7.

Process Capability Analysis

In practice, we want to know whether the products meet the engineering requirement. A
question that is commonly asked is that whether the distribution of individual
measurements will fit between the specification limits. We also want to know whether the
process is optimized in a sense that the mean is closed to the center of the specification
limits. Process capability study is aimed at answering these questions.

7.1. Introduction

Control charts can be used to monitor the process with respect to its real process
characteristics. However, there are engineering specifications for the process characteristic
that have to be met. A process can be in statistical control, but at the same time, it can
perform far from the requirements. Process capability measures the process variation and
it is important to compare with the natural tolerance or specification. It is defined as six
times the estimated standard deviation for the population, 6.

A commonly used formula for the estimation of  is

  R / d2

where the value of d2 which depends on the sample size can be obtained from standard
tables.

This is usually compared with the upper specification limit (USL) and lower specification
limit (LSL), or more exactly, the difference between USL and LSL. This difference is also
called the tolerance. If 6 <USL-LSL, the process variability is considered small
compared with the tolerance.

On the other hand, if 6 >USL-LSL, it is likely that for individual values to fall outside
the specification limits, and this is a situation that is not desirable. That is, the process is
not capable of producing items meeting the specifications. If this is the case, something
has to be done.
Quality Engineering (SYE3102) 64
[email protected]

However, unlike the situation when the control chart shows an out-of-control signal, when
the process is not capable meeting the specifications, the specification limits could be
adjusted if it is not achievable. This is especially when items produced are for internal use
and certain design changes can be made.

LSL Mean USL

Figure 7.1. Process capability is to measure the variation of

the process compared to engineering specifications.

7.2. Process Capability Indices

Process capability measures the process variation in relation to engineering specification.

The 6 concept, although easy to interpret, is difficult to use when comparing different
processes with each other. A relative measure is needed for this purpose.

The most common measure is the Cp index, which is defined as

USL  LSL
Cp 
6

where USL and LSL stand for upper specification limit and lower specification limit,
respectively.

It can be seen that 6 is greater or less than USL-LSL depending on whether Cp is greater
or less than one. However, the traditional rule of thumb is that for a process to be
considered capable, we should have a Cp value to be greater than 1.33. Sometimes an even
larger value of Cp might be required.
Quality Engineering (SYE3102) 65
[email protected]

Cp=(USL-LSL)/6

LSL Mean USL

Figure 7.2. Cp compares the population standard deviation

with the different between USL and LSL, or the tolerance.

Sometimes, we use the process capability ratio, Cr, which is defined as

6
Cr 
USL  LSL

This is just a related measure of process capability. The interpretation is however

different. The standard requirement is that Cr has to be less than 0.75 and the smaller, the
better.

Both Cr and Cp require the knowledge of USL and LSL. In some cases, only one
specification limit is available. This is usually the case for the-smaller the-better or the-
larger the-better process characteristics. In that case, we can define

USL  X X  LSL
C pU  , C pL 
3 3

to deal with one-sided specification limit. In case when both are available, this is still very
useful as to guide the process change and a common process capability index is

C pk  min(C pU , C pL )

This definition of process capability index is widely used also because it measures the
process performance in term of the target value.
Quality Engineering (SYE3102) 66
[email protected]

Mean-LSL
USL-Mean

LSL Mean USL

Figure 7.3. Cpk compare the population standard deviation

with the minimum of the USL-Mean and Mean-LSL.

7.3. Use and misuse of process probability indices

The capability of a process is a measure of its ability to satisfy customer requirements.

Process capability studies are carried out to compare customer requirements with process
performance. Process capability assessments are not limited to obtaining information
about product conformance. At the same time, information about significant process
features can be obtained. It should be noted that there is no evidence to suggest that any
process can have the capability of producing zero defects; there is always some risk of
producing bad items.

When using process capability indices, it is important to note that the specification limits
are not the same as the control limits and they should not be mixed up. Control limits are
based on the statistical analysis of the process while specification limits are based on
engineering judgment and requirement.

Specification limits are needed to ensure the products meeting the requirement for the use
at a later stage. Control limits, on the other hand, are used to distinguish between random
and assignable causes of process quality variation. Only when process is capable, we can
continue to use control chart as otherwise, the process has to be modified to meet the
specification.
Quality Engineering (SYE3102) 67
[email protected]

It can be noted that Cpk=Cp when the process is centered, that is, when the mean is equal
to (USL+LSL)/2. Because the process can never be exactly centered, Cpk<Cp. Hence, it is
sometimes only required for Cpk>1. A very small value of Cpk can be observed and even a
negative one is possible. This is the case when the process mean is outside of the
specification limits.

