CHAPTER ONE

1 Quality Improvement in the Modern Business Environment

1.1The Meaning of Quality and quality Improvement
We may define quality in many ways. Most people have a conceptual understanding of quality as
relating to one or more desirable characteristics that a product or service should possess.
Although this conceptual understanding is certainly a useful starting point, we will give a more
precise and useful definition.
Quality has become one of the most important consumer decision factors in the selection among
competing products and services. The phenomenon is widespread, regardless of whether the
consumer is an individual, an industrial organization, a retail store, or a military defense
program. Consequently, understanding and improving quality key factors are leading to business
success, growth, and enhanced competitiveness. There is a substantial return on investment from
improved quality and from successfully employing quality as an integral part of overall business
strategy. In this section we provide operational definitions of quality and quality improvement.
We begin with a brief discussion of the different dimensions of quality and some basic
terminology.
1.1.1 Dimensions of Quality
The quality of a product can be evaluated in several ways. It is often very important to
differentiate these different dimensions of quality. Garvin (1987) provides an excellent
discussion of eight dimensions of quality. The following are his key points concerning these
dimensions.

1. Performance (will the product do the intended job?)

2. Reliability (how often does the product fail?)
3. Durability (how long does the product last?)
4. Serviceability (how easy is it to repair the product?)
5. Aesthetics (what does the product look like?)
6. Features (what does the product do?)
7. Perceived quality (what is the reputation of the company or its product?)
8. Conformance to standards (is the product made exactly as the designer intended?)
We usually think of a high quality product as one that exactly meets the requirements placed on
it.
Consequently, a simple answer to questions such as "What is quality?" or "What is quality
improvement?" is not easy. The traditional definition of quality is based on the viewpoint that
products and services must meet the requirements of those who use them.
1.1.2 Definition of Statistical Quality Control (SQC)
Quality improvement is the reduction of variability in processes and products.
Statistical Quality Assurance (SQA) refers to a program for the systematic monitoring and
evaluation of the various aspects of a project, service, or facility to ensure that standards of
quality are being met. In short, it is the set of activities that ensures the quality levels of products
and services are properly maintained and that supplier and customer quality issues are properly
resolved. Two key principles characterized QA: "fit for purpose" (the product should be suitable
for the intended purpose) and "right first time" (mistakes should be eliminated). QA includes
regulation of the quality of raw materials, assemblies, products and components; services related
to production; and management, production and inspection processes.
Statistical quality control or SQC emphasizes testing of products to uncover defects, and
reporting to management who make the decision to allow or deny the release. That is, SQC is the
area of statistics concerned with applying statistical methods to improving the quality of
products/services. Hence, the role of statistical quality control is to collect and analyses relevant
data for the purpose of detecting weather the process is under control on not. If not, what can
possibly be the reason for faults?

Example 1: A MOHA company produced bottles of various soft drinks. Bottles of Pepsi are
supposed to have 300ml of drink. Bottles are filled by a machine so some will have more than
300ml, some less if the machine is set incorrectly then the bottle will filled too much the
company lose money or the bottle will filled too little customer complain. A company will take a
sample of bottles. If the mean volume is significantly different from 300ml it includes something
wrong in the process.

Example 2: An electrical company makes light bulbs that are suppose to have an average life
time of 1000 hours. A buyer intends to buy a batch of 10,000 bulbs. The buyer first test a sample
of 100 light bulbs if they have a mean life time of ≥ 1000 hours he accepts the batch otherwise
the batch is rejected.

Example 3: A company makes cups and plates. Some of cups might be chipped or cracked. A
sample will be taken to cheek the proportion of cracked cups. If this is above a certain value the
process must be adjusted. In all above three examples

 There is a cost involving in sampling.

 There is a decision to be made that can result an error. For these case wrongly rejects a
good products it results supplier or producer risk as a result of good products are rejected.
Whereas wrongly accepted a bad products a consumers or buyers at risk as a result of bad
or defective products are distributed to in a certain market. The main aims of statistical
quality control (SQC) is that to develop and control statistical procedures that have
 Low cost i.e. the ability to minimize cost.
 Low producer’s risk. By taking remedial action and by controlling the production process, it
reduces supplier or generally company risk.
 Low buyer’s risk. This implies by checking and controlling the production process it
minimize or reduces the consumer’s risk, by taking a carefully designed hypothesis testing.
In statistical quality control (SQC) we focus on three major areas: SPC, DE and AS.
Statistical Quality Control (SQC): involves investigating the procedures used during the
manufacturing process. The growth of SPC coincided with the growth of factories a way that of
producing goods. A control chart is one of the first techniques of SPC. It is a very useful process
monitoring methods when unusual sources of variability are present.
A designed experiment is extremely helpful in discovering the key variables influencing the
quality characteristics of interest in the process. A designed experiment is an approach to
systematically varying the controllable input factors in the process and determining the effect
these factors have on the output product parameters. Statistical designed experiments are
invaluable in reducing the variability in the quality characteristics and in determining the levels
of the controllable tables that optimize process performance. Often significant breakthroughs in
process performance and product quality also result from using designed experiments.
Acceptance Sampling (AS): defined as the inspection and classification of a sample of units
selected at random form a larger batch or lot and the ultimate decision about disposition of the
lot, usually occurs at two points: incoming raw materials or final productions. In sequential
sampling after we have sampled the “ith” item there are three possibilities.

 Accept the batch

 Reject the batch
 Sample another item.
1.1.2 A Brief History of Quality Control and Improvement

Statistical methods and their application in quality improvement have had a long history. How
long? It is safe to say that when manufacturing began and competition accompanied
manufacturing, consumers would compare and choose the most attractive product. If
manufacturer A discovered that manufacturer B's profits soared, the former tried to improve
his/her offerings, probably by improving the quality of the output, and/or lowering the price.
Improvement of quality did not necessarily stop with the product - but also included the process
used for making the product.

On the other hand, statistical quality control is comparatively new. The science of statistics itself
goes back only two to three centuries. And its greatest developments have taken place during the
20th century. The earlier applications were made in astronomy and physics and in the biological
and social sciences. It was not until the 1920s that statistical theory began to be applied
effectively to quality control as a result of the development of sampling theory. The first to apply
the newly discovered statistical methods to the problem of quality control was Walter A.
Shewhart of the Bell Telephone Laboratories. He issued a memorandum on May 16, 1924 that
featured a sketch of a modern control chart. In 1920s Shewhart introduces the control charts, this
controls the quality of a product as it is being made rather than after it is made. This case is key
components of statistical process control (SPC).
In 1940s because of a war there were a lot of researchers and researches in to statistical process
control would introduce sequential sampling that is was introduced by Wald. For example, take
a sample of 100 items if > 20 items are bad or defective we will reject the batch. This implies
non–sequential sampling; in non–sequential sampling we need try to sample a fixed number of
items. In sequential sampling after we have sampled the “i th” item there are three possibilities:
Accept the batch, Reject the batch, and Sample another item.
In the 1950s, Japan’s industry has been destroyed by wars, they invited American statistician,
“Deming” to introduce statistical methods to industry. Deming introduced “14 point” of good
management and partiality. As a result Japan’s industry has grown. Now statistical quality
control and Deming methods are used in many industries’ worldwide. A brief discussion of
W.E. Deming “14 points” are the following.

This philosophy is summarized in his 14 points for management. We now give a brief statement
and discussion of Deming's 14 points:

1. Create a constancy of purpose focused on the improvement of products and services.

Constantly try to improve product design and performance. Investment in research,
development, and innovation will have long-term payback to the organization.
2. Adopt a new philosophy that recognizes we are in a different economic era. Reject poor
workmanship, defective products, or bad service. It costs as much to produce a defective unit
as it does to produce a good one (and sometimes more). The cost of dealing with scrap,
rework, and other losses created by defectives is an enormous drain on company resources.
3. Do not rely on mass inspection to "control" quality. All inspection can do is sort out
defectives, and at this point it is too late because we have already paid to produce these
defectives. Inspection typically occurs too late in the process, it is expensive, and it is often
ineffective. Quality results from prevention of defectives through process improvement, not
inspection.
4. Do not award business to suppliers on the basis of price alone, but also consider quality. Price
is a meaningful measure of a supplier's product only if it is considered in relation to a
measure of quality. In other words, the total cost of the item must be considered, not just the
purchase price. When quality is considered, the lowest bidder frequently is not the low-cost
supplier. Preference should be given to suppliers who use modem methods of quality
improvement in their business and who can demonstrate process control and capability.
5. Focus on continuous improvement. Constantly try to improve the production and service
system. Involve the workforce in these activities and make use of statistical methods,
particularly the statistically based problem-solving tools discussed in this book.
6. Practice modem training methods and invest in on-the-job training for all employees.
Everyone should be trained in the technical aspects of their job, and in modem quality- and
 productivity-improvement methods as well. The training should encourage all employees to
practice these methods every day.
7. Improve leadership, and practice modem supervision methods. Supervision should not consist
merely of passive surveillance of workers but should be focused on helping the employees
improve the system in which they work. The number one goal of supervision should be to
improve the work system and the product.
8. Drive out fear. Many workers are afraid to ask questions, report problems, or point out
conditions that are barriers to quality and effective production. In many organizations the
economic loss associated with fear is large; only management can eliminate fear.
9. Break down the barriers between functional areas of the business. Teamwork among different
organizational units is essential for effective quality and productivity improvement to take
place.
10. Eliminate targets, slogans, and numerical goals or the workforce. A target such as "zero
defects" is useless without a plan for the achievement of this objective. In fact, these slogans
and "programs" are usually counterproductive. Work to improve the system and provide
information on that.
11. Eliminate numerical quotas and work standards. These standards have historically been set
without regard to quality. Work standards are often symptoms of management’s inability to
understand the work process and to provide an effective management system focused on
improving this process.
12. Remove the barriers that discourage employees from doing their jobs. Management must
listen to employee suggestions, comments, and complaints. The person who is doing the job
knows the most about it and usually has valuable ideas about how to make the process work
more effectively. The workforce is an important participant in the business, and not just an
opponent in collective bargaining.
13. Institute an ongoing program of education for all employees. Education in simple, powerful
statistical techniques should be mandatory for all employees. Use of the basic SPC problem
solving toots, particularly the control char, should become widespread in the business. As
these charts become widespread and as employees understand their uses, they will be more
likely to look for the causes of poor quality and to identify process improvements. Education
is a way of making everyone partners in the quality improvement process.
14. Create a structure in top management that will vigorously advocate the first 13 points.
Quality Systems and Standards
The International Standards Organization (founded in 1946 in Geneva, Switzerland), known as
ISO, has developed a series of standards for quality systems. The first standards were issued in
1987. The current version of the standard is known as the ISO 9000 series. '
It is a generic standard, broadly applicable to may type of organization, and it is often used to
demonstrate a supplier's ability to control its processes. The three standards of ISO 9000 are:
 ISO 9000:2000 Quality Management System--Fundamentals and Vocabulary
 ISO 9001:2000 Quality Management System--Requirements
 ISO 9004:2000 Quality Management System--Guidelines for Performance Improvement
 ISO 9000 is also an American National Standards Institute and an ASQ standard.
1.2 Modeling Process Quality
The use of statistical methodology in quality control and improvement can be done by first how
simple tools of descriptive statistics can be used to be express variation quantitatively in quality
characteristics when a sample of data on these characteristics is available. The second is based on
probability distributions and show how they provide a tool for modeling or describing the quality
characteristics of process.
1.2.1 Describing Variation
There are several graphical methods that are very useful for summarizing and presenting data.
Some of the most useful graphical techniques are the steam-and-leaf plot, the histogram, the box-
plot, scatter diagram and soon. The steam-and-leaf plot display allows us to quickly determine
some important features of the data that are not obvious from the data table whereas histogram is
more compact summary of data than a steam-and-leaf plot. The Box-plot can display several
important features of the data simultaneously like: measure of central tendency, variability,
quartiles, and minima-maxima.
Graphical methods listed in above provide a visual display of the shape, central tendency and
variability of the distribution of the data set only. Here numerical summary of the data is also
helpful to use numerical measures of central tendency and scatter. The most important numerical
measures of central tendency and scatter in the sample data are mean and standard deviation
respectively.
1.2.2 Probability Distributions [Reading Assignment]
Probability distribution is a mathematical model that relates the value of the variable with the
probability of occurrence of that value in the population. There are two types of probability
distributions-discrete and continuous.
1. Continuous distributions: When the variable being measured is expressed on a continuous
scale, its' probability distribution is called a continuous distribution The probability
distribution of metal layer thickness is continuous.
2. Discrete distributions: When the parameter being measured can only take on certain values,
such as the integers 0, 1, 2..... the probability distribution is called a discrete distribution. For
example, the distribution of the number of non-conformities or defects in printed circuit
boards would be a discrete distribution.
Examples
1. A Discrete Distribution
A manufacturing process produces thousands of semiconductor chips per day. On the average,
1% of these chips do not conform to specifications. Every hour, .an inspector selects a random
sample of 25 chips and classifies each chip in the sample as conforming or nonconforming. If we
let x be the random variable representing the number of nonconforming chips in the sample, then

the probability distribution of x is: p(x) = (25x )(0.01) ( 0.99)

x 25− x
, where x= 0,1,2 ..... 25.

This is a discrete distribution, since the observed number of non-conformances is x = 0, 1, 2 ....,

25, and is called the binomial distribution.
We may calculate the probability of finding one or fewer nonconforming parts in the sample as:
1

x=0 x ( )
P(≤ 1¿ = P(x = 0) + P(x=1) = ∑ 25 (0.01) (0.99)
x 25− x

= (0.99)25*25*0.01*(0.99)24
= 0.7778+0.1964
= 0.9742
2. A Discrete Distribution
Suppose that x is a random variable that represents the actual contents in ounces of a l-lb bag of
1
coffee beans. The probability distribution of x is assumed to be: f(x) = , 15.5 < x < 17.0
1.5
This is a continuous distribution, since the range of x is the interval [15.5, 17.0]. This distribution
is called the uniform distribution.
Note that the area under the function f(x) corresponds to probability, so that the probability of a
17.0
1 1
bag containing less than 16.0 oz is: P(x < 16.o) = ∫ dx = 0.3333.
15.5 1.5 1.5

Important Discrete and Continuous Distributions

Several discrete and continuous probability distributions arise frequently in statistical quality
control. Some of these distributions are hyper-geometric, binomial, Poisson, Pascal and negative
binomial from discrete and the normal, the standard normal, the lognormal, exponential, the
gamma, and the Weibull from continuous probability distributions. In certain quality problems, it
is sometimes useful to approximate one probability distribution with another. This is particularly
helpful in situations where the original distribution is difficult to manipulate analytically. The
binomial approximations to the hyper-geometric, the Poisson approximation to the binomial, and
the normal approximation to the binomial are some useful approximation.
Assignment I [5%]
1. A manufacturer of electronic calculators offers a 1-year warranty. If the calculator fails for
any reason during this period, it is released. The time to failure is well modeled by the
following probability: f(x) = 0.125 f ( x )=0.125 e−0.125 x , x >0. Then, what percentage of the
calculators will fail within the warranty period? 0.118
2. A production process operates with 2% nonconforming output. Every hour a sample of 50
units of product is taken, and the number of nonconforming units is counted. If one or more
nonconforming units are found, the process is stopped and the quality control technician must
search for the cause of nonconforming production. Then what is the probability that the
process is stopped? 0.636
3. A random sample of 100 units is drawn a production process every half hour. The fraction of
nonconforming product manufactured is 0.03. What is the probability that ^p ≤ 0.04 if the
fraction nonconforming really is 0.03? 0.818
4. An electronic component for a laser range finder is produced in lots of size N=25. An
acceptance testing procedure is used by the purchaser to protect against lots that contain too
many nonconforming components. The procedure consists of selecting five components at
random from the lot (without replacement) and testing them. If none of the components is
nonconforming the lot is accepted.
A) If the lot contains three nonconforming components what is the probability of acceptance
the lot? 0.496
B) Calculate the desired probability in (A) using the binomial approximation. Is this
approximation satisfactory? Why or why not? 0.528
5. The life of an automotive battery is normally distributed with a mean of 900 days and
standard deviation of 35 days. What fraction of these batteries would be expected to survive
beyond 100 days? 0.21%
CHAPTER TWO

2 Methods of Statistical Process and Capability Analysis

2.1 Methods and Philosophy of Statistical Process Control

If a product is to meet or exceed customer expectations, generally it should be produced by a

process that is stable or repeatable. More precisely, the process must be capable of operating
with little variability around the target or nominal dimensions of the product's quality
characteristics. Statistical process control (SPC) is a powerful collection of problem-solving tools
useful in achieving process stability and improving capability through the reduction of
variability.
SPC builds an environment in which all individuals in an organization seek continuous
improvement in quality and productivity. This environment is best developed when
managements becomes involved in the process. Once this environment is established, the
organization is well on its way to achieving its quality improvement objectives.

SPC is an effective method of monitoring a process through the use of control charts. Control
charts enable the use of objective criteria for distinguishing background variation from events of
significance based on statistical techniques. Much of its power lies in the ability to monitor both
process center and its variation about that center, by collecting data from samples at various
points within the process. Variations in the process that may affect the quality of the end product
or service can be detected and corrected, thus reducing waste as well as the likelihood that
problems will be passed on to the customer. With its emphasis on early detection and prevention
of problems, SPC has a distinct advantage over quality methods, such as inspection, that apply
resources to detecting and correcting problems in the end product or service.

The underlying concept of statistical process control is based on a comparison of what is

happening today with what happened previously. We take a snapshot of how the process
typically performs or build a model of how we think the process will perform and calculate
control limits for the expected measurements of the output of the process. Then we collect data
from the process and compare the data to the control limits.

In addition to reducing waste, SPC can lead to a reduction in the time required to produce the
product or service from end to end. This is partially due to a diminished likelihood that the final
product will have to be reworked, but it may also result from using SPC data to identify
bottlenecks, wait times, and other sources of delays within the process. Process cycle time
reductions coupled with improvements in yield have made SPC a valuable tool from both a cost
reduction and a customer satisfaction standpoint.

2.1.1 Chance and Assignable Causes of Quality Variation

No production process is good enough to produce all items exactly alike in spite of sophisticated
technology and precession of modern scientific equipments; no two articles produced by the
same machine are perfectly identical in measurable characteristics. Some amount of variation,
through of an infinitesimal nature, is inherit and inevitable in every repetitive process in industry.
These variation are broadly termed as resulting from Chance and Assignable causes
Chance cause: In any production process, regardless of bow well designed or carefully
maintained it is, a certain amount of inherent or natural-variability will always exist. This natural
variability or background noise is the cumulative effect of many small, essentially unavoidable
causes. In the framework of statistical quality control, this natural variability is often called a
stable system of chance cause. The natural variation is beyond the control of human hand and
cannot detect or prevented and it is tolerable and does not martially affect the quality of product.
A process that is operating with only chance causes of variation present is said to be under
statistical quality control or in statistical control. Some stable patterns of variation or constant
cause system are inherent in every scheme of production and inspection. The variation due to
these chance causes can be regard as independent normal variation zero means and same error
variance says σ 2
.

Assignable cause: an assignable cause system is one in which the cause of variation are not
distributed normally but not sporadically. Assignable cause also known as non random cause and
the variation due to these cause is termed as chaotic or erratic or preventive variation. Some of
the important factors of assignable cause of variation generally are improperly adjusted or
controlled machines, operator error and/or substandard or defective raw materials (new
techniques or operations, negligence of operation, wrong or improper handling of the machines ,
faulty equipment, unskilled or in experienced technical staff, and so on). A process that is
operating in the presence of assignable causes is said to be out-of-control. These cause can be
identified and eliminated and have to be traced in any production process before the production
becomes defective. If the article show marked deviation from the given specification of product,
the utility of articles is jeopardy. In that situation, one has to make a search for the cause of
responsible for the large variation in the product. Hence, the role of SQC is to collect and
analysis relevant data for the purpose of detecting weather the process is under control or not.
The control chart is a primary process monitoring techniques widely used for this purpose.

2.1.2 Statistical Basis of the Control Chart

A major objective of statistical process control is quickly detest the occurrence of assignable
causes of process shifts so that investigation of the process and corrective action may be
undertaken before many nonconforming units are manufactured. The control chart is an on-line
process-monitoring technique widely used for this purpose. Control charts may also be used to
estimate the parameters of a production process, and, through this information, to determine
process capability. The control chart may also provide information useful in improving the
process. Finally, remember that the eventual goal of statistical process control is the elimination
of variability in the process. It may not be possible to completely eliminate variability, but the
control chart is an effective too1 in reducing variability as much as possible.
Basic Principles
One of the most important tools of production management and control of quality in the
manufactured product is the ‘control charts’ technique. The discovery and development of
control charts were made by a young physicist Walter A. Shewhart of bell telephone laboratories
in 1924 and the following years and is regards as a land mark in the history of industrial quality
control. Based on the theory of probability and sampling, it enables us to detect the presence of
assignable causes of erratic variation in the process. These causes are then identified and
eliminated and the process is established and controlled at desired performances. In other words,
the control chart is a graphical display of quality characteristics that has been measured or
computed from a sample various sample number of time. Control charts in this section are
classified and described according to three general types: variables(X-bar, range, s and median
charts), attributes (p, np, and c charts) and cumulative (CUSUM) chart.
If quality characteristic has been measured or computed from a sample, the control chart shows
the value of the quality characteristic versus the sample number or versus time. In general, the
chart contains a center line that represents the mean value (target value) of the quality
characteristics corresponding to the in-control state or process. Two other horizontal lines,
called the upper control limit (UCL) and the lower control limit (LCL) are also shown on the
chart. These control limits are chosen so that almost all of the data points will fall within these
limits as long as the process remains in-control. As long as the point plot within the control
limits, the process is assumed to be in control, and no action is necessary; otherwise the process
is out-of-control, and correction actions are required to find and eliminate the assignable causes.

