3/21/2012

Chapter 7

Process Capability

Introduction

• A “capable” process is one for which the distributions

of the process characteristics do lie almost entirely
within the engineering tolerances or customer’s needs.
• Process capability indices
– Simple
– Careful in use and interpretation
• Two phases in a process capability study
– Determining how to data are to be collected, and then
collecting the data
– Selecting 1 or more indices and performing the computations

1
 3/21/2012

7.1 Data Acquisition for

Capability Indices
• Data must come from an in-control process
• The sample must be representative of the population
• The sample size must be large enough
– To assess the extent of the non-normality
– To allow a non-normal distribution to be fit to the data

• Process capability indices <> Process Performance

indices

7.2 Process Capability Indices

• Should be easy to compute

• Should not be undermined by slight-to-moderate
departures from normality

2
 3/21/2012

7.2.1 𝐶𝑝

• One of the first capability indices

𝑈𝑆𝐿 − 𝐿𝑆𝐿
𝐶𝑝 =
6𝜎
• Example:
 = 180, 𝜎 = 25/6, USL = 100, LSL = 75 𝐶𝑝 = 1.0,
Fraction of non-conforming: ~100.00%
 = 87.5 𝐶𝑝 = 1.0, Fraction of non-conforming:~0.27%
• 𝐶𝑝 is a poor measure of process capability since it is
meaningless to consider the difference between the
specifications as a multiple of 6𝜎 while ignoring the process
mean, or a target value of the mean.
• 𝐶𝑝 measures only the potential process capability.

7.2.2 𝐶𝑝𝑚
𝑈𝑆𝐿 − 𝐿𝑆𝐿 (𝑈𝑆𝐿 − 𝐿𝑆𝐿)/6𝜎 𝐶𝑝
𝐶𝑝𝑚 = = =
6 𝜎 2 + (𝜇 − 𝑇)2 1 2 2 𝜇−𝑇
( ) 𝜎 + (𝜇 − 𝑇) 1 + ( 𝜎 )2
𝜎
(7.1)
• 𝐶𝑝𝑚 may be interpreted in terms of % loss resulting from
𝜇≠𝑇
• Example:
 = 75, 𝜎 = 5, USL = 110, LSL = 70, T = 80 𝐶𝑝𝑚 = 0.94,
Fraction of non-conforming: ~15.87%
 = 45, 𝜎 = 5, USL = 60, LSL = 20, T = 40 𝐶𝑝𝑚 = 0.94,
Fraction of non-conforming: ~0.135%
• It does not necessarily reflect process performance

3
 3/21/2012

7.2.3 𝐶𝑝𝑘

• Most frequently used process capability index

1 𝑈𝑆𝐿 − 𝜇 𝜇 − 𝐿𝑆𝐿
𝐶𝑝𝑘 = 𝑍𝑚𝑖𝑛 𝑤𝑕𝑒𝑟𝑒 𝑍𝑚𝑖𝑛 = 𝑚𝑖𝑛*𝑍1 = , 𝑍2 = +
3 𝜎 𝜎

𝑈𝑆𝐿 − 𝜇 𝜇 − 𝐿𝑆𝐿
𝐶𝑝𝑘 = 𝑚𝑖𝑛*𝐶𝑃𝑈 = , 𝐶𝑃𝐿 = +
3𝜎 3𝜎

(𝑑 − 𝜇 − 𝑚 ) 𝑈𝑆𝐿 − 𝐿𝑆𝐿 𝑈𝑆𝐿 + 𝐿𝑆𝐿

𝐶𝑝𝑘 = 𝑤𝑕𝑒𝑟𝑒 𝑑 = ,𝑚 =
3𝜎 2 2
• If the limits are at 𝜇 ± 𝑗𝜎, then 𝑍𝑚𝑖𝑛 = j and 𝐶𝑝𝑘 = 𝑗/3
• The use of 𝐶𝑝𝑘 is preferable when the limits are not
equidistant.

