Statistical Quality Control

By statistical Quality Control (SQC) we mean the various statistical methods used for
the maintenance of quality in a continuous flow of manufactured products. In any
manufacturing process, it is not possible to produce goods of exactly the same quality as
some variation is unavoidable. These variations are broadly classified as being due to two
causes, i.e. (i) chance causes and (ii) assignable causes.

Chance Causes of variation: Some variation are inherent in any production and inspection
procedure which may be due to many minor causes which behave in a random manner. The
variation due to these causes is beyond the control of human hand and cannot be prevented or
eliminated under any circumstances. One has to allow such variation, which are called
allowable variation. The range of such variation is known as ‘natural tolerance of the
process’.

Assignable Causes of variation: Sometimes there will be variation which occurs when the
process goes wrong. The causes of such variation are preventable and are known as
preventable variation or assignable cause of variation. These may be due to bad quality raw
material or inexperienced operator or wear and tear of the machine etc.

The main purpose of SQC is to device statistical methods for separating chance
causes of variation from assignable causes of variation, so that an appropriate action can be
taken whenever assignable causes of variation are operating in the process. In other words, an
attempt is made to remove any systematic causes of variation as soon as they occur, so that it
can be assumed that any variation present is only due to chance cause alone.

The main objective of any producer is to maintain a satisfactory level of quality of the
manufactured products so that he can attract more customers. Maintenance of the quality of
the products is done by two types of control known as process control and product control.

Process control: The procedure of controlling the number of defective items during the
course of production itself is known as process control. This is achieved mainly through the
technique of control charts.

Product control: The procedure which ensures that the lots of manufactured goods when they
are delivered to the customers do not contain excessively large proportions of defective goods
is known as product control. This is achieved through sampling inspection plans.

Process control and product control are two distinct problems, because even when the
process is in control, so that the proportion of defectives for the entire output over a long
period may not be large, but an individual lot of items may not be of satisfactory quality.

Variables and Attributes: Many quality characteristics are measured quantitatively, for
example diameter of a ball, length of a screw, chemical composition of a drug, life in hours of
an electric bulb, etc. Such characteristics are known as variables.
 Some quality characteristics cannot be measured and can only classified as good (or
non-defective) and defective. For example a bolt which does not fit the nut is termed as
defective or a nail whose length does not conform to the specification etc are termed as
defectives. Such characteristics are known as attributes.

Process control

We shall first consider how a process is controlled, i.e how the quality of the articles
produced in a factory are controlled during the production stage itself. As mentioned earlier
they are this is done through control charts.

Control Charts: Control chart was discovered by Dr. Walter A. Shewart in 1924. Based on
the theory of probability and sampling, Shewart’s control charts provide a powerful tool of
discovering and correcting the assignable causes of variation outside the ‘stable pattern’ of
chance causes. The main idea in Shewart’s control chart technique is the division of
observations into rational sub-groups. These are to be done in such a manner that the
variation within a sub-group may be attributed entirely to chance causes, while systematic
variation, if it exists, can occur only from one sub-group to another . The basis of selection of
sub-groups is the order of production, as each sub-group will then consist of the product of a
machine or a homogeneous group of machines for a short period of time, so that there cannot
be any remarkable change in the cause system within that period. The use of such sub-groups
would reveal assignable causes of variation that come and go.

Shewart’s control chart technique is a particular diagrammatic method of finding assignable

causes of variation. It consists of (1) a central line denoted as CL, (2) an upper control limit
denoted as UCL, (3) a lower control limit denoted as LCL, together with a number of sample
points.

The UCL and the LCL are usually plotted as dotted lines and central line CL is plotted as a
bold (dark) line. If we focus on some parameter of the 𝜃 of the population and if T be its
corresponding estimator. If the process is in control, then 𝜃 must be the same from subgroup
to subgroup and consequently, the fluctuations in the values of T from sample to sample
should be due to chance variation alone. Then the control limits are defined as follows.

UCL = 𝐸(𝑇) + 3𝑆. 𝐸(𝑇)

CL = 𝐸(𝑇)

LCL = 𝐸(𝑇) − 3𝑆. 𝐸(𝑇).

The value of the statistic T corresponding to each subgroup or sample is then plotted and if
any of these points lie outside the lower or upper control limits, then the process is said to be
out of control, which means the presence of assignable causes of variation.