Note that:

 Even if, all the point plots inside the control limits, if they behave in a systematic or non
random manner, then this could be an indication that the process is out-of-control. (If the
process is in control, all the plotted points should have random pattern.)
 There is a close connection between control charts and hypothesis testing. Say, if the
current values of mean plots are between the control limits, then the process mean is in
control that is it is equal to some target value. On the other hand, if mean is out of either
control limits, then the process is out-of-control that is it is not equal to the target value.
Therefore, a point plotting within the control limits is equivalent to failing to reject the
hypothesis of statistical control; otherwise it is equivalent to rejecting the hypothesis of
statistical control.
Some analysts are using two tests of limits on control charts. These limits are found between the
control limits that used to decide the process is in control with warring and called warring limits.
The following figure shows that the theoretical control chart with only the control limits.
Choice of Control Limits

Widening the control limits will increase the risk of a point falling between the control limits
when the process is really out-of-control. Whereas, if we move the control limits closer to the
center line, the opposite effect is obtained, i.e., increase the risk of a point falling beyond the
control limits when the process is really in control. Therefore, how do we choose the limits?
First note that we do not need the center line. But it is common to have it and it is regarded as the
target line (i.e., value we want to obtain). We want choose the control limits so that: P(target
value is outside the control limits/ process is in control) is small. Similarly, if we want have
warring limits we want choose the warring limits so that: P (target value is outside the warring
limits/ process is in control) to be small.

During the 1920's, Dr. Walter A. Shewhart proposed a general model for control charts as
follows. Let w be a sample statistic that measures some continuously varying quality
characteristic of interest and suppose that the mean of w is w, with a standard deviation of w.
Then the center line, the UCL and the LCL are:

UCL = w + k w
Center Line = w
LCL = w - k w , where k is the distance of the control limits from the center line,
expressed in terms of standard deviation units.

Let our target value is X-bar, then we might use the 95%/ and 99% rule, i.e., solve

P(LCL < X-bar < UCL/ process is in control) = 0.99 and P(LWL < X-bar < UWL/ process is in
control) =0.95. If the distribution of X is normal with mean μ and variance σ 2, then the
distribution of X-bar is normal with mean μ and variance σ 2/n. After some mathematical
operations the confidence interval for X-bar with 95% and 99% confident becomes:

σ σ σ σ
P(μ –1.96* < X-bar < μ+1.96* ) = 0.95 & P(μ–2.58* < X-bar < μ+2.58* ) = 0.99.
√n √n √n √n
Hence we choose:
σ
UCL = μ+2.58* ,
√n
σ
UWL = μ+1.96* ,
√n
CL = μ,
σ
LWL = μ-1.96* and
√n
σ
LCL = μ - 2.58* . It is known as European Convention.
√n
Another way of choosing the limits is to use 2-sigma and 3-sigma methods. Therefore, if we use
the 2 and 3-sigma methods the above formula becomes:
σ
UCL = μ+3* ,
√n
σ
UWL = μ+2* ,
√n
CL = μ,
σ
LWL = μ-2* and
√n
σ
LCL = μ - 3* . It is known as Anglosaxon Convention.
√n
Possible Decision Rules

The most common decision rule is say: out-of-control, in control or issue warring. If we do not
use warring limits we can use the rule say out-of-control and in control. The following control
chart shows the decision rules what we considered.
If a data point falls outside the control limits, we assume that the process is probably out of
control and that an investigation is warranted to find and eliminate the cause or causes. Does this
mean that when all points fall within the limits, the process is in control? Not necessarily! If the
plot looks non-random, that is, if the points exhibit some form of systematic behavior, there is
still something wrong. For example, if seven points in a row either above or below the center line
and/or five points in a row either increasing-Run up: a series of observations are increasing and
Run down: a series of observations are decreasing or, we would wish to know why this is so.
Statistical methods to detect sequences or nonrandom patterns can be applied to the
interpretation of control charts. To be sure, "in control" implies that all points are between the
controls limits and they form a random pattern. This type of method is that detecting the
Statistical process control known as Monogram convention. How do we Measure the
Performance of a Chart?
There are two common methods of measuring the performance of a control chart the power
function and average run length.
I. The Power Function
The power function is equal to the probability of the process is out-of-control at sample i
regarded as a function of parameter the mean and variance of i. if the control chart is good the
power function will be small (close to 0) when the process is in control and large (close to 1)
when it is out-of-control.
II. Average Run Length
The time until a process is found to be out-of-control is called a run. This time is a random
variable. Its expectation is called the ARL. Essentially, the ARL is the average number of points
that must be plotted before a point indicates an out-of-control condition. Ideally we want the
ARL to be large when the process is in control and small when it is out-of-control. If the process
observations are uncorrelated, then for any control chart, the ARL can be calculated easily from
ARL= 1/p, where p is probability that any point exceeds the control limits. This equation is can
be used to evaluate the performance of the control chart.
To illustrate, for the chart with three-sigma limits, p = 0.0027 is the probability that a single point
falls outside the limits when the process is in control. Therefore, the average run length of the X
1 1
chart when the process is in control (called ARL0) is: ARL0 = = =370.
p .0027
That is, even if the process remains in control, an out-of-control signal will be generated every
370 sample, on the average.
It is also occasionally convenient to express the performance of the control chart in terms of its
average time to signal (ATS). If samples are taken at fixed intervals of time that are h hours
apart, then ATS = ARLh.
Use of Control Chart
Control chart have had a long history of use in U.S. industries and in many offshore industries as
well. There are at least five reasons for their popularity.
1. Control charts are a proven technique for improving productivity
2. Control charts are effective in defect prevention
3. Control charts prevent unnecessary process adjustment
4. Control charts provide diagnostic information
5. Control charts provide information about process capability.

II.2 Control Chart for Variables

As pointed out above, ‘control charts’ is the technique of process control. A number of control
charts have been developed depending on the ways of assessing the quality of manufactured
product. Control charts for variables are designed to achieved and maintain a satisfactory quality
level for a process whose product is amenable to quantitative measurements like the thickness,
length or diameter of screw or nut, weight of the bolts, tensile strength of yarn or steel pipes,
resistance of a wire, etc. the observation on such units can be expressed in specific unit of
measurements. In such cases the quality control involves the control of variation both in
measures of central tendency and dispersion of the characteristics. Variable control charts are
used when quality is measured as variables (length, weight, tensile strength, etc.). The main
purpose of the variable control charts is to monitor the process mean and the standard deviation.
Control charts for variables are: Control chart for sample mean ( X ), Control charts for sample
range (R), and Control charts for sample standard deviation(s).

II.2.1 Control Chart for X-bar

A typical quality control plan requires sampling one or more items from a production process
periodically, and making the appropriate quality measurements. Usually more than one items are
measured each time to increase accuracy and measure variability. The X chart helps the quality
control person decide whether the center (or average, or the location of central tendency) of the
measurement has shifted.

Suppose that n observations are to be made at each time in which the process is checked. Let Xi
be the ith observation (i = 1,…,n) at the specified time point. Let X j be the average of n
observations at time j. If E(Xi) = μ and Var(Xi) = σ 2 for a process in control, then X i should be
2
σ
approximately normally distributed with E( X i ) = μ and E( X i ) = .
n

Now since X i has approximately a normal distribution, we can find the interval that will have a
probability of 1- α of containing X i . This interval is μ± z α / 2 ( )
σ
√n
. If the mean and variation
of the population (the µ and σ) is known then, we can set the lower and the upper control limits

( )
at μ−z α / 2 σ
√n
and μ+ z
α / 2
σ .
√n ( )
In practice, we usually will not know it μ and σ . Therefore, they must be estimated from
preliminary samples or subgroups taken when the process is thought to be in control. These
estimates should usually be based on at least 20 to 25 samples. Suppose that m samples are
available, each containing n observations on the quality characteristic. Typically, n will be small,
often either 4, 5, or 6. These small sample sizes usually risk from the construction of rational
subgroups and from the fact that the sampling and inspection costs associated with variables
measurements are usually relatively large. Let X 1 , X 2 , ... X m be the average of each sample. Then
the best estimator of it, the process average, is the grand average--say,
X 1+ X 1 +… X m
X́ = ....................................................................2.2
m
Thus, X́ whould used as the center line on the X 1 chart.
 To construct the control limits, we need an estimate of σ . Here we may estimate σ from either
standard deviations or the ranges of the m samples. Possibilities to estimate σ are the following.

√
m
1
i. ∑ S 2, square root of mean of sample variances, where j=1, 2, 3, …, m.
m j=1 j
1
m
Rj Rj
ii. ∑
m j=1 d 2 ,n
, where d2,n is a constant and chosen to make
d 2 ,n
is an unbiased estimate of

σ.
1
m
Sj Sj
iii. ∑
m j=1 c 2, n
, where c2,n is a constant and chosen to make
c 2 ,n
is an unbiased estimate of

σ.

For now, we will use the range method and if X 1, X2, X3, …,Xn is a sample of size n from the m
samples, the R1, R2, R3,….Rm be the range of the m samples and R is the mean of the range. We
may now give the formula for constructing the control limits on the X-bar chart as follows.

R̄
Though { R̄ has a relationship with σ, it is not an unbiased estimator. However, the ratio is! Here, d2,n ¿is obtained from the tables. Thus, the mew control limits becomes:¿¿
d2,n
R̄ 3
X̄¯ ∓ z α /2 or { X̄¯ ∓ A2 R̄ , where A 2 = using 3-sigma method .¿
d2 , n √ n d 2, n √ n
 II.2.2 Control Chart for R

Deciding whether the center of the distribution of quality measurements has shifted up or down
may not be enough. It is frequently of interest to decide if the variability of the process
measurements has significantly increased or decreased. A process that suddenly starts turning out
highly variable products could cause severe problems in the operations. Process variability may
be monitored by plotting values of the sample, R or the sample standard deviation, S on a control
chart. Because the control limits on the X-bar chart depend on the process variability, unless
process variability is in control, these limits will not have much meaningless. Since we already
have established the fact that there is a relationship between the ranges (Rj) and σ, it would be
natural to base our control chart for variability on the Rj’s. A control chart for variability could be
based on the adjusted standard deviation also (Sj), but use of ranges provides nearly as much
accuracy for much less computation. Using the same argument as control chart for central
tendency (x-bar), one could construct that the center line and control limits for R chart as
follows.
Now E ( R j ) =d 2 , n σ ; Var ( R j )=d32 , n σ 2 ; where d 2, n and d 3, n are found in standard tables .
The confidence interval, then, is: d 2, n σ±z α / 2 d3 , n σ =σ ( d 2, n ±z α / 2 d 3 , n ) .

As in the previous chart (X-bar chart), if σ is not specified, it must be estimated from the data.
The best estimator of σ based on the range is R̄/d , and the estimator of the control limits then
2 ,n
becomes:

( )
R̄ d
( d 2, n ±z α /2 d 3 , n ) ; or { R̄ 1±z α /2 3 , n once again, if we use the 3-sigma method we have z α /2 =3 . ¿
d 2, n d2 , n

( ) ( )
d3 , n d
Letting, 1−3 =D 3 ; and 1+3 3, n =D4 , the control limit becomes ( R̄ D3 , R̄ D4 ) . Values of D3 and
d2 , n d 2, n
D4 are found in statistical tables.

Note that the lower control limit for R-chart can give a negative value but R could not be
negative so this does not make sense. Therefore, we replace the lower control limit by zero if we
get a negative value.
Examples

1. A machine is set to deliver packets of a given weight. 10 sample of size 5 each were
recorded. Below a given relevant data:
Sample Number 1 2 3 4 5 6 7 8 9 10
Mean ( ) 15 17 15 18 17 14 18 15 17 16
Range (R) 7 7 4 9 8 7 12 4 11 5
Calculate the central line and the control limits for X-bar and R-charts and also draw the
charts and then comment the state of control.
2. You are given the values of sample means ( ) and the sample range ( R) ten samples of
five size of each. Draw the X-bar and R-charts and comment on the state of control of the
process.
Sample Number 1 2 3 4 5 6 7 8 9 10
43 49 37 44 45 37 51 46 43 47
R 5 6 5 7 7 4 8 6 4 6
3. Consider the Glass container strength data (Psi) and construct the X-bar and R-
charts and then comment the statistical state of control.
 II.2.2.1 Changing Sample Size on the X-bar and R-charts
We have presented the development of X and R charts assuming that the sample size n is
constant from sample to sample. However, there are situations in which the sample size n is not
constant one situation is that of variable sample size; that is, each sample may consist of a
different number of observations. The X and R charts are generally not used in this case because
they lead to a changing center line on the R chart, which is difficult to interpret for many users.
Another situation is that of making a permanent (or semi-permanent) change in the sample size
because of cost or because the process has exhibited good stability and fewer resources are being
allocated for process monitoring. In this case it is easy to recompute the new control limits
directly from the old ones without collecting additional samples based on the new sample size.
Let
Rold = average range for the old sample size
Rnew = average range for the new sample size
n old = old sample size
n old = new sample size
d1(old) = factor d2 for the old sample size
d2(new) = factor d2 for the new sample size
For the X chart the new control limits are

UCL = X́ + A2 [ ] d 2(new)
R
d 2 (old) old

LCL = X́ - A2
[ ]
d 2(new)
d 2 (old)
R old where the center line is unchanged and the factor As is selected for

the new sample size. For the R chart, the new parameters are:

UCL = D2 [ ]
d 2(new)
R
d 2 (old) old

CL = Rnew = [ ]
d2 (new)
d 2(old)
R old

LCL = max{0, D2
[ ] d 2(new)
d 2 (old)
R old } ,where D3 and D4 are selected for the new sample size.

Exercise!: consider the X-bar and R-charts developed in the above examples. These charts were
based on a sample of five in size. Suppose that since the process exhibits good control, the
process personnel want to reduce the sample size to three, then reconstruct the X-bar and R-
charts.

II.2.2.2 Interpreting the X-bar and R-charts

We have noted previously that a control chart can indicate an out-of-control condition even
though no single point plots outside the control limits, if the pattern of the plotted points exhibits
nonrandom or systematic behavior. In many cases, the pattern of the plotted points will provide
useful diagnostic information on the process, and this information can be used to make process
modifications that reduce variability (the goal of statistical process control).
To effectively interpret X-bar and R-charts, the analyst must be familiar with both the statistical
principles underlying the control chart and the process itself. General rules to determine whether
a process is in control:
1. No points outside the control limits.

2. The number of points above and below the center line are about the same.

3. Points seem to fall randomly above and below the center line.

4. Most points are near center line, and only a few are close to control limits.

One point outside control limits: measurement or calculation error, power surge, a broken tool,
incomplete operation.

Sudden shift in process average: new operator or inspector, new machine setting.

Cycles: operator rotation or fatigue at the end of shift, different gauges, seasonal effects such as
temperature and humidity.

Trends: x-bar-chart – learning effect, dirt or chip buildup, tool wear, aging of equipment; R-
chart (increasing trend) – gradual decline in material quality; R-chart (decreasing trend) –
improved skills, better materials.

Hugging the center line: sample taken over various machines canceling out the variation within
the sample.

Hugging the control limits: sample taken over various machines not canceling out the variation
within the sample.

Instability: difficult to identify causes. Typically, over-adjustment of machine.

Note that, in interpreting patterns on the X-bar and R-charts, one should consider the two charts
jointly. If X is normally distributed with mean μ and σ 2 , then X-bar and R are statistically
independent. If there is correlation between the X-bar and R values that is, if the points on the
two charts “follow” each other then this indicates that the underlying distribution is skewed.

II.2.3 Control Chart for S

Since standard deviation is an ideal measure of dispersion, a combinations of control chart for X-
bar and S-charts, theoretically more appropriate than the combination of and R charts for
controlling the process average and process variability. The construction of S-chart, if σ 2 is
unknown variance of a probability distribution, then an unbiased estimator of σ 2 is the sample
n
variance: S2 = ∑ ¿¿ ¿ . However, the sample standard deviation, S, is a biased estimator of σ . If
i=1

the distribution is normal, then S actually estimates C4σ , where

is a constant that depends on the sample size n and also a tabulated

value in most text books. Furthermore, sd(S) = σ √ 1−C 24 and this information can be used to
establish control chart for S. let σ is known, then E(S) = C4σ and sd(S) = σ √ 1−C , therefore,
2
4

the center line and the control limits of the S-chart are as follows.

UCL = C4σ + 3σ √ 1−C 24

CL = C4σ

LCL = C4σ - 3σ √ 1−C 24. Let we define B5 = C4 - 3 √ 1−C 24 and B6 = C4 + 3 √ 1−C 24, then

UCL = B6σ , CL = C4σ and LCL = B5σ .

In practice, σ is unknown, then it must be estimated by analyzing past data. Suppose that m
preliminarily samples are available, each of size n, and let S j be the standard deviation of the j th
m

sample. Then S =
∑ Sj. We know that the statistic
S
is an unbiased estimator of σ .
j =1 C4
m
Therefore, the parameters of the S-chart would be:
 3 3
now we let assign B3 = 1 - C4
√ 1−C 2
4
and B4 = 1 + C4
√ 1−C , then
2
4

UCL = B4 S, CL = S and LCL = B3 S.

The other steps in the construction of s-chart are same as the case of R-chart. The sample number
is taken along horizontal scale and the statistic standard deviation is taken along the vertical
scale. The sample standard deviation s1, s2, s3….sk are plotted as points against the corresponding
sample numbers. The central line is drawn at S (grand) and UCLs and LCLs are plotted as
control limits at the computed values given.
Similarly, the parameters of the chart would be

3
If we assign here also A3 = C , then UCL = X́ + A3 S, CL = X́ and LCL = X́ - A3 S. , the
4 √n

"grand" mean is the average of all the observations and A3, B3 and B4 are tabulated values.
Remark: S-charts are interpreted the same as R-charts. Theoretically s-chart appears to be more
appropriate than R-charts for controlling dispersion of the process. But the difficulty of
computation makes the use of s-chart almost impractical in most industry work. In practice, R-
chart is preferred to s-chart because of its computational case.
Example: consider the following data and construct the X-bar and S-charts and then comment on
the statistical sate of control.
The X-bar and S-charts with Variable Sample Size
The X and s control charts are relatively easy to apply in cases where the sample sizes are
variable. In this case, we should use a weighted average approach in calculating X́ and S. If ni is
m

∑ ni X i
i=1
th
the number of observations in the i sample, then use: X́ = m and S = ¿ ¿as the center fines
∑ ni
i=1

on the X and s control charts, respectively. The control limits would be calculated from the
above formula, but the constants A 3, B3 and B4 will depend on the sample size used in each
individual subgroup.
Note that if the nj are not very different, the constants A 3, B3 and B4 are based on an average
sample size of n.
Exercise!: Consider the following data and construct the X-bar and S-charts and also
comment for the statistical sate of control.
 II.2.4 Control Chart for S2
Most quality engineers use either the R chart or the s chart to monitor process variability, with s
preferable to R for moderate to large sample sizes. Some practitioners recommend a control chart
based directly on the sample variance s2. The parameters for the s 2 control chart are:
2
S 2
UCL = χα
n−1 2
CL = S2
2
S 2
LCL = χ 1−α ,
n−1 2
2 2 α
where χ α and χ 1−α , n-1denote the upper and lower percentage points of the chi-square
2 2 2
distribution with n - 1 degrees of freedom, and S2 is an average sample variance obtained from
the analysis of preliminary data.
n
Note that if S2 = ∑ ¿¿ ¿ is the sample variance of a random sample from a normal distribution
i=1

2
2 (n−1) S
with mean μ and variance σ , then has a chi-square distribution with n-1, that is,
σ2
2
(n−1) S χ 2α
,(n−1).
σ2 2

Exercise! Construct the S2 –chart by using one of the above example!

II.2.5 The Shewhart Control Chart for Individual Measurement
What if you could not get a sample size greater than 1 (n =1)?
There are many situations in which the sample size used for process monitoring is n = 1; that is,
the sample consists of an individual unit. Some examples of these situations are as follows:
Automated inspection and measurement technology is used, and every unit manufactured is
analyzed.
The production rate is very slow, and it is inconvenient to allow samples sizes of N > 1 to
accumulate before analysis

Repeat measurements on the process differ only because of laboratory or analysis error, as in
many chemical processes.

The individual control charts are useful for samples of sizes n = 1.

The moving range (MR) is defined as the absolute difference between two successive
observations: MRi = |xi - xi-1|, which will indicate possible shifts or changes in the process from
one observation to the next.

Interpretation of the Charts.

• X Charts can be interpreted similar to X-bar charts. MR charts cannot be interpreted the
same as X-bar or R charts.

• Since the MR chart plots data that are “correlated” with one another, then looking for
patterns on the chart does not make sense.

• MR chart cannot really supply useful information about process variability.

• More emphasis should be placed on interpretation of the X chart.