The process capability index Cpk was initially developed because process capability ratio
Cr does not adequately deal with the case of a process with mean that is not centered
between the specification limits. However, Cpk alone is still an inadequate measure of
process centering. To characterize process centering satisfactorily, the index Cpk has to be
used together with Cp and be compared with to Cp.

An important assumption underlying the theory of process capability indices is that their
usual interpretation is based on a normal distribution of process characteristic. If the
underlying distribution is not normal, then the statements about expected process fallout
attributed to a particular value of process capability indices might be wrong. So, the values
have to be interpreted based on the justification of the normality. This can usually be done
by looking at the histogram constructed using some earlier observations.

Theoretical process capability index assumes that the process mean and variation are
known. In practice, they have to be estimated and inserted in the formulas. The accuracy
of process capability assessment depends on the estimation of the underlying process
mean and variance. The estimation is affected by sample size, by the randomness of the
sample selection, by the process' state of control, among others. Histogram can also help
assessing process capability in practice.
Quality Engineering (SYE3102) 68
[email protected]

55
50
45
40
35
30
USL
25
20
LSL
(LSL+USL)/2
15
10
5

0
14.00
14.05
14.10
14.15
14.20
14.25
14.30
14.35
14.40
14.45
14.50
14.55
14.60
14.65
14.70
14.75
14.80
14.85
14.90
14.95
15.00
15.05
15.10
15.15
15.20
15.25

Figure 7.4. A histogram used together with specification limits. It can

be seen that although the process variation is small, the process is not
centered and the process capability can be improved by better centering.

Process improvement can be achieved by process centering and in many cases and more
effectively, by reducing the variation as illustrated below.

Improved process

Original process

LSL USL
Quality Engineering (SYE3102) 69
[email protected]

8. Acceptance Sampling

The quality of products after they are produced is an important aspect of quality control.
Although it is commonly agreed that controlling and improving the quality of the process
is what we should focus on at the beginning, there are situations that final inspection is
unavoidable. In this chapter, commonly used acceptance sampling procedures are
presented together with the discussion of their theoretical background. We focus on
sampling plans for attributes as they are the most common used.

8.1. Introduction to acceptance sampling

In most cases, 100 percent inspection of final products is not possible because of the time
and cost constraints. For example, inspection might destroy the item which means that it is
impossible to test all items at all. What we can do is to select a sample of a number of
items and inspect them. Decision can then be made based on the result of the sample.

Sampling takes less time and it less costly, so it is often preferred to the 100 percent
inspection. The major concern in acceptance sampling study is to assess product or
process quality characteristic by inspecting a random sample of certain size. Sample size
should be selected to yield an acceptable level of accuracy when we use it to predict the
quality level. The sample should be randomly selected and be representative of the entire
population of interest. A random sample is one wherein each element in the population has
an equal chance of being selected.

The goal of all sampling procedures is to estimate the properties of a population from the
properties of the sample. In particular, we wish to test the null hypothesis

H0: Lot is of acceptable quality

with the alternative given as

H1: Lot is of unacceptable quality

In acceptance sampling, we normally assume that the meaning of acceptable or

unacceptable quality is specified. This is usually done by specifying two quality levels:
Quality Engineering (SYE3102) 70
[email protected]

acceptable quality level p0 and unacceptable quality level p1 (and below). We also use
AQL to denote the acceptable quality level, p0 while p1 is usually called lot tolerance
percent defective (LTPD). AQL is the quality level of a good lot and it is the percent
nonconforming that can be considered satisfactory. LTPD is the quality level of a bad lot.
LTPD represents a level of quality that the consumer wants accepted with a low
probability of acceptance.

When the decision is based on a sample rather than the whole population, we are taking a
risk. Products of high quality might be rejected because the sample might happen to be a
bad one. On the other hand, poor quality products might be accepted if we happen to have
selected some good items by chance. What we should try to do is to minimize the risks.

The probability of accepting bad lots is called the consumer's risk and it is usually denoted
by . The probability of rejecting good lots is called the producer's risk and it is usually
denoted by . Hence

 =P(Reject H0 when H0 is true)=P(Reject lot when lot is good)

and

 =P(Accept H0 when H1 is true)=P(accept lot when lot is bad)

It is clear that a producer wants to keep the value of  small and while a consumer wants
to keep the value of  small. Hence, they are given the name Producer's risk and
Consumer's risk, respectively.