3-𝝈 control limits

Suppose 𝑥1 , 𝑥2 , ⋯ , 𝑥𝑛 are sample observations and T being a statistic, T = 𝑇(𝑥1 , 𝑥2 , ⋯ , 𝑥𝑛 ) is

a function of these observations. In such a case suppose
2
𝐸(𝑇) = 𝜇 𝑇 and 𝑉𝑎𝑟(𝑇) = 𝜎𝑇2 ,

one may take any value of T lying outside the limits 𝜇 𝑇 − 3𝜎𝑇 and 𝜇 𝑇 + 3𝜎𝑇 as an indication
of the presence of assignable cause of variation. The reason for this is the fact that, if T is
normally distributed and if the process is under control (i.e. only chance cause of variation is
present), then

𝑃𝑟[|𝑇 − 𝜇 𝑇 | ≤ 3𝜎𝑇 ] = 0.9973, approximately.

Even when T is non-normal, then from Chebyshev’s inequality,

𝑉𝑎𝑟(𝑇) 𝜎𝑇2
𝑃𝑟[|𝑇 − 𝜇 𝑇 | ≤ 3𝜎𝑇 ] > 1 − = 1 − 2 = 8/9.
(3𝜎𝑇 )2 9𝜎𝑇

Thus, the probability of a random value of T goes outside the 3-𝜎 limits, i.e. 𝜇 𝑇 ± 3𝜎𝑇 is
only 1 – 0.9973 = 0.0027, which is very small. Hence, if the process is under control, then the
limits of variation should be between 𝜇 𝑇 − 3𝜎𝑇 and 𝜇 𝑇 + 3𝜎𝑇 . Thus the control limits are
defined as

UCL = 𝜇 𝑇 + 3𝜎𝑇

CL = 𝜇 𝑇

LCL = 𝜇 𝑇 − 3𝜎𝑇 .

Thus for a process if all the sample points fall within these limits, then the process is said to
be under control.

Control Charts for Variabless: These charts may be applied to any characteristics that is
measurable. In order to control a measurable characteristic, control on the estimates of both
the location and the dispersion parameters should be exercised. Usually 𝑋̅ and R charts are
used to control the location (mean) and dispersion of the characteristic.

̅ and R charts
𝑿

As no process is perfect enough to produce all items identically, to identify the variation
arising is due to assignable cause, one type of procedure is using 𝑋̅ and R charts. This is
because variation in the location (mean) may relate to machine settings and variation in
dispersion may be due to negligence on the part of the operator. Also even if the mean is
close enough to what is expected, but if the dispersion characteristic is too large also implies
that there is a vide variation in the products, which is not desirable.

Steps for 𝑋̅ and R charts

Measurement. As the conclusions drawn from control chart are mainly dependant on the
measurements made on the products, mistakes in reading measurement or errors in recording
data should be minimized.

3
Selection of samples or subgroups. For making an effective control chart analysis it is
necessary to pay attention to the selection of samples or subgroups. The choice of sample size
n and the frequency of sampling, i.e. the time between the selections of two groups depend
upon the process and no hard and fast rules can be laid down for these. Usually n is taken to
be 4 or 5 and the frequency of sampling may be 15 or 30 minutes until a state of control is
maintained and later on it can be relaxed.
Remark.While collecting data it may not be necessary to go exactly at a fixed interval, as this
may make periodicities to coincide and the workers may be more alert during those instance.
Calculation of 𝑋̅ and R for each sub-group. Let 𝑋𝑖𝑗 , 𝑗 = 1,2, ⋯ , 𝑛 be the measurements on the
ith sample (𝑖 = 1,2, ⋯ , 𝑘). The mean 𝑋̅𝑖 , the range 𝑅𝑖 and the standard deviation 𝑠𝑖 for the ith
sample are given by:
1 1 2
𝑋̅𝑖 = ∑𝑛𝑗=1 𝑋𝑖𝑗 , 𝑅𝑖 = max 𝑋𝑖𝑗 − min 𝑋𝑖𝑗 and 𝑠𝑖2 = ∑𝑛𝑗=1(𝑋𝑖𝑗 − 𝑋̅𝑖 ) . (1)
𝑛 𝑗 𝑗 𝑛

Next, the overall mean 𝑋̅ , mean range 𝑅̅ and mean standard deviation 𝑠̅ as follows.
1 1 1
𝑋̅ = 𝑘 ∑𝑘𝑖=1 𝑋̅𝑖 , 𝑅̅ = 𝑘 ∑𝑘𝑖=1 𝑅𝑖 and 𝑠̅ = 𝑘 ∑𝑘𝑖=1 𝑠𝑖 . (2)