Summary

It is better to summarize in one place the various computational formulas for the major types
of variables control charts in this section. The formula for X-bar, R, S and S 2-charts are given
in the following table.
Formulas for control charts when σ is known
Chart Center Line Control Limits
X ( μ∧σ known) μ μ ∓ Aσ
R(σ known) d2,n σ UCL = D2σ , LCL = D1σ
S (σ konwn ) C4 σ UCL = B6σ , LCL = B5σ
Formulas for control charts when σ is unknown
X (using R ) X́ X́ ∓ A2 R
R R UCL=D4 R ,LCL=D3 R
S S UCL=B4 S, LCL = B3 S
X (using S ) X́ X́ ∓ A3 S
S2 S
2
S
2
2
UCL= χα , LCL =
n−1 2 , ,(n−1)
2
S 2
χ 1−α
n−1 2 , ,(n−1)
X (Using moving Range) X 3
X∓ MR
d 2 ,n
MR MR UCL=D4 MR , LCL=D3 MR
 Tabulated values for constants

n A2 d2 Dl D2 D3 D4
2 1.88 1.128 0 3.686 0 3.268
3 1.023 1.693 0 4.358 0 2.574
4 0.729 2.059 0 4.698 0 2.282
5 0.577 2.326 0 4.918 0 2.114
6 0.483 2.534 0 5.078 0 2.004
7 0.419 2.704 0.204 5.204 0.076 1.924
8 0.373 2.847 0.388 5.306 0.136 1.864
9 0.337 2.970 0.547 5.393 0.184 1.816
10 0.308 3.078 0.687 5.469 0.223 1.777
11 0.285 3.173 0.811 5.535 0.256 1.744
12 0.266 3.258 0.922 5.594 0.284 1.717
13 0.249 3.336 1.025 5.647 0.308 1.692
 II.3 Control Charts for Attributes

Introduction

Many quality characteristics cannot be conveniently represented numerically. In such cases, we

usually classify each item inspected as either conforming or nonconforming to the specifications
on that quality characteristic. The terminology "defective" or "non defective" is often used to
identify these two classifications of product. More recently, the terminology "conforming" and
"nonconforming" has become popular. Quality characteristics of this type are called attributes.

Attributes charts are generally not as informative as variable charts because there is typically
more information in a numerical measurement than in merely classifying a unit as conforming or
nonconforming. However, attribute charts do have important applications. They are particularly
useful in service industries and in no remanufacturing quality improvement efforts because so
many of the quality characteristics found in these environments are not easily measured on a
numerical scale.

2.3.1 The Control Chart for Fraction Nonconforming (Control Chart for P)

The fraction nonconforming is defined as the ratio of the number of nonconforming items in a
population to the total number of items in that population. The items may have several quality
characteristics that are examined simultaneously by the inspector. If the item does not conform to
standard our one or more of these characteristics, it is classified as nonconforming.

The statistical principles underlying the control chart for fraction nonconforming are based on
the binomial distribution. Suppose the production process is operating in a stable manner, such
that the probability that any unit will not conform to specifications is p, and that successive units
produced are independent. Then each unit produced is a realization of a Bernoulli random
variable with parameter p. If a random sample of n units of product is selected, and if D is the
number of units of product that are nonconforming, then D has a binomial distribution with

parameters n and p; that is, P{D = x} = ( nx) p (1-p¿

x
n −x
, x = 0, 1, 2, ….n. The mean and variance

of the random variable D are np and np(1 -p), respectively. The sample fraction nonconforming
is defined as the ratio of the number of nonconforming units in the sample D to the sample size
 D
n; that is, ^p = . The distribution of the random variable ^p can be obtained from the binomial.
n
p (1− p)
Furthermore, E( ^p ¿ = μ = p and Var( ^p ¿ = σ 2^p = . This background is sufficient to
n
develop the control chart for proportion or fraction nonconforming. The chart is called the p-
chart.

Development and Operation of the Control Chart

If the true fraction conforming p is known (or a standard value is given), then the center line and
control limits of the fraction nonconforming control chart are:

When the process fraction (proportion) p is not known, it must be estimated from the available
data. This is accomplished by selecting m preliminary samples, each of size n. If there are Di
defectives in sample i, the fraction nonconforming in sample i is

And the average of these individuals sample fractions is

The statistic P estimates the unknown fraction nonconforming p. The center line and control
limits of the control chart for fraction nonconforming are computed as follows:
Fraction Nonconforming Control Chart: No Standard Given

UCL = P+3
√ P(1−P)
n
CL = P

LCL = P−3
√ P (1−P)
n
Generally a control chart for fraction of defectives is used when sample unit as a whole classified
as goods or bads, i.e. defective or non-defective. Construction of p chart is as in the case of
and R charts we take the sample number along the horizontal scale and the statistic ‘p’ along the
vertical scale. The sample fraction defective p 1, p2, p3…pk are plotted against the corresponding
sample numbers as points (dots).
Remarks:
1. Since p cannot be negative, if LCLp computed from the above formula comes out to be
negative then it is taken as zero.
2 Interpretation of p charts: p charts interpreted similarly as an or R charts. If all the sample
points fall within the control limits, the process is termed to be statistically in control. If one or
more of the points go out sides the control limits, it is said to be statistically out-of-control. If the
point goes above the UCLp, it reflects the lack of the statistical control since it has changed for
the worse. Such a point, known as high spot, indicates deterioration in the lot quality. Reasons
for this should be traced and eliminated. If the point goes below the LCLp, it again exhibits lack
of control as the process has changed for the better. Such a point, known as low spot, indicates
improvement in the lot quality. Reasoning for this should be looked for and incorporated in to
the process if possible.
Example:

Frozen orange juice concentrate is packed in 6-oz cardboard cans. These cans/containers are
formed on a machine by spinning them from cardboard stock and attaching a metal bottom panel.
By inspection of a can, we may determine whether, when filled, it could possibly leak either on
the side seam or around the bottom joint. Such a nonconforming can has an improper scat on
either the side seam or the bottom panel. We wish to set up a control chart to improve the
fraction of nonconforming cans produced by tiffs machine.
To establish the control chart, 30 samples of n = 50 cans each were selected at half-hour intervals
over a three-shift period in which the machine was in continuous operation. The results are given
in the following table.
Solution: since the 30 samples of n=50 cans each were selected at half-hour intervals over a
three-period in which the machine was in continuous operation.

Therefore,
2.3.2 The np Control Chart

It is also possible to base a control chart on the number nonconforming rather than the fraction
nonconforming. This is often called an np control chart.
If the sample size is constant for all samples, say n, then the sampling distribution of the statistic,
d = np number of nonconforming/defective in the sample is given by,

E (d) =np

S.E (d) = √ np(1−p )

Hence, the 3-sigma method control limits for np charts are given by

E (d) ± sd(d) = np ±3 √ np(1−p)

If ^p is the known or specified values of p then

UCL = n ^p +3 √ np(1−p)
CL = n ^p

LCL = n ^p -3 √ np(1−p)
If p is not known, then its unbiased estimate provided by (grand) given in previous charts i.e. p-
charts.
 UCL= n +3 √ P̄(1− P̄ )
CL= n

LCL= n -3 √ P̄(1− P̄ )
Remarks: if the sample size is varies from sample to sample, then np chart would be quite
uncomfortable to use because the central lines as well as the control limits would vary from
sample to sample. In such a case p chart would be better to use. However, in case of constant
sample size for all samples any one of np or p charts may be used but, in practice, p chart
commonly used. In the other way, many non-statistically trained personnel find the np chart
easier to interpret than the usual fraction nonconforming control chart.
Example: consider the data in the example above for the fraction nonconforming Orange Juice
concentrate cans.

Solution: For the data = 0.2313 and n = 50.

2.3.3 Variable Sample Size

In some applications of the control chart for fraction nonconforming, the sample is a 100%
inspection of process output over some period of time. Since different numbers of units could be
produced in each period, the control chart would then have a variable sample size. There are
different approaches to constructing and operating a control chart with a variable sample size.

A. Variable-Width Control Limits

The first and perhaps the most simple approach is to determine control limits for each individual
sample that are based on the specific sample size. That is, if the i th sample is of size n i, then the

upper and lower control limits are: p ∓ 3

√ p(1− p)
ni
. Note that the width of the control limits is

inversely proportional to the square root of the sample size. To illustrate this approach, consider
the data given below for the 25 samples.
Solution:

B. Control Limits Based on an Average Sample Size

The second approach is to base the control chart on an average sample size, resulting in an
approximate set of control limits. This assumes that future sample sizes will not differ greatly
from those previously observed. If this approach is used, the control limits will be constant, and
the resulting control chart will not look as formidable to operating personnel as the control chart
with variable limits.
Example: For the data in the above example, construct control limits based on an average sample
size.
Solution:
Thus, the resulting control chart is given below.

C. The Standardized Control Chart

The third approach to dealing with variable sample size is to use a "standardized" control chart,
where the points are plotted in standard deviation units. Such a control chart has the center 1ine
at zero, and upper and lower control limits of+3 and -3, respectively. The variable plotted on the
^pi− p
chart is Zi =
√ p(1− p) , but if p is unknown we can estimate it by the process average fraction
ni
^pi− p
nonconforming in the in control state, i.e., Zi =
√ p(1− p) .
ni
Example: construct the standardized control chart for the above example.
2.3.4 Control Charts for Number of Nonconformities (Defects) per unit
(Control Chart for c-Chart)
A nonconforming item is a unit of product that does not satisfy one or more of the specifications
for that product. Each specific point at which a specification is not satisfied results in a defect or
nonconformity. Consequently, a nonconforming item will contain at least one nonconformity.
However, depending on their nature and severity, it is quite possible for a unit to contain several
nonconformities and not be classified as nonconforming. As an example, suppose we are
manufacturing personal computers. Each unit could have one or more very minor flaws/errors in
the cabinet finish and since these flaws do not seriously affect the unit's functional operation, it
could be classified as conforming. However, if there are too many of these flaws, the personal
computer should be classified as nonconforming, since the flaws would be very noticeable to the
customer and might affect the sale of the unit. There are many practical situations in which we
prefer to work directly with the number of defects or nonconformities rather than the fraction
nonconforming.
It is possible to develop control Charts for either the total number of nonconformities in a .unit or
the average number of nonconformities per unit. These control charts usually assume that the
occurrence of nonconformities in samples of constant size is well modeled by the Poisson
distribution. Essentially, this requires that the number of opportunities or potential locations for
nonconformities be infinitely large and that the probability of occurrence of nonconformity at
any location be small and constant. Furthermore, the inspection unit must be the same for each
sample. That is, each inspection unit must always represent an identical area of opportunity for
the occurrence of nonconformities. In addition, we can count nonconformities of several
different types on one unit, as long as the above conditions are satisfied for each class of
nonconformity.
In most practical situations, these conditions will not be satisfied exactly. The number of
opportunities for the occurrence of nonconformities may be finite, or the probability of
occurrence of nonconformities may not be constant. As long as these departures from the
assumptions are not severe, the Poisson model will usually work reasonably well.
Procedures with Constant Sample Size
The c chart is used when we count the number of defects per unit rather than classifying a unit as
good or bad, i.e. defective or non-defective. In a number of manufacturing processes we come
across situations where:
(i) The opportunities for the occurrence of a defect in any unit are very large.
(ii) The probability of occurrences of a defect in any spot is very small.
Consider the occurrence of nonconformities in an inspection unit of product. In most cases, the
inspection unit will be a single unit of product, although this is not necessarily always so. The
inspection unit is simply an entity for which it is convenient to keep records. It could be a group
of 5 units of product, 10 units of product, and so on. Suppose that defects or nonconformities
−c x
e c
occur in tiffs inspection unit according to the Poisson distribution; that is, p(x)= x = 0,
x!
1,2 ...., where x is the number of nonconformities and c > 0 is the parameter of the Poisson
distribution. Furthermore, we recall that both the mean and variance of the Poisson distribution
are the parameter c. Therefore, a control chart for nonconformities with three-sigma limits would
be defined as: UCL = c + 3√ c , CL = c and LCL = c - 3√ c . Objective of this procedure is to
monitor the number of defects per item and to detect situations where this variable is out of
control.
If no standard is given, then c may be estimated as the observed average number of
nonconformities in a preliminary sample of inspection units--say, c . In this case, the control chart
has parameters defined as: UCL = c + 3√ c , CL = c and LCL = c - 3√ c . The control chart for
nonconformities is also sometimes called the c chart.
Example: Consider the number of nonconformities observed in 26 successive samples of 100
printed circuit boards data and construct the control chart for c. Note that, for reasons of
convenience, the inspection unit is defined as 100 boards.
Solution:
total number of nonconformities 516
We estimate c by: c = = 26 = 19.58
number of sample
Then the control limits are given by:

The control chart is shown above and two points plot outside the control limits, samples 6 and
20. Investigation of sample 6 revealed that a new inspector had examined the boards in this
sample and that he did not recognize several of the types of nonconformities that could have
been present. Furthermore, the unusually large number of nonconformities in sample 20 resulted
from a temperature control problem in the wave soldering machine, which was subsequently
repaired. Therefore, it seems reasonable to exclude these two samples and revise the trial control
472
limits. The estimate of c is now computed as: c = = 19.78 and the revised control chart
24
limits are:
These become the standard values against which production in the next period can be compared.
Note that, with complex products such as automobiles, computers, or major appliances, we
usually find that many different types of nonconformities or defects can occur. Not all of these
types of defects are equally important. A unit of product having one very serious defect would
probably be classified as nonconforming to requirements, but a unit having several minor defects
might not necessarily be nonconforming. In such situations, we need a method to classify
nonconformities or defects according to severity and to weight the various types of defects in a
reasonable manner Demerit schemes can be of value in these situations. One possible demerit
scheme is defined as follows.
Class A Defects-Very Serious: The unit is either completely unfit for service, or will fail in
service in such a manner that cannot be easily corrected in the field, or will cause persona/injury
or property damage.
Class B Defects-Serious: The unit will possibly suffer a Class A operating failure, or will
certainly cause somewhat less serious operating problems, or will certainly have reduced life or
increased maintenance cost.
Class C Defects-Moderately Serious: The unit will possibly fail in service, or cause trouble that
is less serious than operating failure, or possibly have reduced life or increased maintenance
costs, or have a major defect in finish appearance, Or quality of work.
Class D Defects-Minor: The unit will not fail in service but has minor defects in finish,
appearance, or quality of work.
Guidelines for Implementing Control Charts
Almost any process will benefit from SPC, including the use of control charts. In this section, we
present some general guidelines helpful in implementing control charts. Specifically, we deal
with the following:
1. Determining which process characteristics to control
2. Determining where the charts should be implemented in the process
3. Choosing the-proper type of control charts
4. Taking actions to improve processes as the result of SPC/control chart analysis
5. Selecting data-collection systems and computer software
The guidelines are applicable to both variables and attributes control charts. Remember, control
charts are not only for process surveillance; they should be used as an active, on-line method for
reduction of process variability.
The u-Chart
A control chart for nonconformities is constructed with the sample size exactly equal to one
inspection unit. The inspection unit is chosen for operational or data- collection simplicity.
However, there is no reason why the sample size must be restricted to one inspection unit. In
fact, we would often prefer to use several inspection units in the sample, thereby increasing the
area of opportunity for the occurrence of nonconforminities.
The sample size should be chosen according to statistical considerations, such as specifying a
sample size large enough to ensure a positive lower control limit or to obtain a particular
probability of detecting a process shift. Alternatively, economic factors could enter into sample-
size determination. There are two general approaches to constructing the revised chart once a
new sample size has been selected. One approach is simply to redefine a new inspection unit that
is equal to n times the old inspection unit. The second approach involves setting up a control
chart based on the average, number of nonconformities per inspection unit. If we find x total
nonconformities in a sample of n inspection units, then the average number of nonconformities
per inspection unit is u = x/n, where x is a Poisson random variable, consequently, the
parameters of the control chart for the average number of nonconformities per unit are as:
UCL = u + 3√ u/n, CL = u and LCL = u - 3√ u.
Determining which Characteristics to Control and where to Put the Control Charts
At the start of a control chart program, it is usually difficult to determine which product or
process characteristics should be controlled and at which points in the process to apply control
charts. Some useful guidelines follow.
1. At the beginning of a control chart program, control charts should be applied to any
product characteristics or manufacturing operations believed to be important. The charts
will provide immediate feedback as to whether they are actually heeded.
2. The control charts found to be unnecessary should be removed, and others that
engineering and operator judgment indicates may be required should be added. More
control charts will usually be employed at the beginning than after the process has
stabilized.
 3. Information on the number and types of control charts on the process should be kept
current. It is best to keep separate records on the variables and attributes charts. In
general, after the control charts are first installed, we often find that the number of control
charts tends to increase rather steadily. After that, it will usually decrease: When the
process stabilizes, we typically find that it has the same number of charts from one year
to the next. However, they are not necessarily the same charts,
4. If control charts are being used effectively and if new knowledge is being gained about
the key process variables, we should find that the number of X and R charts increases and
the number of attributes control charts decreases.
5. At the beginning of a control chart program there will usually be more attributes control
charts, applied to semi finished or finished units near the end of the manufacturing
process. As we learn more about the process, these charts will be replaced with X and R
charts applied earlier in the process to the critical parameters and operations that result in
nonconformities in the finished product. Generally, the earlier that process control can be
established, the better. In a complex assembly process, this may imply that process
controls need to be implemented at the vendor or supplier level.
6. Control charts are an on-line, process-monitoring procedure. They should be
implemented and maintained as close to the work center as possible, so that feed- back
will be rapid. Furthermore, the process operators and process engineering should have
direct responsibility for collecting the process data, maintaining the charts, and
interpreting the results. The operators and engineers have the detailed knowledge of the
process required to correct process upsets and use the control chart to improve process
performance. Microcomputers can speed up the feedback and should be an integral part
of any modem, on-line, process-control procedure.
7. The out-of-control-action plan (OCAP) is a vital pan of the control chart. Operating and
engineering personnel should strive to keep OCAPs up-to-date and valid.
Choosing the Proper Type of Control Chart
A. X and R (or X and S) charts: Consider using variables control charts in these situations:
1. A new process is coming on stream, or a new product is being manufactured by an
existing process.
 2. The process has been in operation for some time, but it is chronically in trouble or
unable to hold the specified tolerances.
3. The process is in trouble, and the control chart can be useful for diagnostic purposes
(troubleshooting).
4. Destructive testing (or other expensive testing procedures) is required.
5. It is desirable to reduce acceptance-sampling or other do wrist ream testing to a
minimum when the process can be operated in control.
6. Attributes control charts have been used, but the process is either out of control or in
control but the yield is unacceptable.
7. There are very tight specifications, overlapping assembly tolerances, or other difficult
manufacturing problems.
8. The operator must decide whether or not to adjust the process, or when a setup must be
evaluated.
9. A change in product specifications is desired.
10. Process stability and capability must be continually demonstrated, such as in regulated
industries.
B. Attributes Charts (p charts and c charts): Consider using attributes control charm in these
situations:
1. Operators control the assignable causes, and it is necessary to reduce process fail out.
2. The process is a complex assembly operation and product quality is measured in terms of
the occurrence of nonconformities, successful or unsuccessful product function, and so
forth.
3. Process control is necessary, but measurement data cannot be obtained.
4. Attributes control charts, such as p charts and c charts are very effective for summarizing
information about the process for management review.
5. Remember that attributes charts are generally inferior to charts for variables. Always use
X and R or X and s charts whenever possible.
C. Control Charts for Individuals: Consider using the control chart for individuals in
conjunction with a moving-range chart in these situations:
1. It is inconvenient or impossible to obtain more than one measurement per sample, or
repeat measurements will only differ by laboratory or analysis error.
2. Automated testing and inspection technology allow measurement of every unit produced.
3. The data become available very slowly, and waiting for a larger sample will be
impractical or make the control procedure too slow to react to problems.
2.4 Process and Measurement System Capability Analysis

Statistical techniques can be helpful throughout the product cycle, including development
activities prior to manufacturing, in quantifying process variability, in analyzing this variability
relative to product requirements or specifications, and in assisting development and
manufacturing in eliminating or greatly reducing this variability. This general activity is called
process capability analysis.

Process capability refers to the uniformity of the process. Obviously, the variability of critical-
to- quality characteristics in the process is a measure of the uniformity of output. There are two
ways to think of this variability:

1. The natural or inherent variability in a critical-to-quality characteristic at a specified time;

that is, "instantaneous" variability
2. The variability in a critical-to-quality characteristic over time

We are often required to compare the output of a stable process with the process specifications
and make a statement about how well the process meets specification. To do this we compare
the natural variability of a stable process with the process specification limits. A capable process
is one where almost all the measurements fall inside the specification limits. It is customary to
take the 6-sigma spread in the distribution of the product quality characteristic as a measure of
process capability. The following figure shows a process for which the quality characteristic has
a normal distribution with mean μ and standard deviation σ . The upper and lower natural
tolerance limits of the process fall at μ + 3σ and μ - 3 σ , respectively; that is,

UNTL = μ + 3σ

LNTL = μ - 3 σ
For a normal distribution, the natural tolerance limits include 99.73% of the variable, or put
another way, only 0.27% of the process output will fall outside the natural tolerance limits. Two
points should be remembered:

1. 0.27% outside the natural tolerances sounds small, but this corresponds to 2700
nonconforming parts per million.
2. If the distribution of process output is non-normal, then the percentage of output falling
outside μ ± 3σ may differ considerably from 0.27%.