Assuming that we inspect n items in a lot and reject the lot when there are more than c
nonconforming items in the sample. We can easily determine the probability of accepting
or rejecting the lot for any given quality level p. Recall that the population is finite and we
can use hypergeometric probability distribution in this case. The probability that X=c is
given by

 M  N  M 
  
 c  n  c 
P ( X  c) 
N
 
n
Quality Engineering (SYE3102) 71
[email protected]

where N is the population size and M is the number of the items in the whole population
that fall into the class of interest. Also,

a a!
  
 b  b!(a  b)!

Note that the hypergeometric distribution can be approximated by the binomial

distribution when N is very large and the sample size is small, i.e., n<<N,. The binomial
distribution with parameter n and p=M/N is a good approximation to the hypergeometric
distribution in this situation. Because the sample size is usually much smaller than the
population this approximation is almost always a good one in acceptance sampling.

If a random variable X follows binomial distribution, the probability that X=c is given as

n!
P ( X  c)  p c (1  p) n c
c!(n  c)!

where p is commonly known as the proportion nonconforming in the population and n is

the sample size.

In some cases, we may use the normal approximation with mean np and variance np(1-p)
as an approximation to the binomial. The normal distribution to the binomial distribution
is good for p close to 0.5 and n>10. For other values of p, larger values of n are needed. It
is also possible to use Poisson distribution to approximate the binomial distribution. When
n is large and p small, we can use Poisson distribution with parameter ,=np to
approximate the binomial approximation. However, it is required that n is very large
which may not be the case in acceptance sampling.

Using the binomial approximation, the producer's risk can be calculated as

n
n
  P[ X  c / p  p0 ]    k  p k
0 (1  p 0 ) n  k
k  c 1  

and the consumer's risk is given by

Quality Engineering (SYE3102) 72
[email protected]

c
n
  P[ X  c / p  p1 ]     p1k (1  p1 ) n k
k 0  k 

An operating characteristic curve is a graphical description of a specific sampling plan

which shows the probability that the plan will accept lots of various quality levels. Figure
8.1 illustrates the operating characteristic (OC) curves.


 =producer's risk
Probability(accept)



 =consumer's risk
p

p0=AQL p1=LTPD

Figure 8.1. A typical OC curve.

The OC curve is a function of the true proportion of nonconforming in the lot. It decreases
as p increases. For the case of single sampling plan, we have that

c
n!
OC ( p )  P( X  c / p)   p k (1  p) n k
k 0 k ! ( n  k )!

It can be noted that the consumer's risk and the producer's risk can be obtained directly
from the OC curve. The consumer's risk is the value on the OC curve that corresponds to
p=p0 while the producer's risk is the value on the OC curve that corresponds to p=p1

It is clear that both the consumer's risk and the producer's risk decreases when the sample
size increases. This can be seen if a set of OC curves for different sample size are shown
together as in Figure 8.2.
Quality Engineering (SYE3102) 73
[email protected]


 increases
Probability(accept)



n  decreases

p

p0=AQL p1=LTPD

Figure 8.2. Both producer's risk and consumer's risk decrease

as the sample size increase.

An average outgoing quality (AOQ) curve shows the average expected quality in all
outgoing lots after the rejected lots from the sample have been 100 percent inspected and
all nonconforming items removed. Incoming lots with a small percentage of
nonconforming item will be passed with a resultant high outgoing quality. Those with a
slightly larger proportion of nonconforming items will result in the worst outgoing quality
because lots that have a large proportion of nonconforming will end up undergoing 100
percent inspection with only the acceptable items being passed.

The AOQ curve has the true percentage of nonconforming items in lots undergoing
inspection on the x axis and the percentage of nonconforming in lots of size N after
inspection on the y axis.
Quality Engineering (SYE3102) 74
[email protected]




AOQ


Proportion of nonconforming p

Figure 8.3. A typical AOQ curve.

In general, we have that

Pa Pd ( N  n)
AOQ 
N

where Pa is the probability of accepting the lot for various values of the percent
nonconforming (which can be obtained from the OC curve) and n is the sample size.

8.2. Common sampling procedures

There are different ways to classify acceptance sampling plans. One major classification is
by attributes and variables. Attributes are quality characteristics that are expressed "go,
no-go" basis. Here we only focus on acceptance sampling plans for attributes. Sampling
plans for variables can be studied using normal distribution and it is relatively more
straightforward.