Setting control limits. Suppose 𝜎 is the standard deviation of the population from which the
samples are drawn then standard error (S.E) of the of the sample mean, when samples of size
n are drawn is 𝜎/√𝑛, i.e 𝑆. 𝐸(𝑋̅𝑖 ) = 𝜎/√𝑛, (𝑖 = 1,2, ⋯ , 𝑘).
Also from the sampling distribution of range,
𝐸(𝑅) = 𝑑2 𝜎,

where 𝑑2 is a constant depending on the sample size. Thus an estimate of 𝜎 is obtained from
𝑅̅ as
1 1 1
𝐸(𝑅̅ ) = 𝑘 ∑𝑘𝑖=1 𝐸(𝑅𝑖 ) = 𝑘 ∑𝑘𝑖=1 𝑑2 𝜎 = 𝑘 𝑘𝑑2 𝜎 = 𝑑2 𝜎.

Hence, 𝑅̅ is an unbiased estimate of 𝑑2 𝜎 and consequently, an unbiased estimate of 𝜎 is given

by 𝜎̂ as

𝜎̂ = 𝑅̅ /𝑑2 . (3)

Also, 𝑋̅ gives an unbiased estimate of the population mean 𝜇, since

1 1 1
𝐸(𝑋̅) = 𝑘 ∑𝑘𝑖=1 𝐸(𝑋̅𝑖 ) = 𝑘 ∑𝑘𝑖=1 𝜇 = 𝑘 𝑘𝜇 = 𝜇.

̅ -chart:
Control Limits for 𝑿

Case 1. When the standards are given, i.e. both 𝜇 and 𝜎 are known.

The 3- 𝜎 control limits for 𝑋̅ chart are given by

𝐸(𝑋̅) ± 3𝑆. 𝐸(𝑋̅) = 𝜇 ± 3𝜎/√𝑛 = 𝜇 ± 𝐴𝜎, where 𝐴 = 3/√𝑛.

4
If 𝜇 = 𝜇′ and 𝜎 = 𝜎′, where 𝜇′ and 𝜎′ are known or specified values, then the control limits of
𝑋̅ chart are given by

𝑈𝐶𝐿𝑋̅ = 𝜇 ′ + 𝐴𝜎′

𝐶𝐿𝑋̅ = 𝜇 ′

and 𝐿𝐶𝐿𝑋̅ = 𝜇 ′ − 𝐴𝜎′, where 𝐴 = 3/√𝑛 is a constant depending on n and its tabulated values
are available.

Case 2. When the standards are not given, i.e. both 𝜇 and 𝜎 are unknown.

If both 𝜇 and 𝜎 are unknown, then they are estimated by 𝑋̅ and 𝜎̂ given in (2) and (3) and the
3- 𝜎 control limits are given by
3 3
𝑋̅ ± 3𝜎̂/√𝑛 = 𝑋̅ ± 3(𝑅̅ /𝑑2 )/√𝑛 = 𝑋̅ ± (𝑑 𝑛) 𝑅̅ = 𝑋̅ ± 𝐴2 𝑅̅ , (𝐴2 = 𝑑 𝑛 ). Thus the
2√ 2√
control limits are

𝑈𝐶𝐿𝑋̅ = 𝑋̅ + 𝐴2 𝑅̅

𝐶𝐿𝑋̅ = 𝑋̅ and

𝐿𝐶𝐿𝑋̅ = 𝑋̅ − 𝐴2 𝑅̅. (4a)

Since 𝑑2 is a constant depending on n, 𝐴2 is also a constant depending on n and its tabulated

values are available.

If the estimate of 𝜎 is to be obtained in terms of the sample standard deviation, then

𝐸(𝑠̅) = 𝐶2 𝜎.