A process capability study usually measures functional parameters or critical-to- quality

characteristics on the product, not the process itself. When the analyst can directly observe the
process and can control or monitor the data-collection activity, the study is a true process
capability study, because by controlling the data collection and knowing the time sequence of the
data, inferences can be made about the stability of the process over time. However, when we
have available only sample units of product, perhaps obtained from the supplier, and there is no
direct observation of the process or time history of production, then the study is more properly
called product characterization. In a product characterization study we can only estimate the
distribution of the product quality characteristic or the process yield (fraction conforming to
specifications); we can say nothing about the dynamic behavior of the process or its state of
statistical control.

Process capability analysis is a vital part of an overall quality-improvement program. Among the
major uses of data from a process capability analysis are the following:

1. Predicting how well the process will hold the tolerances

2. Assisting product developers/designers in selecting or modifying a process
3. Assisting in establishing an interval between sampling for process monitoring
4. Specifying performance requirements for new
5. Selecting between competing suppliers and other aspects of supply chain management
6. Planning the Sequence of production processes when there is an interactive effect of
processes on tolerances
7. Reducing the variability in a manufacturing process

Thus, process capability analysis is a technique that has application in many segments of the
product cycle, including product and process design, supply chain management, production or
manufacturing planning, and manufacturing. Note that, the estimate of process capability may be
in the form of a probability distribution having a specified shape, center (mean), and spread
(standard deviation). Therefore, some reasons for poor process capability may poor process
centering and/or excess process variability.

The primary techniques are used in process capability analysis: histograms or probability plots,
Process capability ratios, control charts, and designed experiments and we will consider in
this section only some.

2.4.1 Process Capability Analysis Using Histogram

The histogram can be helpful in estimating process capability. Alternatively, a stem-and- leaf
plot may be substituted for the histogram. At least 100 or more observations should be available
for the histogram (or the stem-and-leaf plot) to be moderately stable so that a reasonably reliable
estimate of process capability may be obtained. The histogram, along with the sample average X-
bar and sample standard deviation s, provides information about process capability.

To illustrate the use of a histogram to estimate process capability, consider the following
histogram, which presents the bursting strength (psi) of 100 glass containers.
Analysis of the 100 observations gives: X = 264.06 and s = 32.02. Consequently, the process
capability would be estimated as: X ± 3s or 264.06 + 3(32.02) ≈ 264 + 96 psi.

Furthermore, the shape of the histogram implies that the distribution of bursting strength is
approximately normal. Thus, we can estimate that approximately 99.73% of the bottles
manufactured by this process will burst between 168 and 360 psi. Note that we can estimate
process capability independent of the specifications on bursting strength.

An advantage of using the histogram to estimate process capability is that it gives an immediate,
visual impression of process performance. It may also immediately show the reason for poor
process performance.

2.4.2 Process Capability Analysis Using Probability Plotting

Probability plotting is an alternative to the histogram that can be used to determine the shape,
center, and spread of the distribution. It has the advantage that it is unnecessary to divide the
range of the variable into class intervals, and it often produces reasonable results for moderately
small samples (which the histogram will not). Generally, a probability plot is a graph of the
ranked data versus the sample cumulative frequency on special paper with a vertical scale chosen
so that the cumulative distribution of the assumed type is a straight line. The normal probability
plot can also be used to estimate process yields and fallouts. Care should be exercised in using
probability plots. If the data do not come from the assumed distribution, inferences about process
capability drawn from the plot may be seriously in error.

To illustrate the use of a normal probability plot in a process capability study, consider following
20 observations on glass container bursting strength: 197, 200, 215, 221, 231, 242, 245, 258,
265, 265, 271,275, 277, 278, 280, 283, 290, 301, 318, and 346.

The following figure is the normal probability plot of 20 observations of burst-strength data.
Note that the data lie nearly along a straight line, implying that the distribution of bursting
strength is normal.
We know that the mean of the normal distribution is the fiftieth percentile, which we estimate
from the above probability plot as approximately 265 psi, and the standard deviation of the
distribution is the slope of the straight line. It is convenient to estimate the standard deviation as
the difference between the eighty-fourth and the fiftieth percentiles. For the strength data shown
above and using the figure, we find that σ^ = 84th percentile - 50th percentile =298 -265 psi=33 psi.
Note that ^μ = 965 psi and σ^ = 33 psi are not far from the sample average X = 264.06 and
standard deviation s = 32.02.

2.4.3 Process-Capability Ratios using Cp

It is frequently convenient to have a simple, quantitative way to express process capability. One
way to do so is through the process capability ratio (PCR) Cp in which computed as:

USL−LSL
Cp = , where USL and LSL are the upper and lower specification limits, respectively.
6σ

In a practical application, the process standard deviation σ is almost always unknown and must
be replaced by an estimate σ . To estimate σ we typically use either the sample standard
R
deviation, s or ^ p = USL−LSL .
. This results in an estimate of Cp, say, C
d 2 ,n 6 σ^
Example: consider the burst-strength data and estimate C p. For a certain process the USL = 20
and the LSL = 6. The observed process average, = 16, and the standard deviation, s = 2. From

^ p = USL−LSL =
20−6
this we obtain: C = 1.167.
6 σ^ 6 (2)
The PCR Cp in the above equation has a useful practical interpretation, namely, P = ( C1 )100 is
p

the percentage of the specification band used up by the process. If we consider the above

example, the percentage of the specification is given as: P = ( 1.167

1
)100 = 85.69 percent of the
specification band.

The above two Equations, (Cp =

USL−LSL
6σ
and P =
1
Cp( )
100) assume that the process has both

upper and lower specification limits. For one-sided specifications, we define one-sided PCRs as:

USL−μ
Cpu = upper specification only and
3σ

μ−LSL
Cpl = lower specification only.
3σ

^ pu and C
Estimates C ^ pl would be obtained by replacing μ and σ in the above equations by
estimates μ^ and σ^ , respectively.

To illustrate the use of the one-sided process-capability ratios, consider the container bursting-
strength data in the above example. Suppose that the lower specification limit on bursting
strength is 200 psi. We will use ^μ ¿ X =264 and s =32 as estimates of μ and σ , respectively, and
μ−LSL 264−200
the resulting estimate of the one-sided lower process-capability ratio is: C pl = =
3σ 3 (32)
= 0.67.

The fraction of defective containers produced by this process is estimated by finding the area to
LSL−^μ 200−264
the left of Z = = = -2 under the standard normal distribution. The estimated
σ^ 32
fallout is about 2.28% defective, or about 22,800 nonconforming containers per million.

2.4.4 Process Capability Ratio for Off-Center Process

The process capability ratio Cp does not take into account where the process mean is located
relative to the specifications. C p simply measures the spread of the specifications relative to the
6-sigma spread in the process.
This situation may be more accurately reflected by defining a new process capability ratio that
takes process centering into account. This quantity is: C pk = min (Cpu, Cpl). Note that Cpk is simply
the one-sided PCR for the specification limit nearest to the process average.
Example:
Suppose that a stable process has upper and lower specifications at USL = 62 and LSL = 38 with
a sample mean of 53 and sample standard deviation of 2. Then compute the PCR for off-center
process.
Solution:
USL−μ μ−LSL 62−53 53−38
Cpk = min (Cpu, Cpl) = min (Cpu = , Cpl = ) = min ( =1.5 , =2.5) =
3σ 3σ 3∗2 3∗2
1.5.
Generally, if Cp = Cpk, the process is centered at the midpoint of the specifications, when C pk <
Cp the process is off-center, and when pk < 0 the implication is that the process mean lies
outside the specifications.

2.4.5 Confidence Intervals on Process Capability Ratios

We can remember that what we actually observe in practice is an estimate of the PCR. This
estimate is subject to error in estimation, since it depends on sample statistics. English and
Taylor (1993) report that large errors in estimating PCRs from sample data can occur, so the
estimate one actually has at hand may not be very reliable. It is always a good idea to report the
estimate of any PCR in terms of a confidence interval. Therefore, it is easy to find a confidence
interval for the “first generation" ratio C p. If we replace σ by s in the equation for Cp, we produce
^ p. If the quality characteristic follows a normal distribution, then a
the usual point estimator C
100(1 - α )% confidence interval on Cp is obtained from:

√ √ √ √
2 2 2 2
USL−LSL χ α χ α χ α χ α
1−
2 , n−1 ≤Cp ≤ USL−LSL 2 ,n−1
^p
= C 1−
2 , n−1 ≤ Cp ≤ ^ 2, n−1 , where
6σ Cp
n−1 6σ n−1 n−1 n−1
χ
2 2 α α
1−
α and χ α are the lower and upper percentage points of the chi-square distribution
2 ,n−1 2 , n−1 2 2
with n - 1 degrees of freedom.
Example: Suppose that a stable process has upper and lower specifications at USL = 62 and
LSL = 38. A sample of size n = 20 from this process reveals that the process mean is centered
approximately at the midpoint of the specification interval and that the sample standard deviation
s = 1.75, then find the 95% confidence interval for a point estimate of Cp.

^p = USL−LSL 62−38
Solution: C = = 2.29 and
6σ 6∗1.75

√ √
2 2
χ α χ α
^p
the 95% CI for Cp = C 1−
2 , n−1 ≤ Cp ≤ ^ 2, n−1
Cp
n−1 n−1

= 2.29
√ 8.91 ≤ ≤
20−1
Cp 2.29
32.85
20−1 √
=1.57 ≤Cp ≤3.01, where χ 20.975 ,19 = 8.91 and χ 20.025 ,19 =32.85.

The confidence interval on Cp in the above example is relatively wide because the sample
standard deviation s exhibits considerable fluctuation in small to moderately large samples. This
means, in effect, that confidence intervals on Cp based on small samples will be wide. Note also
R
that the confidence interval uses s rather than to estimate σ . This further emphasizes that the
d 2 ,n
process must be in statistical control for PCRs to have any real meaning. If the process is not in
R
control, s and could be very different, leading to very different values of the PCR.
d 2 ,n

For more complicated ratios such as Cpk the approximate confidence intervals will be, if the
quality characteristic is normally distributed, then an approximate 100(1 - α )% confidence
interval on Cpk is given as:

C
[ √
^ pk 1−Z 2
α
1
9nC
+
1
^ pk 2(n−1)
2
]
^ pk 1+ Z 2
≤Cpk ≤ C
α
1
9nC [ √
2
+
1
^ pk 2( n−1) ]
Example: consider the above example and construct the 95% CI for C pk by assuming the
observed sample mean is 53.

^ pu = (62-53)/(3*1.75) = 1.714 and C

Solution: C ^ pl = (53-38)/(3*1.75) =2.857

^ pk = min (C
C ^ pu, C
^ pl) = 1.714
Therefore, the 95% CI for Cpk is

= C
[ √
^ pk 1−Z 2
α
1
9nC
+
1
^ pk 2(n−1)
2
]
≤Cpk ≤ C
[ √
^ pk 1+ Z 2
α
1
9nC
2
+
1
^ pk 2(n−1) ]
=
[
1.714 1−1.96
√ 1
9∗20 ¿ 1.96 2
+
1
]
2(20−1) [
≤Cpk ≤ 1.714 1+ 1.96
√ 1
9∗20 ¿ 1.96 2
+
1
2(20−1) ]
=1.154 ≤Cpk ≤2.274.

2.2.6 The Process is not Approximately Normally Distributed

The process capacities that we considered thus far are based on normality of the process
distribution. This poses a problem when the process distribution is not normal. Without going
into the specifics, we can list some remedies.

1. Transform the data so that they become approximately normal.

2. Use or develop another set of process capacity, that apply to non normal distributions.
One statistic is called Cnpk (for non-parametric Cpk). Its estimator is calculated by:

Where p (0.995) is the 99.5th percentile of the data and p(.005) is the 0.5th percentile of
the data. There is, of course, much more that can be said about the case of non normal
data. However, if a transformation can be successfully performed, one is encouraged to
use it.
 CHAPTER THREE
3 Other Statistical Process monitoring and Control Techniques
3.1 Introduction
In the previous chapter, we have concentrated on the basic methods of statistical process-control
(SPC) and capability analysis. Many of these techniques, such as the Shewhart control charts,
have been in use for well over 50 years. But the major disadvantage of a Shewhart control chart
is that it uses only the information about the process contained in the last sample observation and
it ignores any information given by the entire sequence of points. This feature makes the
Shewhart control chart relatively insensitive to small process shifts. However, the increasing
emphasis on variability reduction, yield enhancement, and process improvement along with the
success of the basic methods has led to the development of many new techniques for statistical
process monitoring and control. The two very effective alternatives to the Shewhart control chart
when small shifts are important are the cumulative sum (or CUSUM) and the exponentially
weighted moving average (EWMA) controls charts.
In this chapter we will describe the cumulative sum (CUSUM) and the exponentially weighted
moving average (EWMA) control charts. These procedures are not really new, since they date
from the 1950s, but they are generally considered somewhat more advanced techniques than the
Shewhart charts. CUSUM and EWMA control charts are excellent alternatives to the Shewhart
control chart for process monitoring situations and therefore these control charts are the subject
of this chapter.
3.2 The CUSUM Control Chart for Monitoring the Process Mean
The CUSUM chart directly incorporates all the information in the sequence of sample values by
plotting the cumulative sums of the deviations of the sample values from a target value. For
example, suppose that samples of size n ≥1 are collected, and X j is the average of the jth sample.
Then if μ0 is the target for the process mean, the cumulative sum control chart is formed by
i
plotting the quantity: Ci = ∑ (X j ¿−μ0 )¿ against the sample number i. Ci is called the cumulative
j=1

sum up to and including the i th sample. Because they combine information from several samples,
cumulative sum charts are more effective than Shawhart charts for detecting small process shifts.
Furthermore, they are particularly effective with samples of size n = 1. This makes the
cumulative sum control chart a good candidate for use in the chemical and process industries
where rational subgroups are frequently of size one.
Cumulative sum control charts were first proposed by Page (1954) and have been studied by
many authors; in particular, see Ewan (1963), Page (1961), Gan (1991), Lucas. (1976), Hawkins
(1981, 1993a), and Woodall and Adams (1993).The book by Hawkins and Olwell (1998) is
highly recommended. In this section, we concentrate on the cumulative sum chart for the process
mean. It is possible to devise cumulative sum procedures for other variables, such as Poisson and
binomial variables for modeling nonconformities and fraction nonconforming. We will show
subsequently how the CUSUM can be used for monitoring process variability.
We note that if the process remains in control at the target value μ0, the cumulative sum defined
i 2
σ
in equation Ci = ∑ (X j ¿−μ0 )¿ is a random walk with mean zero and variance i , i.e., E(Ci) = 0
j=1 n
2
σ
and Var(Ci) = i . Therefore, if the process is in control the graph should be neither increasing
n
nor decreasing. However, if the mean shifts upward to some value; μi > μ0, say, then an upward
or positive drift will develop in the cumulative sum C i. Conversely, if the mean shifts downward
to some μi < μ0, then a downward or negative drift in C i will develop. In short, a positive or
negative slope implies the process is out of control. Therefore, if a significant trend develops in
the plotted points either upward or downward, we should consider this as evidence that the
process mean has shifted, and a search for some assignable cause should be performed.
Example: to demonstrate the cusum, let as consider the data: 9.45, 7.99, 9.29, 11.66, 12.16,
10.18, 8.04, 11.46, 9.20, 10.34, 9.03, 11.47, 10.51, 9.40, 10.08, 9.37, 10.62, 10.31, 8.52, 10.84,
10.90, 9.33, 12.29, 11.50, 10.60, 11.08, 10.38, 11.62, 11.31, and 10.52.
To apply the CUSUM to these observations, we would take X = Xi (since our sample size is n =
i
1) and let the target value μ0 = 10. Therefore, the CUSUM becomes: C i = ∑ (X j ¿−10)¿ = (Xi -
j=1

i−1
10) + ∑ (X j ¿−10)¿ = (Xi - 10) + Ci-1, where the starting value for the CUSUM, C o, is taken to
j=1

be zero. The summary data and its CUSUM plots are given below.
How do we decide if the process is out of control? Informally, we will say it is out of control if
there is a slope upwards or downwards. A formal graphicsal method can be based on looking 10
steps ahead and draw a point. Note that for the first 20 observations where μ = 10, the CUSUM
tends to drift slowly, in this case maintaining values near zero. However, in the last 10
observations, where the mean has shifted to μ = 11, a strong upward trend develops.
Of course, the CUSUM plot below is not a control chart because it lacks statistical control limits.
There are two ways to represent CUSUMs, the tabular (or algorithmic) CUSUM, and the V-mask
form of the CUSUM. Of the two representations, the tabular CUSUM is preferable.
3.2.1 The Tabular or Algorithmic CUSUM for Monitoring the Process Mean
We now show how a tabular CUSUM may be constructed for monitoring the mean of a process.
CUSUM may be constructed both for individual observations and for the averages of rational
subgroups.
Let Xi be the ith observation on the process. When the process is in control, X i has a normal
distribution with mean μ0 and standard deviation σ . We assume that either σ is known or that a
reliable estimate is available.
Sometimes we think of μ0 as a target value for the quality characteristic x. This view point is
often taken in the chemical and process industries when the objective is to control to a particular
target value. If the process drifts or shifts off this target value, the CUSUM will signal, and an
adjustment is made to some manipulatable variable to bring the process back on target. Also, in
some cases a signal from a CUSUM indicates the presence of an assignable cause that must be
investigated just as in the Shewhart chart case.
The tabular CUSUM works by accumulating derivations from μ0 that are above target with one

statistic C+ and accumulating derivations from μ0 that are below target with another statistic C -.

The statistics C+ and C- are called one-sided upper and lower CUSUMs, respectively. They are
computed as follows.
The tabular CUSUM:
+¿¿ +¿¿
Ci = max(0, Xi – ( μ0 + k) + C i−1 ) and
−¿¿ −¿¿ +¿¿
Ci = max(0, ( μ0 - k) - Xi + C i−1 ), where the starting values are C 0 = C−¿=0
0
¿
.
In the above equations, k is usually called the reference value (or the allowance, or the slack
value), and it is often chosen about half way between the target μ0 and the out- of-control value
of the mean μ1 that we are interested in detecting quickly.

Thus, if the shift is expressed in standard deviation units as μ1 = μ0 +δ σ (or δ =

|μ1−μ 0|, then k
σ
δ |μ −μ |
is one-half the magnitude of the shift or k = ∗σ = 1 0 .
2 2
+¿¿ −¿¿
Note that C i and C i accumulate deviations from the target value μ0 that are greater than k,
+¿¿ −¿¿
with both quantify reset to zero on becoming negative. If either C i or C i exceed the decision
interval H (the recommended value of the decision interval is H = 5 σ ), the process is considered
to be out of control.
Example: Recall that the above data, the target value is μ0 = 10, the subgroup size is n = 1, the
process standard deviation is σ = 1, and suppose that the magnitude of the shift we are interested
in detecting is 1.0σ = 1.0(1.0) = 1.0. Therefore, the out-of-control value of the process mean is μ1
= 10 + 1 = 11. We will use a tabular CUSUM with k = ½ (because the shift size is 1.0 σ and σ =1)
and H = 5 (because the recommended value of the decision interval is H = 5σ = 5(1) = 5).
+¿¿ −¿¿
To illustrate the calculations, consider period 1. The equations for C i and C i are:
+¿¿ +¿¿
C 1 = max(0, X1 – ( μ0 + k)+ C 0 ) = max(0, 9.45 – (10 + 0.5)+ 0) = max(0, -1.05) = 0 and
−¿¿ −¿¿
C 1 = max(0, ( μ0 - k) – X1+ C 0 ) = max(0, (10 – 0.5) – 9.45 + 0 ) = max(0, 0.05) = 0.05
For period 2, we would use
+¿¿ +¿¿
C 2 = max(0, X2 – ( μ0 + k)+ C 1 ) = max(0, 7.99 – (10 + 0.5)+0) = max(0, -2.51) = 0
and
−¿¿ −¿¿
C 2 = max(0, ( μ0 - k) – X2+ C 1 ) = max(0, (10 – 0.5) – 7.99 +0.05) = max(0, 1.56) = 1.56 and
the likes.

The quantities N+ and N- in Table indicate the number of consecutive periods that the CUSUMs
+¿¿ −¿¿
Ci or C i have been nonzero.
The CUSUM calculations in the above Table show that the upper-side CUSUM at period 29 is
+¿¿ +¿¿
C 29 = 5.28. Since this is the first period at which C i > H = 5, we would conclude that the
process is out of control at that point. The tabular CUSUM also indicates when the shift probably
occurred. The counter N+ records the number of consecutive periods since the upper-side
+¿¿
CUSUM C i rose above the value of zero. Since N+ = 7 at period 29, we would conclude that
the process was last in control at period 29 - 7 = 22, so the shift likely occurred between periods
22 and 23.
It is useful to present a graphical display for the tabular CUSUM. These charts are sometimes
+¿¿ −¿¿
called CUSUM status charts. They are constructed by plotting C i and C i versus the sample
+¿¿ −¿¿
number and each vertical bar represents the value of C i and C i in period i.