7.2.3 𝐶𝑝𝑘

• Since 𝐶𝑝𝑘 is a function of the process average, it is superior

to 𝐶𝑝
• 𝐶𝑝𝑘 could be negative if the distribution of a process
characteristics is very badly off-center and is either greater
than USL or less than LSL
• Neither 𝐶𝑝𝑘 nor 𝐶𝑝 is useful as a measure of process
capability when there is a target value for the process
mean. It is possible to have the value of 𝐶𝑝𝑘 or 𝐶𝑝 increase
because 𝜎 is decreasing and yet have the center of the
process distribution moving farther away from the target
value.

4
 3/21/2012

7.2.3 𝐶𝑝𝑘

• Another problem with 𝐶𝑝𝑘 is that there is not a 1-to-1

correspondence between 𝐶𝑝𝑘 and the % of non-conforming
• Example:
 = T = 30, 𝜎 = 10/3, USL = 40, LSL = 20, 𝐶𝑝𝑘 = 1.00,
Fraction of non-conforming: ~0.27%
 = 35.17, 𝜎 = 5/3, 𝐶𝑝𝑘 = 0.966,
Fraction of non-conforming: ~0.1878%

7.2.4 𝐶𝑝𝑚𝑘
𝜇 − 𝐿𝑆𝐿 𝑈𝑆𝐿 − 𝜇
𝐶𝑝𝑚𝑘 = min* , + (7.2)
3 𝜎2 + 𝜇 − 𝑇 2 3 𝜎2 + 𝜇 − 𝑇 2

• If 𝜇 = 12(𝐿𝑆𝐿 + 𝑈𝑆𝐿), then 𝐶𝑝𝑚 = 𝐶𝑝𝑚𝑘 .

• If 𝜇 is not equidistant from the specification limits, then
𝐶𝑝𝑚𝑘 < 𝐶𝑝𝑚 . 𝐶𝑝𝑚𝑘 is a better indicator.
• When 𝜇 = T, 𝐶𝑝𝑚𝑘 is equivalent to 𝐶𝑝𝑘
• 𝐶𝑝𝑚𝑘 is equivalent to 𝐶𝑝 when there are symmetric
specification limits and 𝜇 = T

5
 3/21/2012

7.3 Estimating the Parameters in

Process Capability Indices
• Process stability must first be established before the value
of an index is calculated.
• A logical approach would be to compute parameter
estimates using control chart data.
• It is desirable to use a large sample size.
• From X-Chart
𝑠
𝜎= 𝑎𝑛𝑑 𝜇 = 𝑥
𝑐4
• From 𝑋-Chart
𝑆 𝑅
𝜎= 𝑜𝑟 𝑎𝑛𝑑 𝜇 = 𝑥
𝑐4 𝑑2

7.4 Distributional Assumption for

Capability Indices
• It is assumed that the observations have come from a
normal distribution.
• A normal distribution is also assumed when the capability
indices are used.

6
 3/21/2012

7.5 Confidence Intervals for

Process Capability Indices
• Unless the sample size was large, it is desirable to also
report a confidence interval for the index.
• Lower confidence bound is more appropriate than a 2-
sided confidence interval.
• It is assumed that individual observations are used in
computing the parameter estimates (7.5.1~7.5.4)

7.5.1 Confidence Interval for 𝐶𝑝

• If the individual observations are assumed to have normal

distribution, then 𝑛 − 1 𝑆 2 /𝜎 2 has a 𝜒 2 distribution with
(𝑛 − 1) degrees of freedom.
2
𝑛 − 1 𝑆2
𝑃 𝜒 𝑛−1,𝛼 ≤ 2
≤ 𝜒 2 𝑛−1,1−𝛼 = 1 − 𝛼
2 𝜎 2