Hence, 𝑠̅ = 𝐶2 𝜎̂.
𝑛−2
2 ( 2 )!
Or 𝜎̂ = 𝑠̅/𝐶2 , where 𝐶2 = √ . 𝑛−3 is a constant depending on n.
𝑛 ( )!
2

Then the 3- 𝜎 control limits are given by

3
𝑋̅ ± 3𝜎̂/√𝑛 = 𝑋̅ ± 3(𝑠̅/𝐶2 )/√𝑛 = 𝑋̅ ± ( ) 𝑠̅ = 𝑋̅ ± 𝐴1 𝑠̅ .
𝐶2 √𝑛

Thus the control limits are

𝑈𝐶𝐿𝑋̅ = 𝑋̅ + 𝐴1 𝑠̅

𝐶𝐿𝑋̅ = 𝑋̅ and

𝐿𝐶𝐿𝑋̅ = 𝑋̅ − 𝐴1 𝑠̅ . (4b)

5
 3
The factor 𝐴1 = is dependent on only n and its tabulated values are available for different
𝐶2 √𝑛
values of n.

Control Limits for 𝑹-chart: R- Chart is constructed for controlling the variation in the
dispersion(variability) of the product. The procedure of construction of R-chart is similar to
that of 𝑋̅ chart and involves the following steps.

1. Compute the range 𝑅𝑖 = max 𝑋𝑖𝑗 − min 𝑋𝑖𝑗 , (𝑗 = 1,2, ⋯ , 𝑛) for each sample.
𝑗 𝑗
2. Compute the mean of the sample ranges:
1 1
𝑅̅ = 𝑘 ∑𝑘𝑖=1 𝑅𝑖 = 𝑘 (𝑅1 + 𝑅2 + ⋯ + 𝑅𝑘 ).
3. Computation of control limits. The 3- 𝜎 control limits for R-chart are
𝐸(𝑅) ± 3𝜎𝑅 . 𝐸(𝑅) is estimated by 𝑅̅ and 𝜎𝑅 is estimated from the relation
𝑅̅
𝜎𝑅 = 𝑑3 𝜎̂ = 𝑑3 𝑑 [From(3)],
2
where 𝑑2 and 𝑑3 are constants depending on n. Thus the control limits are
3𝑑 3𝑑
𝑈𝐶𝐿𝑅 = 𝐸(𝑅) + 3𝜎𝑅 = 𝑅̅ + 𝑑 3 𝑅̅ = (1 + 𝑑 3 ) 𝑅̅ = 𝐷4 𝑅̅ (5a)
2 2

𝐶𝐿𝑅 = 𝐸(𝑅) = 𝑅̅ and

3𝑑 3𝑑
𝐿𝐶𝐿𝑅 = 𝐸(𝑅) − 3𝜎𝑅 = 𝑅̅ − 𝑑 3 𝑅̅ = (1 − 𝑑 3 ) 𝑅̅ = 𝐷3 𝑅̅. (5b)
2 2

The values of 𝐷4 and 𝐷3 depend only on n and their tabulated values are available for
different values of n.

However if 𝜎 is known, then

𝑈𝐶𝐿𝑅 = 𝐸(𝑅) + 3𝜎𝑅 = 𝑑2 σ + 3𝑑3 σ = (𝑑2 + 3𝑑3 )σ = 𝐷2 σ (5c)

𝐶𝐿𝑅 = 𝐸(𝑅) = 𝑅̅ and

𝐿𝐶𝐿𝑅 = 𝐸(𝑅) − 3𝜎𝑅 = 𝑑2 σ − 3𝑑3 σ = (𝑑2 − 3𝑑3 )σ = 𝐷1 σ. (5d)

Again 𝐷2 and 𝐷1 depend only on n and their tabulated values are available for different
values of n. Also since the range cannot be negative, 𝐿𝐶𝐿𝑅 should always be greater than or
equal to zero. Thus, in case the calculated value of 𝐿𝐶𝐿𝑅 turns out to be neagative, then it
should be taken as zero.

Remark: It should be noted that the control limits of 𝑋̅ and R charts are based on the
assumption that different samples or sub-groups are of constant size n.

4. Construction of control charts for 𝑿 ̅ and R. Control charts are plotted on a

rectangular co-ordinate axis – vertical scale (ordinate) representing the statistical
measures 𝑋̅ and R and the horizontal scale (abscissa) representing the sample number.
The central line in each chart are shown with solid horizontal lines while the upper

6
 and lower control limits by dotted horizontal lines. The values of 𝑋̅ and R
corresponding to different samples are then plotted in the respective charts.