The action taken following an out-of-control signal on a CUSUM control scheme is identical to
that with any control chart; one should search for the assignable cause, take any corrective action
required, and then reinifialize the CUSUM at zero. The CUSUM is particularly helpful in
determining when the assignable cause has occurred; as we noted in the previous example, just
count backward from the out-of-control signal to the time period when the CUSUM lifted above
zero to find the first period following the process shift.
In situations where an adjustment to some manipulatable variable is required in order to bring the
process back to the target value, μ0 , it may be helpful to have an estimate of the new process
mean following the shift. This can be computed from:
^μ = ¿
+¿¿
To illustrate the use of the above equation, consider the CUSUM in period 29 with C i = 5.28,
+¿
C 5.28
then we would estimate the new process average as: ^μ = μ0 +k + +¿i ¿ = 10.0 + 0.5 + =
N ¿ 7
11.25.
So, we would conclude that mean has shifted from 10 to 11.25, and then we would need to make
an adjustment in moving down by 1.25 units.
The Standardized CUSUM
Many users of the CUSUM prefer to standardize the variable x i before performing the
xi −μ 0
calculations. Let: yi = be the standardized value of xi. Then the standardized CUSUMs
σ
+¿¿ +¿¿
are defined as: C i = max(0, yi –k + C i−1 ) and
−¿¿ −¿¿
Ci = max(0, - k - yi + C i−1 ), called standardized two-sided CUSUM.
There are two advantages to standardizing the CUSUM. First, the choices of parameters, H and k
are not scale dependent (that is, they do not depend on σ ). Second, a standardized CUSUM leads
naturally to a CUSUM for controlling variability.
3.2.2 A CUSUM for Monitoring Process Variability
It is possible to construct CUSUM control charts for monitoring process variability. Since
CUSUMs are usually employed with individual observations, the procedure due to Hawkins
(1981) is potentially useful. As before, let x i be the normally distributed process measurement
xi −μ 0
with mean or target value μ0 and standard deviation σ . The standardized value of xi is yi =
σ
.

Hawkins (1981, 1993a) suggests creating a new standardized quantity as: vi = √| y |−0.822 .
i
0.349
He suggests that the vi are sensitive to variance changes rather than mean changes. In fact, the
statistic vi is sensitive to both mean and variance changes. Since the in-control distribution of v i
is approximately N(0,1), two one-sided standardized scale (i.e., standard deviation) CUSUMs
+¿ ¿ +¿ ¿
can be established as: Si = max(0, vi –k + Si−1 ) and
−¿ ¿ −¿ ¿
Si = max(0, - k - vi + Si−1 ), called scale CUSUM.
+¿ ¿ −¿ ¿
The S0 and S0 are 0 and the values of k and h are selected as in the CUSUM for controlling
the process mean. The interpretation of the scale CUSUM is similar to the interpretation of the
+¿ ¿
CNSUM for the mean. If the process standard deviation increases, the values of Si will
−¿ ¿
increase and eventually exceed h, whereas if the standard deviation decreases, the values of Si
will increase and eventually exceed h.
3.3 The Exponential Weighted Moving Average Control Chart
The exponentially weighted moving average (or EWMA) control chart is also a good alternative
to the Shewhart control chart when we are interested in detecting small shifts. The performance
of the EWMA control chart is approximately equivalent to that of the cumulative sum control
chart, and in some ways it is easier to set up and operate. As with the CUSUM the EWMA is
typically used with individual observations, and also give the results for rational subgroups of
size n > 1.
3.3.1 The EWMA Control Chart for Monitoring the Process Mean
The EWMA control chart was introduced by Roberts (1959). The exponentially weighted
moving average is defined as: Zi = λ Xi + (1 - λ )Zi-1, where 0 < λ ≤ 1 is a constant and the starting
value of (required with the first sample at i =1) is the process target, so that: Zo = μo .
Sometimes the average of preliminary data is used as the starting value of the EWMA, so that
Zo= X . To show that the EWMA Z i is a weighted average of all previous sample means, we may
substitute for Zi-1 on the right-hand side of the above equation to obtain:
Zi = λ Xi + (1 - λ )[ λ Xi-1 + (1 - λ )Zi-2] = λ Xi + λ (1 - λ )Xi-1 + (1 - λ )2Zi-2 continuing to substituting
i−1
for Zi-j , j=2, 3, 4, …, t, we obtain: Zi = λ ∑ ¿ ¿ Xt-j + (1 - λ )iZ0.
j=0

The weights, λ ¿ decrease geometrically with the age of the sample mean. Furthermore, the
i−1
weights sum to unity, since λ ∑ ¿ ¿ = λ ¿ = 1−¿ .
j=0

Choice of weighting factor: The parameter λ determines the rate at which 'older' data enter into
the calculation of the EWMA statistic. A value of λ = 1 implies that only the most recent
measurement influences the EWMA (degrades to Shewhart chart). Thus, a large value of λ = 1
gives more weight to recent data and less weight to older data; a small value of λ gives more
weight to older data. The value of λ is usually set between 0.2 and 0.3 (Hunter) although this
choice is somewhat arbitrary. Lucas and Saccucci (1990) give tables that help the user select λ .
N.B.: Since the EWMA can be viewed as a weighted average of all past and current
observations, it is very insensitive to the normality assumption. It is therefore an ideal control
chart to use with individual observations.
If the observations Xi are independent random variables with variance o a, then the variance of Z i
2
is: σ z =σ
i
2
( 2−λλ )¿.
Therefore, the EWMA control chart would be constructed by plotting Z i versus the sample
number i (or time). The center line and Control limits for the EWMA control chart are as
follows.

UCL = μ0 + Lσ
(√ 2−λλ ) ¿ ¿
CL = μ0

LCL = μ0−Lσ
(√ 2−λλ ) ¿ ¿, where the factor L in the above equations is the
width of the control limits.
Note that the term ¿ in the equations above approaches unity as i gets larger. This means that
after the EWMA control chart has been running for several time periods, the control limits will

approach steady-state values given by: UCL = μ0 + Lσ (√ 2−λλ ) and

LCL = μ −Lσ
0 (√ 2−λλ )
However, we strongly recommend using the exact control limits for small values of i instead of
the simplified one.
Example: recall the example discussed above for CUSUM with the target value of the mean is μ0
=10 and the standard deviation is σ =1. Apply EWMA control chart with λ = 0.10 and L=2.7.
Solution: to illustrate the calculation of EWMA, consider X1 = 9.45. Then the first value of
EWMA is: Z1 = λ X1 + (1 - λ )Z0 = 0.1(9.45) + (1 – 0.1)(10) = 9.945; Z2 = λ X2 + (1 - λ )Z1 =
0.1(7.99) + (1 – 0.1)(9.945) = 9.7495; etc. the other values of the EWMA statistic are computed
similarly and given in the following table with its UCLs and LCLs.
sample i Xi Zi Mi
2
Si √ S 2i
1 9.45 9.95 9.45 0.925 0.962
2 7.99 9.75 8.72 1.142 1.068
3 9.29 9.70 8.91 1.045 1.022
4 11.66 9.90 9.5975 1.250 1.118
5 12.16 10.13 10.11 1.539 1.241
6 10.18 10.13 10.256 1.385 1.177
7 8.04 9.92 10.266 1.601 1.265
8 11.46 10.08 10.7 1.633 1.278
9 9.20 9.99 10.208 1.531 1.238
10 10.34 10.02 9.844 1.388 1.178
11 9.03 9.92 9.614 1.329 1.153
12 11.47 10.08 10.3 1.390 1.179
13 10.51 10.12 10.11 1.266 1.125
14 9.40 10.05 10.15 1.182 1.087
 15 10.08 10.05 10.098 1.064 1.031
16 9.37 9.98 10.166 0.995 0.997
17 10.62 10.05 9.996 0.928 0.963
18 10.31 10.07 9.956 0.841 0.917
19 8.52 9.92 9.78 0.952 0.976
20 10.84 10.01 9.932 0.926 0.962
21 10.90 10.10 10.238 0.897 0.947
22 9.33 10.02 9.98 0.856 0.925
23 12.29 10.25 10.376 1.187 1.089
24 11.50 10.37 10.972 1.195 1.093
25 10.60 10.40 10.924 1.079 1.039
26 11.08 10.47 10.96 1.009 1.005
27 10.38 10.46 11.17 0.909 0.953
28 11.62 10.57 11.036 0.927 0.963
29 11.31 10.65 10.998 0.879 0.937
30 10.52 10.63 10.982 0.792 0.890

Note from the Figure below the control limits increase in width as I increases from i = 1, 2,...,

until they stabilize at the steady-state values given by equations UCL = μ0 + Lσ (√ 2−λλ ) and LCL
= μ0−Lσ(√ 2−λλ ), i.e.,
√ 2−0.1
UCL = 10+2.7 (1) (
0.1
) √ 2−0.1 ) = 9.38.
= 10.62 and LCL = 10−2.7(1) (
0.1
Interpretation of EWMA control chart: The jagged line is the EWMA statistic over time and the
control chart tells us that the process is in control because all EWMA lie between the control
limits. However, there seems to be a trend upwards for the last 5 or 6 periods.
3.3.2 Monitoring Variability
MacGregor and Harris (1993) discuss the use of EWMA-based statistics for monitoring the
process standard deviation. Let Xi be normally distributed with mean μ and standard deviation σ .
The exponentially weighted mean square error (EWMS) is defined as:
2
Si = λ ¿
It can be shown that E( S2i ) = σ 2 (for large i) and if the observations are independent and normally
2
Si 2−λ
distributed, then has an approximate chi-square distribution with v = , degrees of
σ
2 λ

freedom. Therefore, if σ 0 represents the in-control or target value of the process standard
deviation, we could plot √ S 2i on an exponentially weighted root mean square or EWRMS control

√ √
2 2
χ α χ 1−α
chart with control limits given by: UCL = ν,
2 and LCL = ν,
2 . MacGregor and Harris
σ0 σ0
ν ν
(1993) point out that the EWMS statistic can be sensitive to Shifts in both the process mean and

the standard deviation. They suggest replacing μ0 in equation μ0 + Lσ (√ 2−λλ ) with an estimate ^μ i

at each point in time. A logical estimates of μ0 turns out to be the ordinary EWMA Z i. They
derive control limits for the resulting Exponentially Weighted Moving Variance (EWMV): S2i
¿ λ ¿.
Example: consider the above example and construct the control chart for monitoring variability
by using α =0.05 .
Solution: the variability statistic drawn in the control chart is given in the above Table of the last
column.
 2−λ 2 2
The degree of freedom v = = (2 - 0.1)/0.1 = 19, χ ν, α = χ 219 ,0.025 = 32.85 and χ ν, 1−α =¿
λ 2 2

√ √
2 2

√
2
χ α 32.85 χ 1−α
χ 19 ,0.975 =8.91. therefore, UCL = ν,
2 =(1) = 1.315 and LCL = ν,
2 =
σ0 19 σ0
ν ν

(1)
√ 8.91 = 0.685.
19

3.3.3 The Moving Average Control Chart

The EWMA chart uses a weighted average as the chart statistic. Occasionally, another type of
time-weighted control chart based on a simple, un weighted moving average may be of interest.
Suppose that individual observations have been collected, and let x l, x2, x3,… denote these
observations. The moving average of span w at time i is defined as:

{
X i+ X i−1+ …+ X i −w+1
, if i≥ 5
w
Mi = i
1
∑ X , if i<5
i j=1 j

That is, at time period i, the oldest observation in the moving average set is dropped and the
newest one added to the set. The variance of the moving average Mi is:
i i
1 1 W 2 σ2 3σ
Var(Mi) = 2 ∑
W j =i−w +1
V (X i ) = 2 ∑
W j =i−w +1
σ 2
=
W
2
σ =
W
fori≥ 5 and
√i
for i < 5.

Therefore, if μ0 denotes the target value of the mean used as the center line of the control chart,
then the three-sigma control limits for Mi are:
 {
3σ
μ0 ± , if i≥ 5
√W
3σ
μ0 ± , if i< 5
√i
The control procedure would consist of calculating the new moving average M i as each
observation xi becomes available, plotting Mi on a control chart with upper and lower control
limits given by the above equations, and concluding that the process is out of control if M i
exceeds the control limits. In general, the magnitude of the shift of interest and w are inversely
related; smaller shifts would be guarded against more effectively by longer-span moving
averages, at the expense of quick response to large shifts.
Example: consider the data given in the above example and use using w =5 in order to construct
the moving average control chart.
Solution: the statistic plotted on the moving average control chart will be: M i =
X i + X i−1 +…+ X i−4
for period i≥ 5. For time periods i < 5, the average of the observations for
w
periods 1, 2, 3, …i is plotted. The values of these moving averages are shown in the above Table
of the 6th column.
The control limits for the moving average control chart may be easily obtained from equations
3σ 3σ
UCL = μ0 +¿ and LCL = μ0−¿ . Since we have μ0=10∧¿ σ = 1.0, then
√W √W
3 3
UCL =10+¿ = 11.34 and LCL = 10−¿ = 8.658.
√5 √5
The control limits for Mi apply for periods i ≥5. For periods 0 < i < 5, the control limits are given
3σ
by, μ0 ± .
√i
The moving average control chart is shown in the above figure. No points exceed the control
limits. Note that for the initial periods i < w the control limits are wider than their final steady-
state value. Moving averages that are less than w periods apart are highly correlated, which often
complicates interpreting patterns on the control chart.
N.B.: The moving average control chart is more effective than the Shewhart Chart in detecting
small process shifts. However, it is generally not as effective against small shifts as either the
CUSUM or the EWMA. The moving average control chart is considered by some to be simpler
to implement than the CUSUM.
3.4 Other Univariate Statistical Process Monitoring and Control Techniques
The widespread successful use of the basic SPC methods described in Shewhart control chart,
the CUSUM and EWMA control Charts in the previous chapters have led to the development of
many new techniques and procedures.
Statistical process-control methods have found wide application in almost every type of business.
Some of the most interesting applications occur in job-shop manufacturing systems, or generally
in any type of system characterized by short production runs. Some of the SPC methods for these
situations are straightforward adaptations of the standard concepts and require no new
methodology. In this section, we present a summary of several techniques that have proven
successful in the short production run situation.
3.4.1 X-bar and R Charts for Short Production Runs
The simplest technique for using X-bar and R charts in the short production run situation was
introduced previously. But here we use deviation from nominal instead of the measured variable
on the control chart. This is sometimes called the DNOM control chart. To illustrate the
procedure, consider the following representation:
Xi = Mi – TA,
where TA is the nominal measurements, Mi represents the ith actual sample measurement and Xi is
the deviations from nominal actual measurement.
Example: consider the data given below, the first four samples represent hole diameters in a
particular part (say, part A). Panel (a) of this table shows the actual diameters in millimeters. For
this part, the nominal diameter is T A = 50 mm. and Panel (b) shows the deviations from nominal
Xi, as well as the X and R values for each sample. Consider the last six samples, these hole
diameters are from a different part number, B, for which the nominal dimension is TB = 25 mm.
The control charts for X-bar and R using deviation from nominal are shown below. Note that
control limits have been calculated using the data from all 10 samples. In practice, we would
recommend waiting until approximately 20 samples are available before calculating control
limits. However, for purposes of illustration let us calculate the limits based on 10 samples to
show that, when using deviation from nominal as the variable on the chart, it is not necessary to
have a long production run for each part number.
sample Part
number number M1 M2 M3 X1 X2 X3 X-bar R
1 A 50 51 52 0 1 2 1.00 2
2 A 49 50 51 -1 0 1 0.00 2
3 A 48 49 52 -2 -1 2 -0.33 4
4 A 49 53 51 -1 3 1 1.00 4
5 B 24 27 26 -1 2 1 0.67 3
6 B 25 27 24 0 2 -1 0.33 3
7 B 27 26 23 2 1 -2 0.33 4
8 B 25 24 23 0 -1 -2 -1.00 2
9 B 24 25 25 -1 0 0 -0.33 1
10 B 26 24 25 1 -1 0 0.00 2
X́
=0.17 R = 2.7
Three important points should be made relative to the DNOM approach:
1. An assumption is that the process standard deviation is approximately the same for all
parts. If this assumption is invalid, use a standardized X-bar and R chart
2. This procedure works best when the sample size is constant for all part numbers.
3. Deviations from nominal control charts have intuitive appeal when the nominal
specification is the desired target value for the process.
3.4.2 Standardized X-bar and R Charts
If the process standard deviations are different for different part numbers, the deviation from
nominal (or the deviation from process target) control charts described above will not work
effectively. However, standardized X-bar and R charts will handle this situation easily. Consider
the jth part number. Let R j and Tj be the average range and nominal value of x for this part
Ri
number. Then for all the samples from this part number, plot Ris = on a standardized R chart
Rj
M i −T j
with control limits at LCL = D 3 and UCL = D4, and plot X is = on a standardized X-bar
Rj
chart with control limits at LCL =- A2 and UCL = + A2.
Note that the center line of the standardized X-bar chart is zero because M i is the average of the
original measurements for subgroups of the jth part number.
The target values R j and Tj for each part number can be determined by using specifications for T j
and taking R j from prior history (often in the form of a control chart, or by converting an
 Sd2
estimate of σ into R j by the relationship R j= ). For new parts, it is a common practice to
C4
utilize prior experience on similar parts to set the targets.
3.5 Multivariate Process Monitoring and Control
3.5.1 Introduction
In previous chapters we have addressed process monitoring and control primarily from the
univariate perspective; that is, we have assumed that there is only one process output variable or
quality characteristic of interest. In practice, however, many if not most process monitoring and
control scenarios involve several related variables. Applying univariate control charts to each
individual variable is inefficient and can lead to multivariate methods that consider the variables
jointly are required. There are many situations in which the simultaneous monitoring or control
of two or more related quality characteristics is necessary. For example, suppose that a bearing
has both an inner diameter (x1) and an outer diameter (x2) that together determine the usefulness
of the part. Because both quality characteristics are measurements, they could be monitored by
applying variable control chart and monitoring these two quality characteristics independently
can be very misleading. Rather when we examine the two variables simultaneously, the unusual
behavior of the point is fairly obvious.
In this section we present control charts that can be regarded as the multivariate extensions of
some of the univariate charts of previous discussions. The Hotelling T 2 chart is the analog of the
Shewhart X-bar chart. These multivariate control charts work well when the number of process
variables is not too large, say, 10 or fewer. As the number of variables grows, however,
traditional multivariate control charts lose efficiency with regard to shift detection. A popular
approach in these situations is to reduce the dimensionality of the problem. We show how this
can be done with principal components.
3.5.2 The Multivariate Normal Distribution
In univariate statistical quality control, we generally use the normal distribution to describe the
behavior of a continuous quality characteristic. The univariate normal probability density
−1 x −μ 2
1 ( )
2 σ
f ( x) = e
function is σ √ 2Π ∞ < x <∞; σ > 0 and −∞ < μ < ∞
The mean of the normal distribution is μ and the variance is σ 2. Note that (apart from the minus
sign) the term in the exponent of the normal distribution can be written as (x- μ)(σ 2)-1(x - μ).
Suppose that we have p variable given by x1, x2, x3, …, xp. Arrange these variables in a p
component vector x' = [x1, x2, x3, …, xp]. Let μ' = [ μ1 , μ 2 , μ3 , … , μ p ] be the vector of the means of
the x's, and let the variances and covariance of the random variables in x be contained in a p x p
covariance matrix Σ . The main diagonal elements of Σ are the variances of the x's and the off-
diagonal elements are the co-variances. Now the squared standardized (generalized) distance
from x to μ is (x- μ ¿' Σ −1(x – μ ¿. Therefore, the multivariate normal probability density function
−1
1 2
¿¿
e
is: f(x) = p 2
1
, −∞ ≤ x j ≤ ∞ , j=1 , 2 ,3 , … , p .
(2 π) |Σ|
2
The Sample Mean Vector and Covariance Matrix
Suppose that we have a random sample from a multivariate normal distribution--say, X 1 X2,
X3...,Xn, where the ith sample vector contains observations on each of these variables x i1, xi2, xi3,
n
1
…, xip. Then the sample mean vector is X = ∑ X and the sample covariance matrix is
n i=1 i
n
1
S= ∑ (X ¿¿ i−¿ X )¿ ¿ ¿ ¿
n−1 i=1

That is, the sample variances on the main diagonal of the matrix S are computed as
n n
1 1
S2j= ∑
n−1 i=1
¿ ¿ ¿ and the sample co-variances are S jk= ∑ (X ¿ ¿ ij−¿ X j)( X ¿ ¿ ik− X k )¿ ¿ ¿.
n−1 i=1

We can show that the sample mean vector and sample covariance matrix are unbiased estimators
of the corresponding population quantifies; that is, E( X ) = μ and E(S) = Σ .
The Hotelling T2 control chart
The most familiar multivariate process-monitoring and control procedure is the Hotelling T 2
control chart for monitoring the mean vector of the process. It is a direct analog of the univariate
Shewhart X-bar chart. We present two versions of the Hotelling T 2 chart: one for sub grouped
data, and another for individual observations.
[Link] Sub grouped Data
Suppose that two quality characteristics x 1 and x2 are jointly distributed according to the bi-
variate normal distribution. Let μ1 and μ2 be the mean values of the quality characteristics, and
let σ 1 and σ 2 be the standard deviations of x1 and x2 respectively. The covariance between x1and
x2 is denoted by σ 12. We assume that σ 1 , σ 2, and σ 12 are known. If X 1 and X 2 are the sample
averages of the two quality characteristics computed from a sample of size n, then the statistic

The process monitoring procedure may be represented graphically. Consider the case in which
the two random variables xl and x2 are independent; that is, σ 12 = 0. If σ 12 = 0, then the above
statistic, χ 20 , defines an ellipse centered at ( μ1, μ2) with principal axes parallel to the X 1 , X 2 axes,
as shown in the figure below. Taking χ 20 equal to χ 2α , 2 implies that a pair of sample averages ( X 1 ,
X 2 ) yielding a value of χ 20 plotting inside the ellipse indicates that the process is in control,

whereas if the corresponding value of χ 20 plots outside the ellipse the process is out of control.
The figure below is often called a control ellipse.
In the case where the two quality characteristics are dependent, then σ 12 ≠ 0 and the
corresponding control ellipse is like the following.