2 2
1 𝜒 𝑛−1,𝛼2 1 1 𝜒 𝑛−1,1−𝛼2
𝑃 ≤ ≤ =1−𝛼
𝑆 𝑛−1 𝜎 𝑆 𝑛−1
2 2
𝑈𝑆𝐿 − 𝐿𝑆𝐿 𝜒 𝑛−1,𝛼2 𝑈𝑆𝐿 − 𝐿𝑆𝐿 𝑈𝑆𝐿 − 𝐿𝑆𝐿 𝜒 𝑛−1,1−𝛼2
𝑃 ≤ ≤
6𝑆 𝑛−1 6𝜎 6𝑆 𝑛−1
= 1−𝛼

7
 3/21/2012

7.5.2 Confidence Interval for 𝐶𝑝𝑘

• Approximate 100(1-)% lower confidence limit for 𝐶𝑝𝑘

2
1 𝐶𝑝𝑘
𝐿𝑜𝑤𝑒𝑟 𝑙𝑖𝑚𝑖𝑡 = 𝐶𝑝𝑘 − 𝑧𝛼 + 𝑤𝑕𝑒𝑟𝑒 𝑎 = 𝑛 − 1
9𝑛 2𝑎 (7.3)
• Assume 𝐶𝑝𝑘 is approximately normally distributed

7.5.3 Confidence Interval for 𝐶𝑝𝑚

• Approximate 100(1-)% lower confidence bound

𝜒2𝜈
𝐿𝑜𝑤𝑒𝑟 𝑏𝑜𝑢𝑛𝑑 = 𝐶𝑝𝑚 𝑤𝑕𝑒𝑟𝑒
𝜈
(𝑛 + 𝜆)2 𝑛(𝑋 − 𝑇)2
𝜈= 𝑎𝑛𝑑 𝜆 =
𝑛 + 2𝜆 𝑆2

8
 3/21/2012

7.5.5 Confidence Intervals Computed

Using Data in Subgroups
𝜎= 𝐴𝑣𝑒 𝑠 2 𝑤𝑖𝑡𝑕 𝑘 𝑛 − 1 𝑑𝑒𝑔𝑟𝑒𝑒𝑠 𝑜𝑓 𝑓𝑟𝑒𝑒𝑑𝑜𝑚
𝑆 𝑅
𝑜𝑟 𝑜𝑟
𝑐4 𝑑2

𝑎𝑛𝑑 𝜇 = 𝑥

7.5.6 Nonparametric Capability

Indices and Confidence Limits
• Some quality characteristics such as diameter, roundness,
mold dimensions, and customer waiting time will be non-
normal, and flatness, runout, and % contamination will
have skewed distributions.
• Process capability indices are not robust to non-normality
in the individual observations.
• 4 approaches for non-normal distributions:
– Robust capability index
– Fit a distribution to a set of data and use percentiles in an index
– Transform the data to approximate normal
– Resample from the n sampled observations

9
 3/21/2012

7.5.6.1 Robust Capability Indices

𝑈𝑆𝐿 − 𝐿𝑆𝐿
𝐶𝜃 =
𝜃𝜎
𝑤𝑕𝑒𝑟𝑒 𝜃 𝑖𝑠 𝑐𝑕𝑜𝑠𝑒𝑛 𝑠𝑜 𝑡𝑕𝑎𝑡 𝑃 𝜇 − 𝜃𝜎 < 𝑋 < 𝜇 + 𝜃𝜎 ≅ .99

• 𝜃 = 5.15 is reasonable for 𝜒 2 distributions with various df

• The value of 𝜃 could be used in 𝐶𝑝𝑚 index
• Robust capability indices would not work for one-sided
indices (𝐶𝑝𝑘 and 𝐶𝑝𝑚𝑘 )

