmark of 1.0
‘was chosen to relate C, to the standard six sigma
spread used on control charts. If a process is exactly
capable then
tice, =
Ist
ene @
upper control limit,
Lee
lower control limit
subgroup sample size on control chart
‘The relation in (2) indicates that sample size adjusted
specification limits are equal to contro! limits for @
process where (, = 1.0.
A primary use of C, is to make various types of
comparisons, Figure 3 shows how different C, values
relate to the spread of a process relative to the spec:
ification width. A minimum value of C, = 1.33is gen-
erally used for an ongoing process (e.g, Juran, Gryna
and Bingham 1979, pp. 9-22). Thisensuresa very low
reject rate (0.007%) and is thus an effective strategy
for prevention of nonconforming items. A value of C,
‘83 is also often used to qualify machinery since
long-term statistical control of a process is generally
not established during qualification trials. Using 1.83
gives some assurance that a C, = 1.0 will be possible
when the additional sources of variation are experi
enced in production processing. It should be noted
thatit is seemingly more natural to use the traditional
indicator
Journal of Quality Technology
% of specification used = (100) 2
‘The motivation for using C, rather than the % of spec
ification (or “capability ratio” from Charbonneau and
Webster (1978)) stems from the natural relationship
between process potential, quantified by C,, and pro:
cess performance, quantified by Cu, which is discussed
in later sections, Also, there are natural extensions to
‘unilateral tolerances and target dimensions using the
G, index,
Machine potential studies often do not consider
whether a machine is in statistical control. This is
clearly not advisable since many simple problems
could be identified if standard control charts were
used, even with short production runs of W parts (ie.,
N being between 30 and 100), Consideration of the
production order on control charts is a natural prob-
lem-identification tool. However, it is not possible to
determine if a machine will be in a state of statistical
control over the long-term with a short-term sampling,
plan, The question arises whether a short-term ca
pability study is of any value in predicting long-term
production performance. If itis assumed that an un-
stable process will have a larger standard deviation
than the same process which is stable, then a short
term process potential study will provide indication
of problem areas. Typically, a process is evaluated
using the estimated capability
‘SL — LS
@
cs
where s is the sample standard deviation computed
from a sample of size V. Unfortunately, short-term
studies have other problems, such as specially pre-
pared raw material, new tools, highly trained opera:
tors, and so on that can make the qualification process
FIGURE 2. Distribution of Individual Parts for @ Capable
Process (C, = 1.0)
Vol. 18, No. 1, Janvary 1986 PROCESS CAPABILITY INDICES 43
Parcent
of Specification 10 78
Rejects 0.27% 68 pew
0.6 Ppa
60 so
>0.1 Fem
FIGURE 3. C, Indices for Varying Widths of the Process Distribution
different from typical production operations. Gener.
ally, the special factors in short-term qualification
studies reduce variability that is likely in production
processing. Thus, these studies mainly serve to iden-
tify large problem areas, The G, index computed from
short-term studies generally identifies the lower bound
of variability (process potential) that can be expected
in the initial production setting. However, it is not
uncommon to be able to improve the initial production
G, index with later process refinements.
1g Considerations
‘The sampling variation of G, in (3) can be easily
studied sinee it is possible to use the chi-square dis
tribution of the sarmple variance (s*). Burr (1976, pp-
241) derives the operating characteristic (OC)
curve for testing the standard deviation using the fact
that (¥~ 1)s?/o® has a chi-square distribution with
(8 — 1) degrees of freedom, In an analogous manner,
itis possible to compute the OC curve for testing pro
cess capability. It should first be noted thatthe test
(o>)
Hea = 0
Hyo <0
is equivalent to a test for process capability (cy > 0)
Hs < ¢y (process is not capable)
@
Hy, > ey (process is capable).