Two disadvantages are associated with the control ellipse. The first is that the time sequence of
the plotted points is lost. This could be overcome by numbering the plotted points or by using
special plotting symbols to represent the most recent observations. The second and more serious
disadvantage is that it is difficult to construct the ellipse for more than two quality
characteristics. To avoid these difficulties, it is customary to plot the values of χ 20 for each
sample on a control chart with only an upper control limit at χ 2α , 2. This control chart is usually
called the chi-square control chart.

It is possible m extend these results to the case where p related quality characteristics are
controlled jointly. It is assumed that the joint probability distribution of the p quality
characteristics is the p-variate normal distribution. The procedure requires computing the sample
mean for each of the p quality characteristics from a sample of size n. This set of quality
characteristic means is represented by the p x 1 vector
 []
X1
X2
X3
X= .
.
.
Xp

The test statistic plotted on the chi-square control chart for each sample is: χ 20 =n(x- μ ¿' Σ −1(x– μ ¿,
where μ' = [ μ1 , μ 2 , μ3 , … , μ p ] is the vector of in-control means for each quality characteristic and
Σ is the covariance matrix. The upper limit on the control chart is χ 2α , p.

Estimating μ and Σ
In practice, it is usually necessary to estimate μ and Σ from the analysis of preliminary samples
of size n, taken when the process is assumed to be in control. Suppose that m such samples are
available. The sample means and variances are calculated from each sample as usual; that is,

where xijk is the ith observation on the jth quality characteristics in the k th sample. The covariance
between quality characteristic j and quality characteristic h in the k th sample is
The { X́ j} are the elements of the vector X́ and the p x p average of sample covariance matrices S
is formed as

The average of the sample covariance matrices S is an unbiased estimate of Σ when the process
is in control.
The T2 Control Chart
Now suppose that S from the above matrix is used to estimate Σ and that the vector X́ is taken as
the in-control value of the mean vector of the process. If we replace μ with X́ and
Σ with S, the test statistic now becomes T2=n(x- X́ ¿' S−1(x– X́ ¿. In this form, the procedure is
usually called the Hotelling T2 control chart. The control limits for the T 2 control chart are given
by

When the chart is used for monitoring future production, the control limits are as follows:

Example: The tensile strength and diameter of a textile fiber are two important quality
characteristics that are to be jointly controlled. The quality engineer has decided to use n = 10
fiber specimens in each sample. He has taken 20 preliminary samples, and on the basis of these
data he concludes that X́ 1 = 115.59 psi, X́ 2 = 1.06 x 10-2 inch, S21 = 1.23, S22= 0.83, and
S12= 0.79. Therefore, the statistic he will use for process-control purposes is

The data used in this analysis, the summary statistics and the Hotelting T 2 control chart are
presented as follows. The preliminary sample is used to calculate the upper control limit. If α =
0.001, then the UCL is
Notice that no points exceed tiffs limit, so we would conclude that the process is in control. If we
had used the approximate chi-square control limit the upper control limit is UCL = 15.16. we
would have obtained, χ 20.001 ,2 = 13.816, which is reasonably close to the correct limit but
somewhat too small.
Interpretation of out-of-control signals
One difficulty encountered with any multivariate control chart is practical interpretation of an
out-of-control signal. Specifically, which of the p variables (or which subset of them) is
responsible for the signal? This question is not always easy to answer.

Ak (1985) suggests using X-bar charts with Bonferroni-type control limits [i.e., replace Z α in the
2

Z
X-bar chart control limit calculation with α
(2 p)
].

Another very useful approach to diagnosis of an out-of-control signal is to decompose the T 2

statistic into components that reflect the contribution of each individual variable. If T 2 is the
2
current value of the statistic, and T (i) is the value of the statistic for all process variables except
2 2
the ith one, then Runger, Att, and Montgomery (1996b) show that d i=T −T (i) is an indicator of
the relative consideration of the ith variable to the overall statistic.
[Link] Individual Observations
In some industrial settings the subgroup size is naturally n = 1. This situation occurs frequently
in the chemical and process industries. Since these industries frequently have multiple quality
characteristics that must be monitored; multivariate control charts with n = 1 would be of
interest. Suppose that m samples, each of size n = 1, are available and that p is the number of
quality characteristics observed in each sample. Let X and S be the sample mean vector and
covariance matrix of these observations, respectively, then the Hotelling T2 statistic is becomes:
T2=(x- X ¿' S−1(x– X ¿ and
the control limits for this statistic are:

When the number of preliminary samples m is large--say, m > 100--many practitioners use an
approximate control limit, either

For m > 100, the first equation is a reasonable approximation. The cid-square limit is only
appropriate if the covariance matrix is known, but it is widely used as an approximation. Lowry
and Montgomery (1995) show that the chi-square limit should be used with caution. If p is
large--say, p ≥ 10, then at least 250 samples must be taken (m ≥ 250) before the cid-square
upper control limit is a reasonable approximation to the correct value.
A significant issue in the case of individual observations is estimating the covariance matrix Σ .
Sullivan and Woodall (1995) give an excellent discussion and analysis of this problem, and
compare several estimators. One of these is the "usual" estimator obtained by simply pooling all
m
1
m observations—say S1= ∑ (V ¿¿ i−¿ X)¿ ¿ ¿ ¿.
m−1 i=1

The second estimator [originally suggested by Holmes and Mergen (1993)] uses the difference
between successive pairs of observations: v i = xi+1 – xi, I = 1, 2, 3, …, m-1. Now arrange these
vectors into a matrix V, where
 The estimator for Σ is one-half the sample covariance matrix of these differences:

The following figure shows the T2 control charts from this example. Sullivan and Woodall
(1995) used simulation methods to find exact control limits for this data set (the false alarm
probability ii 0.155).
3.5.3 The Multivariate EWMA Control Chart
The chi-square and T 2 charts described in the previous section are Shewhart-type control charts.
That is, they use information only from the current sample, so consequently, they are relatively
insensitive to small and moderate shifts in the mean vector. Cumulative sum and EWMA control
charts were developed to provide more sensitivity to small shifts in the univariate case, and they
can be extended to multivariate quality control problems.
Lowry et al. (1992) have developed a multivariate version of the EWMA (or art MEWMA)
control chart. The EWMA is a logical extension of the univariate EWMA and is defined as
follows:

Prabhu and Runger (1997) have provided a thorough analysis of the average run- length
performance of the MEWMA control chm7t, using a modification of the Brook and Evans
(1972) Markov chain approach. They give tables and charts to guide selection of the upper
control limit--say, UCL = H-for the MEWMA.

3.5.4 Principal Components

Conventional multivariate control-charting procedures are reasonably effective as long as p (the
number of process variables to b monitored) is not very large. However, as p increases, the
average run length performance to detect a specified shift in the mean of these variables for
multivariate control charts also increases, because the shift is "diluted" in the p-dimensional
space of the process variables. We will discuss the method of principal components.
Basically, if we let C be the matrix whose columns are the eigenvectors, then C’ Σ C = Λ
The sample covariance matrix of the first 20 observations on the x's, in correlation form, is
The Table below presents the remits of a PCA on the first 20 observations, showing the
eigenvalues and eigenvectors, as well as the percentage and the cumulative percentage of the
variability explained by each principal component.
 CHAPTER FOUR
4. Acceptance Sampling
4.1 Introduction
Inspection of raw materials, semi-finished products, or finished products is one aspect of quality
assurance. When inspection is for the purpose of acceptance or rejection of a product, based on
adherence to a standard, the type of inspection procedure employed is usually called acceptance
sampling.
In the 1930s and 1940s, acceptance sampling was one of the major components of the field of
statistical quality control, and was used primarily for incoming or receiving inspection. In more
recent years, it has become typical to work with suppliers to improve their process performance
through the use of SPC and designed experiments, and not to rely as much on acceptance
sampling as a primary quality assurance tool.
A typical application of acceptance sampling is as follows: A company receives a shipment of
product from a supplier. This-product is often a component or raw material used in the
company's manufacturing process. A sample is taken from the lot, and some quality
characteristic of the units in the sample is inspected. On the basis of the information in this
sample, a decision is made regarding lot disposition. Usually, this decision is either to accept or
to reject the lot. Sometimes we refer to this decision as lot sentencing. Accepted lots are put into
production; rejected lots may be returned to the supplier or may be subjected to some other lot-
disposition action.
Although it is customary to think of acceptance sampling as a receiving inspection activity, there
are other uses of sampling -methods. For example, frequently a manufacturer will sample and
inspect its own product at various stages of production. Lots that are accepted are sent forward
for further processing, and rejected lots may be reworked or scrapped. Three aspects of sampling
are important.
1. It is the purpose of acceptance sampling to sentence lots, not to estimate the lot quality.
Most acceptance-sampling plans are not designed for estimation purposes.
2. Acceptance-sampling plans do not provide any direct form of quality control. Acceptance
sampling simply accepts and rejects lots. Even if all units are of the same quality,
sampling will accept some lots and reject others, the accepted lots being no better than
 the rejected ones. Process controls are used to control and systematically improve quality,
but .acceptance sampling is not.
3. The most effective use of acceptance sampling is nor to "inspect quality into the product,"
but rather as an audit tool to ensure that the output of a process conforms to requirements.
Generally, there are three approaches to lot sentencing: (1) accept with no inspection; (2) 100%
inspection-that is, inspect every item in the lot, removing all defective units found and (3)
acceptance sampling.
Acceptance sampling is most likely to be useful in the following situations:
1. When testing is destructive
2. When the cost of 100% inspection is extremely high.
3. When 100% inspection is not technologically feasible or would require so much calendar
time that production scheduling would be seriously impacted.
4. When there are many items to be inspected and the inspection error rate is sufficiently
high that 100% inspection might cause a higher percentage of defective units to be passed
than would occur with the use of a sampling plan.
5. When the supplier has an excellent quality history, and some reduction in inspection from
100% is desired, but the supplier's process capability is sufficiently low as to make no
inspection an unsatisfactory alternative.
6. When there are potentially serious product liability risks, and although the supplier's
process is satisfactory, a program for continuously, monitoring the product is necessary.
4.1.1 Advantages and Disadvantages of Sampling
When acceptance sampling is contrasted with 100% inspection, it has the following advantages:
1. It is usually less expensive because there is less inspection.
2. There is less handling of the product, hence reduced damage.
3. It is applicable to destructive testing.
4. Fewer personnel are involved in inspection activities.
5. it often greatly reduces the amount of inspection error-
6. The rejection of entire lots as opposed to the simple return of defectives often
provides a stronger motivation to the supplier for quality improvements.
Acceptance sampling also has several disadvantages, however. These include the following:
1. There are risks of accepting "bad" lots and rejecting "good" lots.
 2. Less information is usually generated about the product or about the process that
manufactured the product.
3. Acceptance sampling requires planning and documentation of the acceptance sampling
procedure whereas 100% inspection does not.
There are a number of different ways to classify acceptance-sampling plans. One major
classification is by attributes and variables. Variables, of course, are quality characteristics that
are measured on a numerical scale and attributes are quality characteristics that are expressed on
a "go, no-go" basis. In this chapter we will deal with lot-by-lot acceptance-sampling plans for
attributes and variables sampling plans along with a brief discussion of several special
acceptance-sampling procedures.
4.1.2 Definitions of Basic Acceptance Sampling Terms
A single-sampling plan is a lot-sentencing procedure in which one sample of n units is selected
at random from the lot, and the disposition of the lot is determined based on the information
contained in that sample. For example, a single-sampling plan for attributes would consist of a
sample size n and an acceptance number c. The procedure would operate as follows: Select n
items at random from the lot. If there are c or fewer defectives in the sample, accept the lot, and
if there are more than c defective items in the sample, reject the lot.
Double-sampling plans are somewhat more complicated. Following an initial sample, a
decision based on the information in that sample is made to (1) accept the lot, (2) reject the lot, or
(3) take a second sample. If the second sample is taken, the information from both the first and
second sample is combined in order to reach a decision whether to accept or reject the lot.
A multiple-sampling plan is an extension of the double-sampling concept, in that more than
two samples may be required in order to reach a decision regarding the disposition of the lot.
Sample sizes in multiple sampling are usually smaller than they are in either single or double
sampling.
Sequential sampling is the ultimate extension of multiple sampling, in which units are selected
from the lot one at a time, and following inspection of each unit, a decision is made either to
accept the lot, reject the lot, or select another unit.
Lot Formation
How the lot is formed can influence the effectiveness of the acceptance-sampling plan.
There are a number of important considerations in forming lots for inspection. Some of these are
as follows:
a. Lots should be homogeneous. The units in the lot should be produced by the same
machines, the same operators, and from common raw materials, at approximately the
same time. When lots are non homogeneous, such as when the output of two different
production lines is mixed, the acceptance-sampling scheme may not function as
effectively' as it could. Non homogeneous lots also make it more difficult to take
corrective action to eliminate the source of defective products.
b. Larger lots are preferred over smaller ones. It is usually more economically efficient to
inspect large lots than small ones.
c. Lots should be conformable to the materials-handling systems used in both the supplier
and consumer facilities. In addition, the items in the lots should be packaged so as to
minimize shipping and handling risks, and so as to make selection of the units in the
sample relatively easy.
Rectifying Inspection: Acceptance-sampling programs usually require corrective action when
lots are rejected. This generally takes the form of 100% inspection or screening of rejected lots,
with all discovered defective items either removed for subsequent rework or return to the
supplier, or replaced from a stock of known good items. Such sampling programs are called
rectifying inspection programs, because the inspection activity affects the final quality of the
outgoing product.
Lot Tolerance Percent Defective (LTPD): The LTPD is a designated high defect level that
would be unacceptable to the consumer. The consumer would like the sampling plan to have a
low probability of accepting a lot with a defect level as high as the LTPD.
Type I Error (Producer's Risk): This is the probability, for a given (n,c) sampling plan, of
rejecting a lot that has a defect level equal to the AQL. The producer suffers when this occurs,
because a lot with acceptable quality was rejected. The symbol is commonly used for the Type
I error and typical values for range from 0.2 to 0.01.
Type II Error (Consumer's Risk): This is the probability, for a given (n,c) sampling plan, of
accepting a lot with a defect level equal to the LTPD. The consumer suffers when this occurs,
because a lot with unacceptable quality was accepted. The symbol is commonly used for the
Type II error and typical values range from 0.2 to 0.01.
Average Sample Number (ASN): For a single sampling LASP (n, c) we know each and every lot
has a sample of size n taken and inspected or tested. For double, multiple and sequential LASP's,
the amount of sampling varies depending on the number of defects observed. For any given
double, multiple or sequential plan, a long term ASN can be calculated assuming all lots come in
with a defect level of p. A plot of the ASN, versus the incoming defect level p, describes the
sampling efficiency of a given LASP scheme.
Random Sampling: the units selected for inspection from the lot should be chosen at random,
and they should be representative of all the items in the lot. The random-sampling concept is
extremely important in acceptance sampling. Unless random samples are used, bias will be
introduced. For example, the supplier may ensure that the units packaged on the top of the lot are
of extremely good quality, knowing that the inspector will select the sample from the top layer.
"Salting" a lot in this manner is not a common practice, but if 'it occurs and nonrandom-sampling
methods are used, the effectiveness of the inspection process is destroyed.
4.1.3 Guidelines for Using Acceptance Sampling
An acceptance-sampling plan is a statement of the sample size to be used and the associated
acceptance or rejection criteria for sentencing individual lots. A sampling scheme is defined as a
set of procedures consisting of acceptance-sampling plans in which lot sizes, sample sizes, and
acceptance or rejection criteria along with the amount of 100% inspection and sampling are
related. Finally, a sampling system is a unified collection of one or more acceptance-sampling
schemes. In general, the selection of an acceptance-sampling procedure depends on both the
objective of the sampling organization and the history of the organization whose product is
sampled. Furthermore, the application of sampling methodology is not static; that is, there is a
natural evolution from one level of sampling effort to another.
4.2 Acceptance Sampling for Proportion (Attributes)
Suppose that each item has an attribute (qualitative characteristics)-say cups might be
cracked/not cracked, fuses are defective/not defective. We want to decide whether or not to
accept a batch based on sample observations. Let a lot of size N has been submitted for
inspection based on sample size of n. Since the quality characteristic inspected is an attribute,
each unit in the sample is judged to be either conforming or nonconforming. One or several
attributes can be inspected in the same sample; generally, a unit that is nonconforming to
specifications on one or more attributes is said to be a defective unit. This procedure is called a
single-sampling plan because the lot is sentenced based on the information contained in one
sample of size n.
4.2.1 Single Stage Sampling Plans
Acceptance sampling is a method used to make a decision as to whether to accept or to reject lots
or batches based on inspection of samples. The objective is not to control or estimate the quality
of lots only to pass a judgment on lots. Here we have the following notations: p is a proportion
of defectives in the batches/lots and d is number of defective items in the n sample. If the
number of observed defectives d is less than or equal to c, is known as acceptance number, the lot
will be accepted. If the number of observed defectives d is greater than c, the lot will be rejected.
The quantities n and c can either be determined by lot quality protection approach or by an
average quality protection approach. A single-sampling plan is defined by the sample size n and
the acceptance number c. Thus, if the lot size is N, then taking the sample size n and acceptance
number c means that from a lot of size N a random sample of n units is inspected and the number
of nonconforming or defective items d observed.
The procedure to arrive at a decision about the acceptance or rejection of a lot is as follows:-

(i) Select a random sample of size n from the submitted lot.

(ii) Inspect each and every unit included in the sample and classify them as defective or
non-defective. Suppose the number of defectives in the sample is d. the criteria for the
acceptance or rejection of lot is,

¿c
(a) If d , the lot is accepted. In this case all the defective items in the sample are replaced
by non-defectives and the lot is release for marketing.
(b) If d>c, the lot is rejected. In this situation, inspect the whole lot and replace all
defectives found in the lot by non defectives and release the lot for sale.
Case I: N is Large (Consider N Infinite)
An important measure of the performance of an acceptance-sampling plan is the operating-
characteristic (OC) curve. This curve plots the probability of accepting the lot versus the lot
fraction defective. Thus, the OC curve displays the discriminatory power of the sampling plan.
That is, it shows the probability that a lot submitted with a certain fraction defective will be
either accepted or rejected. Suppose that the lot size N is large (theoretically infinite). Under this
condition, the distribution of the number of defectives d in a random sample of n items is
binomial with parameters n and θ , where θ is the fraction of defective items in the lot. But most
of the time θ is unknown and estimated by P.

Therefore, the probability of observing exactly X defectives is: P(d defectives) = P(d) = ( nd) p ¿.
d

The probability of accepting is simply the probability that X is less than or equal to c, or
c c
OC(p)= Pa = P(d ≤ c/ p ¿ = ∑ P(X =c / p) = ∑ n p ¿ ¿
d =0 d =0 d
d
()
That is,

Pa = (n0) p ¿ + (n1) p ¿ + (n2) p ¿ + ….+ (nc) p ¿

0 1 2 c

P = ¿ + np ¿ + ( ) p ¿ + ….+ ( ) p ¿
n 2 n c
a
2 c
From this we can see that

 It is a decreasing function

 Its value at c=0 is OC( p ) = (1- p ) n

 Its value at p=0 , OC(0) =1

 Its value at p=1 , OC(1) =0

Effect of n and c on OC Curves

The OC curve runs horizontally at a probability of acceptance Pα = 1.00 until a level of lot
quality that is considered "bad" is reached, at which point the curve drops vertically to a
probability of acceptance Pα = 0.00, and then the curve runs horizontally again for all lot fraction
defectives greater than the undesirable level. If such a sampling plan could be employed, all lots
of "bad" quality would be rejected, and all lots of "good" quality would be accepted.
Unfortunately, the ideal OC can almost never be obtained in practice. In theory, it could be
realized by 100% inspection, if the inspection were error free. The ideal OC curve shape can be
approached, however, by increasing the sample size. The OC curve becomes more like the
idealized OC curve shape as the sample size increases. (Note that the acceptance number c is
kept proportional to n.)
Thus, the precision with which a sampling plan differentiates between good and bad lots
increases with the size of the sample. In general, the OC function depends on n and c. We need
to choose n and c to give us a “good” OC function (in which similar as possible as the ideal
curve).
I. Fix c =0
This corresponds to taking a sample of n items and accepting the batch only if there are no
defects in the sample. That is all items are good.
c
OC(θ )=∑
d =0
( nd) θ ¿ ¿, if c = 0, then
d

c
OC(θ )=∑
d =0
(0n) θ ¿ ¿ =¿ , this is P(no defects in the sample/a proportion of θ of the population are
0

defective). We now to choose n in a sensible way.