7.5.6.2 Capability Indices Based

on Fitted Distributions
• Let 𝑃𝛼 denote the 100 percentile of the fitted distribution, a
percentile-based 𝐶𝑝𝑘 would be
𝑈𝑆𝐿 − 𝑃0.5 𝑃0.5 − 𝐿𝑆𝐿
𝐶 ∗ 𝑝𝑘 = min( , )
𝑃0.9987 − 𝑃0.5 𝑃0.5 − 𝑃0.0013
• 𝐶 ∗ 𝑝𝑘 reduces to 𝐶𝑝𝑘 for a normal distribution.
• 𝐶 ∗ 𝑝𝑚𝑘 could be developed similarly.
• The success of this approach depends on the skill exhibited
in identifying and fitting an appropriate distribution.
• Nine distributions available in Minitab
– 1-parameter exponential -- 2-parameter exponential
– Smallest extreme value -- Weibull
– Largest extreme value -- 1/2/3 parameter logistic

10
 3/21/2012

7.5.6.3 Data Transformation

• Data can be transformed so the transformed data will be

approximately normally distributed.
• Lognormal data

7.5.6.4 Capability Indices Computed

Using Resampling Methods
• Resampling methods have been used to approximate
sampling distributions when no assumption is made of the
distribution of the random variable.
• Bootstrapping is one type of resampling.
– Naïve bootstrap: keep the original sample size, resample with
replacement
• The standard bootstrap methods can not be relied on to
produce a lower confidence limit for a capability index.

11
 3/21/2012

7.6 Asymmetric Bilateral

Tolerances
• Modified version of 𝐶𝑝𝑘
𝑇 − 𝐿𝑆𝐿 𝑇−𝜇 𝑈𝑆𝐿 − 𝑇 𝑇−𝜇
𝐶 ′ 𝑝𝑘 = max*min 1− , 1− , 0+
3𝜎 𝑇 − 𝐿𝑆𝐿 3𝜎 𝑈𝑆𝐿 − 𝑇
(7.4)

7.6.1 Example
A B C
USL 62 62 62
LSL 50 50 50
 50 56 62
 1 1 1
T 59 59 59
Cp 2 2 2
Cpk 0 2 0
Cpk’ 0 0 0
Cpm 0.2209 0.6325 0.6325
Cpmk 0 0.6325 0
% def. 0.500000 0.000000 0.500000

12
 3/21/2012

7.7 Capability Indices that are a

Function of % Non-conforming
• Capability Indices that are a function of the actual % Non-
conforming:
𝐶𝑝𝑝 = 13Φ−1 (1−𝑝2)
•  denotes the cumulative normal distribution function, and p
is the proportion non-conforming
• Capability Indices that are a function of the potential % Non-
conforming:
𝐶𝑝∗ = 13Φ−1 (1−𝑝∗
2
)
where p* is the potential proportion non-conforming
• Scaled distance measure
|𝑇 − 𝜇|
𝑘=
min*𝑇 − 𝐿𝑆𝐿, 𝑈𝑆𝐿 − 𝑇+

7.8 Modified 𝑘 Index

𝑇−𝜇 𝜇−𝑇
𝑘𝑁 = max* , +
𝑇 − 𝐿𝑆𝐿 𝑈𝑆𝐿 − 𝑇

13
 3/21/2012

7.11 Process Capability Indices

vs. Process Performance Indices
• A process capability index measures the capability of an in-
control process.
• A process performance index measures the performance of
a process, (either in-control or out of control).
• Estimation of parameters:
𝜎 = 𝑆 𝑎𝑛𝑑 𝜇 = 𝑥 𝑜𝑟 𝑥

Process Capability Indices Process Performance Indices

𝐶𝑝 𝑃𝑝
𝐶𝑝𝑘 𝑃𝑝𝑘

7.13 Software for Process

Capability Indices
• Minitab:
Process Capability Indices Process Performance Indices
𝐶𝑝 𝑃𝑝
𝐶𝑝𝑚
𝐶𝑝𝑘 𝑃𝑝𝑘
𝐶𝑃𝑈 𝑃𝑃𝑈
𝐶𝑃𝐿 𝑃𝑃𝐿

14