𝑋̅ Chart R Chart
* 𝑈𝐶𝐿𝑅
𝑈𝐶𝐿𝑋̅
* * *
𝐶𝐿𝑋̅ 𝐶𝐿𝑅
𝑋̅ * * *
* 𝐿𝐶𝐿𝑋̅ 𝐿𝐶𝐿𝑅
1 2 k 1 2 * k

Sample numbers Sample numbers

̅ and R charts.
Criterion for detecting lack of control in 𝑿

The main object of the control chart is to indicate when a process is not in control. The
following situations show lack of control of the process.

1. A point outside control limits. If any point (or points) lie above upper control limit
(UCL) or below lower control limit (LCL) in either of 𝑋̅ or R chart, then the process is
said to be out of control. This is because any point going outside the control limits is
an indication of presence of assignable cause of variation.
2. A run of seven or more points. Although all points are within the control limits, but if
a specific pattern is formed by the points, then it indicates presence of assignable
variation. One such situation is a run of 7 or more points lying above or below the
central line in the control chart indicates shift in the process level. On R-chart a run of
points above the central line indicates increase in the variability while a run below the
central line indicates reduction in variability.
3. One or more points in the vicinity of control limits or a run of points beyond some
secondary limits, example, a run of 2 or 3 points beyond 2-𝜎 limits or a run of 4or 5
points beyond 1-𝜎 limits.
4. The sample points on 𝑋̅ and R charts are too close to the central line indicates
presence of assignable cause. This situation represents systematic differences within
samples and results from improper selection of samples and biases in measurements.

* *
*
* *
*

5. Presence of trends. The trends exhibited by sample points also indicates presence of
assignable cause. Trend may be upward or downward and indicates gradual shift in
the process.

7
 * * *
* * *

* *
* *

6. Presence of Cycles. In some case the cyclic pattern of points in the control chart
indicates presence of assignable cause of variation. Such patterns are due to material
or/and mechanical reasons.

*
* *

* * * *
* *
*

̅ and R charts
Interpretation of 𝑿

In order to judge if a process is under control, 𝑋̅ and R charts should be examined

together and a process is said to be under statistical control only if both the charts show to be
under state of control. If either of the two charts are out of control, then the process is said to
be out of control.

Control chart for standard deviation (or 𝝈-Chart)

Since standard deviation is an ideal measure of dispersion, a combination of control

chart for mean (𝑋̅) and standard deviation (s), known as 𝑋̅ and s-charts (or 𝑋̅ and 𝜎-charts) is
theoretically more appropriate than 𝑋̅ and R charts for controlling process average and
process variability.

In a random sample of size n from a normal population with standard deviation 𝜎,

𝑛−1 2 [(𝑛−2)/2]!
𝐸(𝑠 2 ) = 𝜎 2 and 𝐸(𝑠) = 𝐶2 𝜎, where 𝐶2 = √𝑛 [(𝑛−3)/2]!.
𝑛

Tabulated values of 𝐶2 are available for different values of n.

Also, 𝑉𝑎𝑟(𝑠) = 𝐸(𝑠 2 ) − [𝐸(𝑠)]2

𝑛−1 𝑛−1
= 𝜎 2 − 𝐶22 𝜎 2 = ( − 𝐶22 ) 𝜎 2 .
𝑛 𝑛

𝑛−1
Hence, 𝑆. 𝐸(𝑠) = 𝐶3 𝜎, where 𝐶3 = √ − 𝐶22 .
𝑛

Thus the control limits are

8
𝑈𝐶𝐿𝑠 = 𝐸(𝑠) + 3𝑆. 𝐸(𝑠) = (𝐶2 + 3𝐶3 )σ = 𝐵2 σ

𝐶𝐿𝑠 = 𝐸(𝑠) = 𝐶2 𝜎 and (6)

𝐿𝐶𝐿𝑠 = 𝐸(𝑠) − 3𝑆. 𝐸(𝑠) = (𝐶2 − 3𝐶3 )σ = 𝐵1 σ.

Tabulated values of 𝐵1 and 𝐵2 are available for different values of n.