Example: suppose we want to be 95% of a certain rejection of a batch that has 5% of defectives,
to do this we can solve:

P(rejecting) = 0.95 means P(accepting) = 1-0.95, that is: OC(θ )¿1 −0.95=0.05

Since OC(θ ) =(1-θ )n

⇒ OC(0.05) = (1-0.05)n

⇒ (0.95)n ¿ 0 . 05

By taking logarithm both sides we get,

n*log0.95¿ log 0 . 05

= n¿ log 0 . 05 \log 0.95 =58.4 ¿ 59

= n¿ 59 n must be an integer so choose n=59, we can use “take a sample of 59 objects and
accept the batch if every object is good otherwise reject if there is ¿ 1 defectives”.
If we want to be even more certain, we might assume c=0, the conditions be 99% of a certain
rejecting a batch with 2% of defectives
Solution: θ=2 % , OC(θ ) ¿ 0 . 01
⇒ OC(0.02) = (1-0.02)n ¿ 0 . 01

⇒ (1-0.02)n ¿ 0 . 01 . By taking natural logarithm both sides we get,

n*log(0.98) ¿ log 0 . 01
⇒ n¿ log 0 . 01 /log 0 . 98 =227.9, that is, n¿ 228 since n must be an integer, so

that we can “take a sample of 228 observation accept the batch if every object is good and reject
if there is¿ 1 defectives”.

II. Fix n
Suppose that instead of fixing c we fix n. For instance, let take n =100, then what value should
we take for c in we want to make sure we are at least 95% of rejecting a batch which has 5% of
defectives?
We choose c to solve OC (0.05)¿ 0 . 05 , try to c is 0, 1, 2, 3…n =100.
c= 0, OC(0.05) = P(d ≤ 0 ¿ = (1-0.05)100 = (0.95)100 = 0.00592

c=1, OC(0.05) = p(d¿ 1) = p(d=0) + p(d=1) by using the binomial distribution we can get

( 100¿) ¿ ¿¿
= (0.95)100 +¿ (0.05) 1
(0.95)99 = 0.037

c=2, OC(0.05) =p(d¿ 2) = p(d=0) + p(d=1) +p(d=2) by using the binomial distribution we can

( 100¿) ¿ ¿¿ ( 100¿) ¿ ¿¿
get: OC(0.05) = (0.95)100 +¿ (0.05) 1
(0.95)99 +¿ (0.05) 2
(0.95)98 = 0.11

Therefore, OC(0.05) = {¿ 5 % ,¿c=2

5 % , c=0 , 1
, 3 , 4 , … , 100
, c = 0 and c = 1 both satisfy the constraint.
 c=0 c=1
Buyer’s risk Lower Higher (we might accept a bad lot)
Producer’s risk Higher (we might reject a good lot) Lower

Of the two, taking c=1 leads to accepting more bad batches, but it also leads to accepting more
good batches. In practice we would probably take c=1. First we put a limit on the buyer’s risk.
Then we minimize the producer’s risk subject to that limit. (This procedure is similar to
hypothesis testing).
c
III. Fix
n
For instance, a buyers test a sample of n fuses from infinite batch (population) of fuses. She will
accept the batch if the sample contains ¿ 4 % defectives. She is considering by taking a sample of
n = 25, 50, 75, 100 etc fuses. How large should “n” be if she wants to be 90% of certain of
rejecting a batch with 10% of defectives? (the null is accepting the batch if d ≤0.04n) i.e., c\
n=0.04 find the value of c and n so that OC (0.1) ≤ 10%. Possible solutions to check: n=25, c=1;
n=50, c=2; n=75, c=3 and n=100, c=4. Therefore, OC(0.1) ¿ 0 . 01 C=1,n=25, implies that

( 25¿) ¿ ¿¿
OC(0.1)=p(x¿ 1) =p(d=0)+p(d=1) = (0.9)25 +¿ (0.1) 1
(0.9)24 =0.27,i.e., OC(0.1) = 0.27 >10%,
that is this is not solution set.

c=2, n=50, OC(0.1) =P(d ¿ 2) =p(d =0)+ p(d =1) +p(d =2)

( 50¿) ¿ ¿¿ ( 50¿) ¿ ¿¿
= (0.9)50 +¿ (0.1) 1
(0.9)49 +¿ (0.1) 2
(0.9)48= 0.11>10%, that is this is also not solution set.

c=3, n=75, OC(0.1) =P(d¿ 3 ) = P(x=0) +…P(d=3)

We finally get as, the same procedure above, = 0.059<10%
⇒ OC(0.1) = 0.059, that is this is a solution set. Therefore, we need n¿ 75 and c=3.
The OC function for the 3 sampling schemes is tabulated below.
c= 1, n=25 c=2, n=50 c=3, n=75
θ 0.01 0.97 0.986 0.993
0.03 0.83 0.811 0.812
0.04 0.74 0.677 0.647
0.05 0.64 0.541 0.480
0.1 0.27 0.111 0.05
0.2 0.027 0.001 7x10-5

With an increasing proportion of defects, θ the probability that the lot is accepted decrease. The
ideal OC function is very close to 1 for small values of θ and very close to 0 for values of θ
greater than an acceptable quality limit. Note that d is follows that a binomial distribution with
the parameters of n and θ, so as n approaches to infinity the weak law of large numbers tells us

d/n approaches to θ in probability, i.e., for ε >0, P( |dn −θ|> ε ¿ →[Link] suppose θ < 0.04, then
d d
we can choose ε >0, so that θ+ ε >0.04. P( > 0.04 ¿ ≤ P( >θ +ε ¿ → 0, i.e., P(reject batch)→ 0
n n
as n→ ∞ if θ < 0.04. Similarly, we can show that P(reject batch)→1 as n∞ if θ >0.04.

Case 2: N is small (consider N is finite, N<∞ )

A sampling plan about the lot sentencing is based on one of the sample that has been inspected. Suppose a

lot consists of N (consider N is finite) items having a proportion of defectives as θ and in all D defectives,

i.e. D=Nθ and a sample of size n drawn from a submitted lot which contains d defectives with the
maximum allowable number of defectives in the sample, c. Here we have a batch of N items, M of
M 1 2 3
this are defectives; θ is proportion of defectives = N , where (θ takes the value: 0, N , N , N ,

N −1
…. ,1), D= number of defective in the sample has hyper- geometric distribution. Then,
N
Then, in single sampling plan, the variable d follows the hyper geometric distribution. Various
probabilities under this plan can be worked out by the following formula.
1. The probability of getting d defective in the sample is:
❑
P(D=d) = ❑ ,where d¿ n , n-d¿ N−M

2. The probability of accepting the lot having not more than c defectives in the sample of n items is,
c

∑ ( Md )( Nn−d
−M
)
d=0
P(D=d) =
( Nn )
Example: Suppose N=1000, c=0, choose n to make sure OC(0.05) ¿ 5 % i.e. we have a batch of
size 1000 we use the rule “accept if the sample contains no defectives”. Choose n so that we are
95% of rejecting a batch is 5% defectives.
We can choose n so that OC(0.05) ¿ 0 . 05
M θ=5 %
⇒ =¿ , N =1000, M=Nθ =50.
N
OC(d =0\M=50) = P(every sample observation have no defective/M=50) = 950\100
¿ 949 ¿×948 ¿×.. .×950−( n−1) ¿−( n−1) .

OC (0.05) = P(d = 0) =
( )(
M N−M
d n−d ) =
( 0 )( n−0 )
50 950

( Nn ) ( 1000
n )
This is solved by substitution and a solution can’t be obtained in a closed form but we want to
choose n large enough to make sure this product is ¿ 0 . 05 .
Example: Suppose the batch size is N= 40 and we take c=0, then how can we choose n so that
OC (0.05) ¿ 5 % , that is for this value of n calculate the OC functions?
If N=40, θ can only take one of 0, 1/40, 2/40, …, , so, if θ=0 .05 , θ = M ¿ then, M=Nθ
=0.05 ¿ 40=2 .
We want to solve: OC(0.05) = P(accept the batch/M=2) ¿ 0 . 05 .
= P(1st item is good )¿ P(2nd item is good/1st item is good) ¿ P(3rd item is good/1st and 2nd item is
good) ¿ ……¿ P(nth item is good /1st ,2nd ,3rd ….(n-1)th item is good).
That is P(no defectives in the sample) is:
 ¿ x )
( d )( n−d )
M N−M
(20)( 40−2
n−0 ) (38n ) 38 ! / 40!
OC(0.05) = P(X = = = = n ! ( 38−n ) ! n ! ( 40−n ) ! =
( Nn ) ( 40n ) (n)
40

( 40−n ) (39−n)!
39 x 40
Then by substitution and inspection, n = 19, 30, 31 OC(0.05) is 0.071, 0.058, 0.046, respectively.
Therefore, we need to n¿ 31 is the minimum sample observation.
4.3.2 Double Stage Sampling Plans
A number of extensions of single-sampling plans for attributes are useful. These include double-
sampling plans, multiple-sampling plans, and sequential-sampling plans were invented to give a
questionable lot another chance. A double-sampling plan is a procedure in which, under, certain
circumstances, a second sample is required before the lots can be sentenced.
Double sampling is the results of the first sample are not conclusive with regard to accepting or
rejecting, a second sample is taken. A double-sampling plan is defined by four parameters.
n1 = sample size on the first sample,
cl = acceptance number of the first sample,
n2 = sample size on the second sample,
c2 = acceptance number for both samples.
Suppose n1 is the first sample size c1 is the first acceptance number, n2 is the second sample size
and c2 is the second acceptance number. Thus, a random sample of n 1 items is selected from the
lot, and the number of defectives in the sample, d 1, is observed. If d1 ≤ c1, the lot is accepted on
the first sample. If d1 > c2, the lot is rejected on the first sample. If c 1 < d1 < c2, a second random
sample of size n2 is drawn from the lot, and the number of defectives in this second sample, d 2, is
observed. Now the combined number of observed defectives from both the first and second
sample, d1 + d2, is used to determine the lot sentence. If d 1 + d2 ≤ c2, the lot is accepted. However,
if d1 + d2 ¿ c2, the lot is rejected. The operation of this double-sampling plan is illustrated
graphically as follows.
The principal advantage of a double-sampling plan with respect to single sampling is that it may
reduce the total amount of required inspection. Double sampling has two potential disadvantages:
First, unless curtailment is used on the second sample, under some circumstances double
sampling may require more total inspection than would be required in a single-sampling plan that
offers the same protection. The second disadvantage of double sampling is that it is
administratively more complex, which may increase the opportunity for the occurrence of
inspection errors.
The OC Curve
The performance of a double-sampling plan can be conveniently summarized by means of its
operating-characteristic (OC) curve. The OC curve for a double-sampling plan is somewhat more
involved than the OC curve for single sampling. A double-sampling plan has a primary OC curve
that gives the probability of acceptance as a function of lot or process quality. The OC curve for
the probability of rejection on the first sample is simply the OC curve for the single4ampling
plan n = nl and c = c2.
Example: Primary and supplementary OC curves for the plan n l = 50, cl = 1, n2 = 100, c2 = 3 are
constructed as follows.
Solution: we assume N is infinite, so use the binomial rather than the hyper-geometric.
If Pa denotes the probability of acceptance on the combined samples, PaI and PaII denote the
probability of acceptance on the first and second samples, respectively, then: Pa = PaI + PaII
 I
Pa is just the probability that we will observe dl ≤ c1 = 1 defectives out of a random sample of n1
1

= 50 items. Thus: P = I
a ∑
d1 =0 (50d ) p ¿ ¿. Now let p = 0.05, then the fraction defective in the
1
d1

incoming lot is: PaI = ∑

d1 =0 (50d ) 0.05 ¿ ¿ = 0.279.
1
d1

To obtain the probability of acceptance on the second sample, we must list the number of ways
the second sample can be obtained. A second sample is drawn only if there are two or three
defectives on the first sample, that is, if c1 <dl ≤ c2.
1. d1 = 2 and d2 = 0 or l; that is, we find two defectives on the first sample and one or less
defectives on the second sample. The probability of this is:
p{d1 =2, d2 ≤ 1} = p{ p(d1=2) p(d2 ≤ 1¿ }

= (502) 0.05 ¿ 2

= (0.261)(0.037) = 0.0097.
2. d1= 3 and d2 = 0; that is, we find thee defectives on the first staple and no defectives on the
second staple. The probability of this is:
p{d1 =3, d2 ¿ 0} = p{ p(d1=3) p(d2 ¿ 0 ¿}

= (503) 0.05 ¿ = (0.22)(0.0059) = 0.001.

3

Thus, the probability of acceptance on the second sample is:

II
Pa = p{d1 =2, d2 ≤ 1} + p{d1 =3, d2 ¿ 0} = 0.0097 + 0.001 = 0.0107.
The probability of acceptance of a lot that has fraction defective p = 0.05 is therefore
Pa = PaI + PaII= 0.279+ 0.0107 = 0.2897.
N.B.: the OC function for the general case (n, m, c1, d1, c2) for d1 and d2 ≤ c2 – d1 is:
c1 d 1−1 c 2−d1

OC(θ ) = ∑
d1 =0 ( )
d1
1

d =c + 1 d 1
1
( )
n θ d ¿ ¿ + ∑ n θ d ¿ ¿ * ∑ m θ d ¿ ¿.
d =0 d 2
1
1

2
() 2

Examples:
1. Find the OC function when n = 100, c1 = 1, d1 = 4, m = 300 and c2 =6.
Solution:
c1 d 1−1 c 2−d2

OC(θ ) = ∑
d1 =0 ( )
n θd ¿ ¿ +
d1
1
∑ n θ d ¿ ¿*
d =c + 1 d 1
1 1
∑( ) m θ d ¿ ¿.
d =0 d 2
1

2
( ) 2
 1 3 6−d1

= ∑
d1 =0 ( )
100 θd ¿ ¿
d1
1
+∑
d =2 d 1
1
( )
100 θ d ¿ ¿
* ∑ 300 θ ¿ ¿.
d =0 d2
d 1

2
( ) 2

4 3
= P(d 1=0)+P(d 1=1 )+P(d 1=2 )* ∑
d2 =0 ( )
300 θd ¿ ¿ d =3
d2
2
+P( 1 ) *∑
d =0
300 θd ¿ ¿
d2 2
.
( ) 2

2. Suppose we intend to use the plan with n 40, c 1 = 0, d1 = 2, c2 = 1, then how should we choose
m if we want OC(0.1) < 5%.
Solution:
c1 d 1−1 c 2−d2

OC(θ ) = ∑
d1 =0 ( )
n θd ¿ ¿ +
d1
1

d =c + 1
1 1 1
( )
∑ dn θ d ¿ ¿* ∑ dm θ d ¿ ¿.
d =0 2
1

2
( ) 2

= ¿ + 40 θ ¿
OC(0.1) = ¿ + 40 x 0.1 ¿
=0.9 40 + 4x0.939 +m< 0.05
= 0.939 +m < ¼ (0.05 - 0.9 40 ¿
= (39 + m) > ln(¼ (0.05 - 0.9 40 ¿)/ln(0.9) = 44.9
m > 5.9, take m = 6.
The Average Sample Number Curve
In single sampling, the size of the sample 'inspected from the lot is always constant, whereas in
double sampling, the size of the sample selected depends on whether or not the second sample is
necessary. The probability of drawing a second sample varies with the fraction defective in the
incoming lot. With complete inspection of the second sample, the average sample size in double
sampling is equal to the size of the first sample times the probability that there will only be one
sample, plus the size of the combined samples times the probability that a second sample will be
necessary. Therefore, a general formula for the average sample number in double sampling, if we
assume complete inspection of the second sample, is:
ASN = n1P1 + (n1 +n2)(1- P1) = n1 + n2(1 – P1), where P1 is the probability of making a lot-
dispositioning decision on the first sample. This is
P1 = P(lot is accepted on the first sample) + P(lot is rejected on the first sample).
If the above equation is evaluated for various values of lot fraction defective p, the plot of ASN
versus p is called an average sample number curve.
4.3.3 Multiple Stage Sampling
Multiple sampling is an extension of double sampling. It involves inspection of 1 to k successive
samples as required to reach an ultimate decision. The rule of thumb suggests that k = 7 is a good
number. Multiple sampling plans are usually presented in tabular form.
Procedure for multiple sampling
The procedure commences with taking a random sample of size n1from a large lot of size N and
counting the number of defectives, d1.
if d1 a1 the lot is accepted.
if d1 r1 the lot is rejected.
if a1 < d1 < r1, another sample is taken.
If subsequent samples are required, the first sample procedure is repeated sample by sample.
For each sample, the total number of defectives found at any stage, say stage i, is

This is compared with the acceptance number ai and the rejection number ri for that stage until a
decision is made. Sometimes acceptance is not allowed at the early stages of multiple sampling;
however, rejection can occur at any stage.
Efficiency measured by the ASN:-
Efficiency for a multiple sampling scheme is measured by the average sample number (ASN)
required for a given Type I and Type II set of errors. The number of samples needed when
following a multiple sampling scheme may vary from trial to trial, and the ASN represents the
average of what might happen over many trials with a fixed incoming defect level.
4.3.4 Sequential Sampling Plans
Sequential sampling is an extension of the double-sampling and multiple-sampling concept. In
sequential sampling, we take a sequence of samples from the lot and allow the number of
samples to be determined entirely by the results of the sampling process. In practice, sequential
sampling can theoretically continue indefinitely, until the lot is inspected 100%. If the sample
size selected at each stage is greater than one, the process is usually called group sequential
sampling. If the sample size inspected at each stage is one, the procedure is usually called item-
by-item sequential sampling.
The operation of an item-by-item sequential-sampling plan is the cumulative observed number of
defectives is plotted on the chart. For each point, the abscissa is the total number of items
selected up to that time, and the ordinate is the total number of observed defectives. If the plotted
points stay within the boundaries of the acceptance and rejection lines, another sample must be
drawn. As soon as a point fails on or above the upper line, the lot is rejected. When a sample
point fails on or below the lower line, the lot is accepted. The equations for the two limit lines for
specified values of p1, 1 -α , P2, and β are:

where

Diagrams of items by item sampling

Description of sequential sampling graph: The cumulative observed number of defectives is

plotted on the graph. For each point, the x-axis is the total number of items thus far selected, and
the y-axis is the total number of observed defectives. If the plotted point falls within the parallel
lines the process continues by drawing another sample. As soon as a point falls on or above the
upper line, the lot is rejected. And when a point falls on or below the lower line, the lot is
accepted. The process can theoretically last until the lot is 100% inspected. However, as a rule
of thumb, sequential-sampling plans are truncated after the number inspected reaches three times
the number that would have been inspected using a corresponding single sampling plan. Note
that, instead of using the graph to determine the fate of the lot, one can resort to generating tables
(with the help of a computer program).
Example: let p1 = 0.01, p2 = 0.06, = 0.05, = 0.10. Thus:

The acceptance and rejection numbers are found by substituting values of n into the equations for
the acceptance and rejection Lines. For example, the calculations for n = 45 are:

Acceptance and rejection numbers must be integers, so the acceptance number is the next integer
less than or equal to xa and the rejection number is the next integer greater than or equal to xr.
Thus, for n =24, the acceptance number is 0 and the rejection number =3. Note that the lot cannot
be accepted until at least 44 units have been tested.
The OC Curve and ASN Curve for Sequential Sampling
The OC curve for sequential sampling can be easily obtained. Two points on the curve are (P l, 1 -
h2
α ) and (P2, β ). A third point, near the middle of the curve, is p = s and Pa = .
h1 + h2

The average sample number taken under sequential sampling is: ASN = Pa = ( CA ) + (1- Pa)( CB ),
and C = plog( ) + (1 – p)log(
( 1−αβ ), B = ( 1−β
α ) 1− p )
p2 1− p2
where A = log .
p 1 1
4.3.5 Comparing and Choosing Sampling schemes
There are many ways of measuring the performance of a scheme such as OC percentiles, quality
and total inspection time.
OC Percentiles
Operating characteristic (OC) curve: the OC curve shows the relationship between the
probabilities of acceptance of a lot P a for variables in the lot quality. It is plotted by taking θ
along horizontal scale and Pa along the vertical scale. The following points (in usual situations)
are common to any O.C curve.
The Rejectable Quality Level (RQL) is the value of θ for which the quality of the batch is
unacceptable.
The Acceptable Quality Level (AQL): is the value of θ for which the quality of the batch can
be regarded as good.
We want OC(RQL) small and we might fix a small number, β (consumer’s risk) and insist
[OC(RQL) < β ]. And also, we want OC(AQL) close to 1and we might fix a small number, α
(producer’s risk-example 5%) and insist [OC(AQL) > 1 - α ].
θ 0 OC(AQL) OC(RQL) 1
OC(θ ) 1 1−α β 0

The points can be easily obtained as: OC (0) =1 and OC (1) =0 since a lot containing no
defective at all, (θ =0) will always be accepted and a lot containing 100% defective items ( θ =1),
will always be rejected.