Using a critical region of G > ¢, the power function
ofthe test (4) is obtained directly from the chi-square
distribution of the sample variance as
2(6,) = Paid, > eC)
Prixte-ny < (N-1)C%AIG,)
Vol. 18, No. 1, Janvory 1986
where xfu-1) is a chi-square random variable with
(— 1) degrees of freedom. Itis convenient to use the
power function for C, rather than the corresponding,
power function for a, since it is not necessary to adjust
for differences in process tolerances and a "standard:
ized” process capability can be studied directly
Using (5) the 0¢ curve OC(C,) = 1 ~ x(C,) can be
computed afd used to compare testing schemes. The
0¢ curve enables evaluation of the unfortunate prac:
tice of sometimes neglecting sampling variation of &
when assessing process capability. Suppose we wish
to evaluate whether a process is capable at the cy
[33 level (i.e., HC, < 1.33) using N = 30 samples
anda sample rejection limit (critical value) of ¢ = 1.33
‘The OC curve (a) in Figure 4 shows that 0C(1.33)
40, which implies that there is a 40% chance of
incorrectly judging the process not capable (accepting
He). The true process capability must be C, = 1.6 before
FIGURE 4. Operating Choracteristic Curve for Sampling
Plan that Rejects Process Copobilty if C, < ¢ Where (al
0; ¢ = 1.33, and (b} N= 70; ¢= 1-46.
Journal of Quality Technology 44 VICTOR E. KANE
there is only a 5% chance of judging a process not,
capable using a critical value of ¢ = 1.93.
Selection of a meaningful eritieal value for a ca
pability test requires specification of an acceptable
quality level (4QL) and rejectable quality level (ROL)
for the C, value (see Burr [1976, p. 336-941)). The AGL
isa sufficiently high process capability such that we
would like to accept processes with capabilities above
the AQL. The RQL is a sufficiently low process capability
such that we would like to reject processes with ca
pabilities below the RQL. Thus,
AQL = G,(high) > Clow) = ROL.
Using (5) the a and 8 risks are
a= x(6,(low))
= Prixiy-n < W — DeKlow)eM0,Clow)]
x(G,(high))
= Prixty-n < WW ~ L)esChigh}e ‘6, high)
Solving for O,(high)/¢,(low) and ¢ gives
Cxlrigh) | xbv-n(1- 8)
G,(low) ~~ xd ka)
+ oftowy =D.
ote)
where xiy- (7) is the 9(100)-th percentile ofthe ehi-
square distribution with (¥ ~ 1) degrees of freedom.
Table gives values of C,Chigh) /Clow)and ¢/C,(low)
for varying sample size and a = equal to 0.10 and
006, The eritical value ¢ and sample size N can be
determined for any specified Chigh)and Glow), For
example, if a ~ B= 008, C,(high) = 16, and C,(low)
2; then O,{high)/C,(low) = 1.38, and from Table
TABLE 1. Sample Size and Critical Value Determination
for Testing C,
ceo cao
fim) ae filo) ereenw)
10 188 Lar 228 137
2» 159 120 170 128
20 14 116 155 12
“0 14 i 145 18
50 130 Lis 10 16
80 Lar uu 138 115
0 125 110 12 1M
50 123 Lio 130 113
% 12 110 12a Le
100 120 09 135 un
Journal of Quality Technology
1:N = 70 and ¢ = (1.14)¢,(low) = 1.37. Thus, using a
critical value of 1.37 and a sample size of 70, there is
2.5% risk of judging a process with C, above 1.6 not
capable (accepting A) and a 5% risk of judging a pro:
cess with C, below 1.2 as capable (rejecting Hs)
In order to guarantee that any process with a C,
below 1.33 (where USL ~ LSL = 80) has a high proba
bility of being judged not capable (a = 0.1), and a
process with aC, above 1,66 (where USL ~ LSL = 10@)
has a high probability of being judged capable (8
1), then ¢,(high)/C,(low) = 1.28, and from Table
IN = 70 and c/C,(low) = 1.10, which gives ¢ = 1.46,
‘The OC curve for this sampling plan is curve (b) in
Figure 4.