The most simple sampling plan for attributes is the single sampling plan. A sample of size
n is selected and if the number of nonconforming items is more than c, then the lot is
rejected. On the other hand, the lot is accepted if the number is less or equal to c. Hence
the numbers (n,c) determine a single sampling plan. Most of the results presented in the
previous section are valid for single sampling plans.
Quality Engineering (SYE3102) 75
[email protected]

A common approach to the design of an acceptance sampling plan is to require that the
OC curve pass through two designated points. Suppose that we wish to construct a simple
sampling plan such that the probability of acceptance is 1- for lots with fraction of
nonconforming p0 and the probability of acceptance is  for lots with fraction
nonconforming p1. Then we can determine the sample size n and acceptance number c by
solving the equations for  and  given as

n
n
   k  p k
0 (1  p 0 ) n  k
k  c 1  

and

c
n
     p1k (1  p1 ) n  k
k 0  k 

However, it should be noted that there is no simple solution as the equations are not linear.
Furthermore, there is no exact solution as n and c can only take discrete values, so some
compromise has to be made. A binomial nomograph can sometimes be used.

Double sampling plan is another type of sampling plan and it might require two stages to
make the final decision. First, a sample of size n1 is selected. If the number of
nonconforming items in the sample is more than c2, the lot is rejected. If the number of
nonconforming items in the sample is less than c1<c2, the lot is accepted. On the other
hand, if the number of nonconforming items in the sample falls between c1 and c2, no
decision is made and another sample of size c2 is selected and inspected. The lot is
accepted or rejected depending on whether the cumulative number of nonconforming
items in both samples is less than or equal to c3. Normally c3=c2 and hence the whole
sampling procedure required (n1,n2,c1,c2) or four numbers.

A double sampling plan is more difficult to apply than a single sampling plan. However, it
can be more useful as it usually require less sampling in the long run than a single
sampling plan for the same levels of the consumer's risk and producer's risk. It is also
psychologically more acceptable because it provides a second chance before rejecting a
lot.
Quality Engineering (SYE3102) 76
[email protected]

A sequential sampling plan is a generalization of the double sampling plan. New samples
are drawn and inspected until a decision is made. After inspecting each sample, there are
three possibilities: (1) accept the lot; (2) reject the lot; (3) take another sample. The
decision after each sample is taken based on the cumulative number of nonconforming
items found in all previous samples. The sequential sampling plan is statistically sound.
However, it is not as commonly used as the single sampling plan because they are also
more complex.

REJECT
Cumulative number of
nonconforming items

TAKE
A NEW
SAMPLE

ACCEPT

Number of items inspected

Figure 8.4. An illustration of sequential sampling procedure.

8.3. The use of sampling standards

In practice, sampling plans are usually selected from quality handbooks. One of the most
widely used one is the MIL-STD-105D. It is commonly used for lot-by-lot inspection by
attributes and it was first developed in 1942. Since then there have been a number of
revisions.

The sampling plans discussed in the previous sections are individual sampling plans. A
sampling scheme is an overall strategy specifying the way in which sampling plans are to
be used. MIL-STD-105D is a collection of sampling schemes. It provides for all three
types of sampling: single sampling, double sampling and multiple sampling. For each type
of sampling plan, a provision is made for either normal inspection, tightened inspection or
Quality Engineering (SYE3102) 77
[email protected]

reduced inspection. Normal inspection is used at the start of the inspection activity.
Tightened inspection is instituted when the vendor's recent quality history has
deteriorated. Tightened inspection implies a more stringent inspection than under normal
inspection. Reduced inspection is instituted when the vendor's recent quality history has
been exceptionally good. The sample size generally used under reduced inspection is less
than that under normal inspection.

The primary item of MIL-STD-105D is the acceptable quality level which is the reference
point. The standard is indexed with respect to a series of AQLs. When the standard is used
for percent nonconforming plans, the AQL range from 0.1 to 10%. A sample size code
letter is selected for given batch size and inspection level. Then using the acceptable
quality level specified, a sampling plan ca be found from the standard.

Different types of sampling plans can be designed so that they produce equivalent results.
That is, these procedures can be designed so that a lot of specified quality has exactly the
same probability of acceptance under different types of sampling plans. Consequently,
when selecting the type of sampling procedure, one must consider factors such as the
administrative efficiency, the type of information produced by the plan, the average
amount of inspection required by the procedure, and the impact that a given procedure
may have on the material flow in the organization.