Average Outgoing Quality (AOQ): A common procedure when sampling and testing is non-
destructive, is to 100% inspect rejected lots and replace all defectives with good units. In this
case, all rejected lots are made perfect and the only defects left are those in lots that were
accepted. If all lots come in with a defect level of exactly θ, and the OC curve for the chosen (n,
c) indicates a probability pa of accepting such a lot, over the long run is:
P(Defective) = P(defective ∩ batch was accepted) +P(defective ∩ batch was rejected)
= P(defective ¿ batch was accepted) P(accepted) +P(defective ¿ batch was rejected)P(rejected)
= θ OC(θ ). This is called the average outgoing quality and AOQ (θ ) = θ OC(θ ¿ in which
AOQ (θ ) < θ OC(θ ¿ for θ < 1.
The AOQ is a function of θ and we want this to be small, but we can not minimize it as θ is
unknown.
Average Outgoing Quality Limit (AOQL): A plot of the AOQ (Y-axis) versus the incoming lot θ
(X-axis) will start at 0 for θ = 0, and return to 0 for θ = 1 (where every lot is 100% inspected). In
between, it will rise to a maximum. This maximum, which is the worst possible long term AOQ,

is called the AOQL and AOQL =max

θ
AOQ(θ).

Example: suppose we want to design a one-stage sampling plan with infinite N and c = 0 such
that OC(AQL) > 80% where AQL = 0.01 and AOQL < 3%, then what values of n satisfy these
two constraints?
Solution: first as c = 0, then the OC (θ ) = (1 - θ ¿n , therefore
(1 – 0.01 ¿n > 0.8 means 0.99n > 0.8
ln ⁡(0.8)
n ln(0.99) > ln(0.8), then n≤ =22.
ln ⁡(0.99)
Second AOQ(θ ¿ = θ OC(θ ¿ = θ (1 - θ ¿n , as c = 0
Here we need to maximize AOQ and equating to zero by differentiation as:
⁡d ⁡d ⁡1
AOQ(θ ) = (θ (1 - θ ¿n ) = 0, gives θ = .
dθ dθ n+1
⁡1
Therefore, the AOQ attains its maximum at θ =
n+1
 ⁡1 ⁡1 n ⁡nn
Implies AOQL =max (θ (1 - θ ¿n ) = (1 - ¿ =
θ n+1 n+1 ¿¿
n
Choose n so that ⁡n < 0.03, solve by trial and error to give n≥ 12.
¿¿
Therefore, take n to be any value between 12 and 22 in order to satisfy these two constraints.

Example: suppose we want to design a one-stage sampling plan with infinite N and then (a) find
n so that AOQL < 2% if c=0 (b) find AOQL if n=25 and c = 0 (c) find AOQL if n =100 and c =1.
Solution:
(a) c = 0 implies OC (θ ¿ = (1 - θ ¿n , then AOQ (θ ) = θ (1 - θ ¿n
Find its unique maximum by differentiating and equating to zero, then AOQ( θ ) has a maximum
⁡1 n n
at θ = . This implies that AOQL = ⁡n and now choose n so that ⁡n < 0.02 as follows.
n+1 ¿¿ ¿¿
n 16 17 18
n 0.022 0.021 0.019
⁡n
¿¿
Therefore, AOQL < 2% for all n values greater than or equal to 18.
n 25
(b) c = 0 and n =25 then AOQL = ⁡n = ⁡25 that is if we take n = 18, we know < 2% of
¿¿ ¿¿
items sold will be defective. If we take n =25, we know < 1.4% will be defective.
(c) c = 1 and n =100, then OC(θ ¿ = (1 - θ ¿100 + 100 θ (1 - θ ¿99 and AOQ(θ ¿=¿ θ OC(θ ¿
This implies that AOQ(θ ¿=¿ θ (1 - θ ¿100 + 100 θ (1 - θ ¿99
Maximize by differentiating and equating to zero is: 9999 θ2 - 98θ –1= 0, it givesθ =
98 ± √ 98 2+ 4 x 9999
, positive root is θ = 0.01604. Therefore, AOQL = AOQ (0.01604) =
2 x 9999
0.01604 (1 – 0.01604 ¿ 100 + 100(0.06104)(1 -0.06104¿ 99 = 0.0084 = 0.84%.
The Average Total Inspection (ATI): if two sampling schemes have approximately the same
OC function we will usually choose the scheme that is cheaper to administer. This is often the
scheme that has fewer items inspected. As the expected number of items sampled, a low ATI is
preferred. For a one-stage sampling scheme ATI = n and for two-stage sampling scheme we
sample: n items with probability P(d 1 ≤c1) + P(d1> c1) and n+m items with probability
P(c1<d1<c2), therefore, ATI = n[P(d1 ≤c1) + P(d1> c1)] + (n+m) P(c1 < d1 < c2)
= n[P(d1 ≤c1) + P(d1> c1) + P(c1 < d1 < c2)] + m P(c1 < d1 < c2)
= n*1 + m P(c1 < d1 < c2) = n + m P(c1 < d1 < c2)
 Exercises:
1. Write down OC function for a binomial sampling scheme with c = 0 (your answer is based
on a function of n). Differentiate this function and show its derivative at 0 equals -n.
2. Suppose we have a binomial sampling scheme and take c = 0, then how large should we
take n if we want to make sure :
a. We are at least 99% certain rejection of a batch which has 8% defectives?
b. We are at least 90% certain of accept a batch which has 1% of defectives?
3. Suppose we have a binomial sampling scheme with n = 40. what value should we take for
c we want to make sure :
a. We are at least 90% of certain rejecting a batch which has 10% of defectives?
b. We are at least 90% certain of accept a batch which has 1% of defectives?
c. Is it possible to create the scheme so that we are 90% of certain rejecting a batch
which has 10% of defectives and 90% of certain of accepting a batch which has1%
of defectives?
4. write down OC function
a. for binomial sampling scheme with c = 1(your answer will be a function of n)
b. Differentiate this function and show that its derivative at 0 equals 0.
c. Calculate OC (0.01), OC (0.03), OC (0.05), OC (0.10) and OC (0.15) for binomial
sampling scheme with c = 1 and n = 80 and sketch the function.
5. Suppose we have a binomial sampling scheme with c = 1. Show that if we want to make
sure we are at least 95% certain rejecting a batch that has 10% of defectives we need to
take n¿ 46 .
6. Suppose we have a batch size N. What is the OC function of the hyper geometric sampling
scheme with c =1.
7. Consider two acceptance sampling scheme defined by the following rules:
(i). the single stage binomial sampling schemes with n = 59 and c = 0. “Take a sample
size of 59 and accept the batch if it contains no defectives”

(ii). in double – stage binomial sampling scheme with n = 70, c 1 = 0, d1 =2 and c2 =

1:”take a sample of size 70. Then

 If it contains no defectives accept the batch.

 If it contains two or more defectives reject the batch.
 If it contains exactly one defective take a further sample of size m and accept the batch
if this sample has no defectives.
Calculate the OC function for each of these schemes. Calculate OC 1 (0.05), OC1 (0.01) and OC1
(0.10). Find the smallest values of m so that OC 2 (0.05) <0.05 and for this value of m calculate
OC2 (0.01), OC2 (0.05) and OC2 (0.10). Sketch both OC curve.
8. Write down the OC function of the one stage acceptance sampling scheme with c= 0 which
values of n define an admissible scheme in each of the following cases:
(a) RQL = 5%, OC (RQL) < 5% (b) AQL = 1%, OC(AQL)>90% (c) AOQL<1%
4.4 Acceptance Sampling for Variables
4.4.1 Introduction
This chapter summarizes acceptance-sampling techniques for variables sampling plans, which
can be used as alternatives to attribute plans when measurement data are available.
The primary advantage of variables sampling plans is that the sample operating-characteristic
curve can be obtained with a smaller sample size than would be required by an attributes
sampling plan. Thus, a variables acceptance-sampling plan that has the same protection as an
attributes acceptance-sampling plan would' require less sampling. The [Link]
required by a variables sampling plan would probably cost more per observation than the
collection of attributes data. However, the reduction in sample size obtained may more than
offset this increased cost.
A second advantage is that measurement data usually provide more information about the
manufacturing process or the lot than do attributes data. Generally, numerical measurements of
quality characteristics are more useful than simple classification of the item as defective or non-
defective.
A final point to be emphasized is that when acceptable quality levels are yew small, the sample
sizes required by attributes sampling plans are very large. Under these circumstances, there may
be significant advantages in switching to variables measurement.
Thus, as many manufacturers begin to emphasize allowable numbers of defective parts per
million, variables sampling becomes very attractive.
Variables sampling-plans have- several disadvantages: the primary- disadvantage is that the
distribution of the quality characteristic must be known. Furthermore, most standard variables
acceptance-sampling plans assume that the distribution of the quality characteristic is normal. ff
the distribution of the quality characteristic is not normal, and a plan based on the normal
assumption is employed, serious departures from the advertised risks of accepting or rejecting
lots of given quality may be experienced.
The second disadvantage of variables sampling is that a separate sampling plan must be
employed for each quality characteristic that is being inspected. Finally, it is possible that the use
of a variables sampling plan will lead to rejection of a lot even though the actual sample
inspected does not contain any defective items.
4.4.2 Types of Sampling Plans Available
There are two general types of variables sampling procedures: plans that a lot or process
parameter (usually the mean).
Consider a variables sampling plan to control the lot or process fraction nonconforming. Since
the quality characteristic is a variable, there will exist either a lower specification limit (LSL), an
upper specification limit (USL), or both, that define the acceptable values of this parameter.
Suppose that the standard deviation σ known. Under this condition, we may wish to sample from
the lot to determine whether or not the value of the mean is such that the fraction defective p is
acceptable. As described next, we may organize the calculations in the variables sampling plan in
two ways.
Procedure 1: Take a random sample of n items from the lot and compute the statistic is:
X−LSL
Z LSL =
σ
Note that ZLSL in the above equation simply expresses the distance between the sample average,
X and the LSL in standard deviation units. The larger is the value of Z LSL, the farther the sample
average, X is from the lower specification limit, and consequently, the smaller is the lot fraction
defective p. If there is a critical value of p of interest that should not be exceeded with stated
probability, we can translate this value of p into a critical distance--say, k--for Z LSL. Thus, if
ZLSL> k, we would accept the lot because the sample data imply that the lot mean is sufficiently
far above the-LSL to ensure that the lot fraction nonconforming is satisfactory. However, if Z LSL
< k, the mean is too close to the LSL, and the lot should be rejected.
Procedure 2: Take a random sample of n items from the lot and compute Z LSL using the above
equation. Use ZSLL to estimate the fraction defective of the lot or process as the area under the

standard normal curve below ZLSL. Actually, using QLSL = ZLSL

√ n
n−1
as a standard normal

variable is slightly better, because it gives a better estimate of p.) Let ^p be the estimate of p so
obtained. If the estimate ^p exceeds a specified maximum value M, reject the lot; otherwise,
accept it.
The two procedures can be designed so that they have equivalent results. When there is only a
single specification limit (LSL or USL), either procedure may be used. Obviously, in the case of
 USL− X
an upper specification limit, we would compute: ZUSL = instead of using Z LSL =
σ
X−LSL
.
σ
When there are both lower and upper specifications, the M method, Procedure 2, should be used.
Advantages and disadvantage of sampling by variables
Advantages
1. Usually sampling by variable requires smaller sample sizes than an equivalent sampling
by attributes (it is important in about destructive sampling).
2. It is easier to draw inferences about the batch if we are sampling by variables.
Disadvantages
1. We need the normality assumption to hold
2. It involves measuring, this is usually more costly than classifying an item as good/bad
3. It cannot be used if we are dealing with quality that cannot be measured
Exercises
1. Suppose we take n = 100, then find k to make sure (i) we are at least 90% certain of
rejecting a batch which has 10% of defectives (ii) we are at least 90% certain of
accepting a batch which has 3% of defectives.
2. Suppose we take k = 2, then find n to make sure (i) we are at least 95% certain of
rejecting a batch which has 7% of defectives (ii) we are at least 99% certain of
accepting a batch which has 1% of defectives.
 5. Reliability and Life Testing
5.1. Definition of reliability
Reliability: The reliability of a product is the probability that the product will meet certain
specifications for a given period of time. For example, suppose we want a new automobile to
perform without malfunction for a period of 2 years or for 20000 miles. The probability that an
automobile will meet this specification is the reliability of the automobile.
The failure time T: the length of life of a product is the length of time until the product fails to
perform according to specifications. When the product fails to perform according to
specifications, it is said to be failed. The time at which a single product item fails is called the
failure time for the item. For example, the length of life of an abrasive grinding wheel is the
length of time until the wheel fails to perform according to specifications. The specifications
may have been determined by the manufacturer or the user may have written his or her own
specifications. The length of time until failure is called the failure time of the wheel. Therefore,
the Failure Time T of a product is a random variable that represents the length of time that the
product performs according to specifications.
The density function for a product failure time is called failure time distribution. If we denote
the failure time density function by the symbol f(t), then the probability that the product will fail
t0

before time t0 is: P(T ≤ t 0 ) = F(t0) = ∫ f ( t ) dt .

0

Suppose that a product is said to be reliable if it survives time t 0. Then the reliability of the
product that is the probability that it will survive until time t 0 is R(t0) = 1 – F(t0). The reliability,
R(t0), is also called the survival function for the product. The probability of failure time and
reliability is the shaded and un-shaded area under the density function of the following figure,
respectively.
Hazard Rates: the failure time distribution for a product enables us to calculate the probability
F(t0) that an item will fail before time t 0 and the probability R(t0) = 1 –F(t0) that an item survive
until time t0. While for some small change in failure time, denoted ∆ t , the probability that an
item will fail in the interval (t, t + ∆ t ) is called the hazard rates. Therefore, the hazard rate for a
f (t) f (t )
product is defined is: z(t) = = , where f(t) is the density function of the product’s
1−F (t) R (t)
failure time distribution.

Realistically, the failure time distribution is a conceptual relative frequency distribution of the
lengths of life of some group of product items of specific interest, say, those manufactured in a
given week, or year, based on an analysis of sample data, we may select one density function
among those of continuous distributions. The family of density functions represented by the
Weibull distribution is often used for this purpose. The Wibull density function and cumulative
distribution function are, respectively,
α
−t
f(t) = α t α −1 e β
, t ≥0, α > 0, β > 0 and
β
t α
−t
F(t) =∫ f ( z ) dz = 1 - e β .
0

Example: a drill bit has a failure time distribution given by the density:
 2
−t
f(t) = 2 t e 100 , t ≥0, then find (a) F(t) (b) R(t) and z(t) (c) use the result of (b) and find R(8)
100
and z(8).
t α 2
−t −t
Solution: (a) F(t) =∫ f ( z ) dz = 1 - e β = 1 - e 100 , t ≥ 0
0
2 2
−t −t
(b) R(t) = 1 – F(t) = 1- (1 - e 100 ) = e 100 , t ≥0
2
−t
2 100
f (t ) 100 t e 2t t
z(t) = = = = , t ≥0.
R (t) −t
2
100 50
e 100
2
−8 −64
(c) R(t=8) = e 100 = e 100 = e−0.64 = 0.5273

8
z(t) = z(8) = = 0.16
50
Exercises
1. Find the hazard rate for the Weibull distribution for shape parameter 1, 2, and 3.
2. The failure time of a computer disk pack is considered to be an initial failure if it occurs
prior to time t = α and a wear-out failure if it occurs after time t = β . Suppose the failure
1
time distribution during the useful life of the disk pack is given by: f(t) = , α ≤ t ≤ β.
β−α
Then find F(t), R(t) and z(t) and for α =100∧β=1500, what is the reliability of the dick
pack at t= 500 hours?
3. Mechanical engineers conducted a reliability analysis of lead-free solder joints used in
microscopic electronic packages. The minimum time required by cracks in the solder
joints to propagate through the weakest joint (i.e., the failure time) was approximated
using a Weibull distribution with shape parameter 3.5 and the mean failure time of 2370
hours. Then find an estimate of the scale parameter, z(t) and the hazard rate at t= 5000
hours.
5.2 Types of reliability tests
A life testing is an experiment conducted to obtain sample values of the lengths of life of some
product items. Typically, a random sample of n items is placed on test under specified
environmental conditions and left on test until they fail. The recorded times to failure, t 1, t2, …tn
provide a random sample of observations on the length of life T of the product. If for
convenience we let t1 represent the smallest failure time and t n the largest, then the times might
appear as points on a time line. In many situations, life tests are conducted to determine the
quality of a manufactured product prior to sale. Waiting for the last few items in a sample to fsil
can be time-consuming and expensive. To reduce the cost of waiting for some long-life items,
tests are often concluded after a specified length T = t c of test time. When we do this, we say that
the life test is censored at time t c. The second type of censored sampling occurs when we
conclude the testing after a fixed number r of items has failed. There are many other types of life
testing procedures. In life testing with replacement, product items are replaced on the test
equipment as soon as an item fails a procedure that makes maximum use of the test equipment.
Estimating parameters of a Weibull failure time distribution
The method of maximum likelihood can be used to obtain estimates of the shape and scale
parameters of Weibull distribution, but the procedure is difficult. The disadvantage of the
method of maximum likelihood is that the estimates of shape and scale parameters are obtained
by solving a complicated pair of simultaneous nonlinear equations. The advantage of the method
is that when the sample size is large, maximum likelihood estimators possess sampling
distributions that are approximately normal with known mean and variance.
In seated of using the method of maximum likelihood to estimate shape and scale parameters, we
will use the method of least-squares. Recall that the following for the Weibull distribution:
α α
−t −t
F(t) = 1 - e β , t≥ 0 , then the probability of survival to time t is R(t) = e β , , t≥ 0 and

1 α
t
= e β.
R (t)
α
t
Taking the natural logarithms on both sides of this equation, we obtain: –ln(R(t)) = and then
β
Ln(-ln(R(t))) = - ln β + α lnt.
To use the method of least-square, we need to estimate the survival function based on life test
data. One way to do this is to place random sample of n items on life test and count the number
of survivors at the end of one unit of time (for example, a week, or a month), after two units of
time, and, in general, after, i units of time, i = 1, 2, 3, …. An estimate of the proportion of
ni
survivors at time i is ^
R (i) = , where ni is number of survivors at the end of the i th time unit and
n
n is total number of items placed on test. We would calculate ^
R (i) for i = 1, 2,3, …. and then fit
the least-square line as: ln [−ln R
^ ( i ) ]= y ,−ln β=a and α ln(i)¿ bx to the data points (xi, yi), i= 1,

2, 3, …the procedure is outlined as follows.

Assume a sample of n items is placed on test and the numbers of items surviving at the end of
several time intervals are recorded.
ni
Step1: for each time interval, find ^
R (i) = ,
n
Step2: for each time interval, compute yi=ln [−ln ^
R ( i) ]
Step3: for each time interval, compute xi=ln(i)
Step4: using the variables computed in steps 2 and 3, fit the simple linear regression model,
E(y) = a +bx
Step5: the Weibull parameter estimates are: α^ = b^ and ^β = e−a^ .
Example: a manufacturer of hydraulic seals conducted a life test during which the seals
subjected to a fluid pressure that was 200% of the pressure normally maintained in hydraulic
systems in which the seal is used. One hundred seals were placed on test and the number of
survivors was recorded at the end of each day for a period of 7 days, as listed in the following
table.

DAY 1 2 3 4 5 6 7

NUMBER OF SURVIVORS 69 48 33 21 13 7 4

(a) Estimate the shape and scale parameters of weibull distribution by using least-square method.
(b) Find the 95% CI for the shape and scale parameters of Weibull distribution.

Solution:

ni
^
R (i) =
i xi =ln (i) N.S. n -ln( ^
R (i)) yi = ln [−ln R
^ (i)]
1 0.00 69 0.69 0.37 -0.99
2 0.69 48 0.48 0.73 -0.31
3 1.10 33 0.33 1.11 0.10
4 1.39 21 0.21 1.56 0.45
5 1.61 13 0.13 2.04 0.71
6 1.79 7 0.07 2.66 0.98
7 1.95 4 0.04 3.22 1.17
(a) The simple linear regression equation is: yi = - 1.05 + 1.11 xi, then after α^ = b^ and ^β =
− a^
implies that α^ = 1.11 and ^β = e−a = e−(−1.05) = 2.86045.
^
e
Therefore, based on the method of least-square, we would use a Weibull distribution with shape
and scale parameters of 1.11 and 2.86 to model the failure time distribution of the hydraulic
seals. So, find F(t), R(t) and z(t) for the hydraulic seals.
(b) The SPSS output gives us the 95% confidence interval as shown below. Therefore, the
95% CI forα is (1.02, 1.197) and for β is (e 0.928, e 1.171) equals (2.5382, 3.2236).
Model Unstandardized Coefficients t 95.0% Confidence Interval for B
B Std. Error Lower Bound Upper Bound
(Constant) -1.050 .047 -22.132 -1.171 -.928
1
xi 1.109 .035 32.124 1.020 1.197
Exercise: Suppose the life length (in year) of a memory chip in a main frame computer has a Weibull
failure time distribution. To estimate the Weibull parameters, 50 chips were placed on test and the
number of survivors was recorded at the end of each year, for a period of 8 years, as listed in the
following table.

YEAR 1 2 3 4 5 6 7 8

NUMBER OF SURVIVORS 47 39 29 18 11 5 3 1

(a) Estimate the shape and scale parameters of weibull distribution by using least-square method.
(b) Find the 95% CI for the shape and scale parameters of Weibull distribution.
(c) Find F(t), R(t) and z(t) for the memory chip based on the result of (a).