‘The above example indicates that some past auto-
motive industry machine qualification practices using,
N = 30 were not generally adequate. Also, current
practices using a minimum ¢ = 1.38 may aecount for
nonsampling problems that make qualification runs
different from production runs, but a minimus of 1.33,
may not adequately account for sampling variability.
Larger N and/or higher minimums 6, values may be
necessary. Charbonneau and Webster (1978, p. 112)
suggest a capability index of 1.5 for new equipment,
but the sample size is not specified.
Tolerance Interval Approach
An alternate approach for assessing the perfor:
mance of a process is to construct an interval with
length 2Xs uch that 100p percent of the process output
is covered by the interval with probability 7. This is
a standard tolerance interval approach to process ca-
pability. Tolerance intervals, together with other types
of statistical intervals, are discussed by Hahn (1970),
It could be argued that the G, index in (4) should be
generalized to
5 _ USL LL
6a ©
G, if K = 8. If we choose N = 70, p = .99
then from Table 2 the tolerance interval
has length 6.042s, which is essentially the same as the
standard approach. However, other intervals would
be obtained for different N, p, and y. For example,
with N= 30 and the same p and y as before; X= 3.35
so the denominator of (6) becomes 6.7s. Thus, the in-
creased uncertainty for N= 30s reflected by a larger
‘multiplier of s in the denominator of (6).
Process Performance
It is apparent that the G, index measures potential
process performance since only the process spread is
related to the specification limits; the location of the
Vol. 18, No. 1, January 1986 PROCESS CAPABILITY INDICES 45
TABLE 2. Tolerance Intervol Factors
443 720
3.615
2.360 4.758
3213 4.499
0 3.126 328
50 3.068 4.206
70 3021 4120
0 2.986 14059
0 2.058 3.990
0 S04 3.954
process mean is not considered, It would be possible
to have any percentage of parts outside the specifi
cation limits with a high C, by merely locating the
process mean sufficiently close to a specification limit,
‘Thus, only potential performance of a process is
quantified by C, and can only be attained for the pro:
‘cess mean equal to the midpoint of the specification
‘The Cy index is related to the C, index, but utilizes
the process mean and can be considered a measure
of process performance. There are two equivalent
forms of the Cy. index. The first formulation is derived
by considering the upper and lower specification limits
separately. The second formulation uses the deviation
of the process mean from the midpoint of the speci
fication limits. Both derivations will be given since a
different insight is gained from each approach.
Upper and Lower Capability Index
Consider the unilateral tolerance situation where
only a single specification limit is given. Figure 5
shows the process distribution related to an upper
specification limit, Relating the actual and allowable
spreads in the same manner as (1) gives @ new CPU
index
allowable upper spread
cpu
‘actital upper spread
USL — w
om
2
_ use
30
o
‘The CPU index was developed in Japan and is utilized
by @ number of Japanese companies. In a similar
manner to (7) the index for the lower specification
limit is
Vol. 18, No. 1, January 1986
cp
Be
Estimates of CPU and GPL are denoted by CPU’ and CPi
and are obtained by replacing u and ¢ by and s,
respectively. The CPU (C, upper) and CPL (C, lower)
indices are related to C, by
(opu + cPLy/2
and are used to measure process capability in the sin:
sle specification limit situation, Of course, in the single
limmit case a CPU = 1.0 implies half as many noncon:
Forming items (0.136%) as when C, = 1.0 in the two
limit case.
Ga Index:
Since CPU and CPL utilize process location, a naturat
index for the bilateral specification case is Cj. where
Gu = Min( CPt, CPU} @
‘Thus, the Gy index relates the scaled distance between
the process mean and the closest specification lirit
A slightly different approach is to reduce the value of
G, for a process by a factor (1 ~ #), a scaled distance
that the process is off-center. Denote the midpoint of
the specification range by
m= (USL + L81)/2
‘The distance between the process mean, u, and the
optimum, the midpoint m, is w —_m where we are as-
suming m sw < USL as shown in Figure 6. The scaled
distance is
tt
eee
ALLOWABLE UPPER SPREAD = USL —»
FIGURE 5. Relationship of CPU Parameters.
Journal af Quality Technology 46 VICTOR E. KANE
aaa aoa
©" Ost ust
—
The absolute value sign was added for the case when
ISL 7 ~ LSL, and GPU = Oif
i? — pl > USL. ~ 7. These indices can be used for a one:
sided specification limit with a target value. The above
equations are analogous to (8) for Cy and correspond
to the CPL and CPU defined previously when 7 = 1.
fe)
t—__t
FIGURE 9. Relotionship of k Parameters Using Torget
Volues.
Journal of Quality Technology
FIGURE 10. Copobility Indices Using a Target Value of
Te 16.
A simple example illustrates the use of the indic
Suppose that USL = 18, ISL = 10,7 = 16nd ¢ = 067,
Note that from (11) G, = 1.0. Figure 10 shows the
capability indices for selected mean values. Notice that
because Cye adjusts for departures from 7, itis possible
to have a0 capability and not have any parts out of
the specification limits, Also, deviations on the “long,
side” (Je., maximum distance between Tand a spec:
ification limit) of T result in the same reduction in Cy
as deviations on the “short side” (e., minimum dis
tance between 7 and a specification limit).
Drawbacks
Experience to date has shown that there are poten
tial problems in using C, and Cj, on a routine basis,
‘These drawbacks generally stem from users having
an incomplete understanding of statistical principles
rather than from problems with the indices. Some of
the drawbacks are described below
(1) Statistical Controt—There is a tendency to want
to know the capability of a process before statistical
control has been established. Capability refers to a
quantification of common cause variation and what
can be expected from a process in the future, The
presence of special causes of variation make predic
Vol. 18, No. 1, Janvery 1986 PROCESS CAPABILITY INDICES 49
tion impossible and the meaning of a capability index
(2) Sampling Plan—Clearly, values of the average
range &, often used to estimate o, depend on the sam:
pling plan. Thus, it could be argued that the value of
a capability index could be made to change easily by
merely changing the sampling plan. The situation is
seemingly complicated further when a remark is con
sidered that W, Edwards Deming made to an engineer
who was noting his achievement at attaining statistical
control, Deming stated that he could make any process
appear in statistical control —merely space out the
‘samples within a subgroup over time. This statement
is recognition of the fact that the number of sources
of variation increases as the time interval between
samples within a subgroup increases, Spreading out
the samples within a subgroup over time will increase
#, widening the control limits, and thus makes achiev:
ing statistical control more likely. However, increasing
A will increase the estimate of and thus decrease
process capability. Conversely, itis possible to increase
process capability by using consecutive piece sampling
Ge, obtaining a small 2) to obtain a minimum esti
mate of o and maximum capability. However, in this
case statistical contral is the most difficult to attain
since the distance between control limits will be nar:
rower. Thus, considering only statistical control or
process capability separately is nota valid assessment
of process performance—both criteria rust be eval
uated jointly
(3) Computation—It is sometimes difficult to com.
pute G, and Cw on the plant floor where all individuals
are not familiar with mathematical formulas, This is
sue can be addressed in a variety of ways, but often
involves training and /or some type of automation,
(4) Nonnormality—A variety of processes result in
a nonnormal distribution for a characteristic. It is
probably reasonable to expect that the capability in
dices are somewhat sensitive to departures from nor
rmality Data transformations may be useful to attain
approximate normality. Also, tis possible to estimate
the percentage of parts outside the specification limits
either directly or with a fitted distribution, This per
centage can be related to an equivalent capability in
dex for a process having a normal distribution.
(5) Toul Wear—In a tool wear situation, & from a
consecutive piece sampling plan is of primary interest
‘Technically, & can be used to estimate C,, but the ca
pability is generally quite high. The performance of
the process depends on the tool change frequency
‘There is not a convenient extension of Cys to this sit
uation
Vol. 18, No. 1, Jonvory 1986
TABLE 3, Summary of Capebilty Indices
S ise = ust process potential for wo
& sed specification limits
coe process performance relative
to upper specification lit
ca process performance relative
to lower specification timit
‘