If the value of σ is not specified or not known, then its estimate is used. Thus the estimate of
σ based on 𝑠̅ defined in (2) is given by σ̂ = 𝑠̅/𝐶2 and the control limits are
𝐶 𝐶
𝑈𝐶𝐿𝑠 = 𝐸(𝑠) + 3𝑆. 𝐸(𝑠) = 𝑠̅ + 3 𝐶3 𝑠̅ = (1 + 3 𝐶3) 𝑠̅ = 𝐵4 𝑠̅ (6a)
2 2

𝐶𝐿𝑠 = 𝐸(𝑠) = 𝑠̅ and (6b)

𝐶 𝐶
𝐿𝐶𝐿𝑠 = 𝐸(𝑠) − 3𝑆. 𝐸(𝑠) = 𝑠̅ − 3 𝐶3 𝑠̅=(1 − 3 𝐶3) 𝑠̅ = 𝐵3 𝑠̅, (6c)
2 2

where 𝐵3 and 𝐵4 are tabulated for different values of n. Also since, s cannot be negative,
hence if calculated value of LCL happens to be negative then it should be taken as zero.

Control Charts for Attributes: When the characteristic under study is of qualitative in nature
or attributes like defective or certain colour etc, then 𝑋̅ and R charts will not be able to deal
such situations. For such situations control charts for attributes are used.

There are two control charts for attributes. They are

(a) Control chart for fraction defective (p- chart) or the number of defective (np or d
chart)
(b) Control chart for number of defects per unit.
Difference between a defective and a defect. An article which does not conform to one or
more of the specifications is called as defective while any instance of articles lack of
conformity to specifications is called a defect.
Control chart for fraction defective ( p- chart).
While dealing with attributes, a process will be adjudged in statistical control if all the
samples or sub-groups are ascertained to have the same population fraction defective P.
If ‘d’ is the number of defectives in a sample of size n, then the sample proportion of
defective is p=d/n. Hence, d is a binomial variate with parameters n and P.
∴ 𝐸(𝑑) = nP and 𝑉𝑎𝑟(𝑑) = nPQ, where Q =1 – P.
1 1 𝑛𝑃𝑄 𝑃𝑄
Thus 𝐸(𝑝) = 𝐸(𝑑/𝑛) = 𝑛 𝐸(𝑑) = P and 𝑉𝑎𝑟(𝑝) = 𝑉𝑎𝑟(𝑑/𝑛) = 𝑛2 𝑉𝑎𝑟(𝑑) = = . (7)
𝑛2 𝑛
Hence, the 3-𝜎 limits for p-chart are given by
𝐸(𝑝) ± 3𝑆. 𝐸(𝑝) = 𝑃 ± 3√𝑃𝑄/𝑛 = 𝑃 ± 𝐴√𝑃𝑄, (8)
where A = 3/√𝑛 .Tabulated values of A are available for different values of n.
Case(i) Standards specified. If the value of P is known or specified as P’, i.e. P=P’, then the
control limits are
𝑈𝐶𝐿𝑝 = 𝑃′ + 𝐴√𝑃′𝑄′ = 𝑃′ + 𝐴√𝑃′(1 − 𝑃′)

9
𝐶𝐿𝑝 = 𝑃′ (8a)
𝐿𝐶𝐿𝑝 = 𝑃′ − 𝐴√𝑃′𝑄′ = 𝑃′ − 𝐴√𝑃′(1 − 𝑃′).
Case(ii) Standards not specified. In this case, let 𝑑𝑖 be the number of defectives and 𝑝𝑖 be the
fraction defective for the ith sample (𝑖 = 1,2, ⋯ , 𝑘) of size 𝑛𝑖 . The population fraction
defective P is estimated by the statistic 𝑝̅ given by:
∑𝑑 ∑ 𝑛𝑖 𝑝𝑖
𝑝̅ = ∑ 𝑛𝑖 = ∑ 𝑛𝑖
. (8b)
𝑖
Then 𝑝̅ is an unbiased estimate of P,since
∑𝑑 ∑ 𝐸(𝑑𝑖 ) ∑ 𝑛𝑖 𝑃
𝐸(𝑝̅) = 𝐸 (∑ 𝑛𝑖 ) = ∑ 𝑛𝑖
= ∑ 𝑛𝑖
= P.
𝑖
Thus the control limits are
𝑈𝐶𝐿𝑝 = 𝑝̅ + 𝐴√𝑝̅ (1 − 𝑝̅)
𝐶𝐿𝑝 = 𝑝̅ (8c)
𝐿𝐶𝐿𝑝 = 𝑝̅ − 𝐴√𝑝̅(1 − 𝑝̅).
Control chart for number of defectives ( d- chart).
If instead of p the sample proportion defective, d the number of defectives in the sample is
used, then the 3-𝜎 control limits are given by:

𝐸(𝑑) ± 3𝑆. 𝐸(𝑑) = 𝑛𝑃 ± 3√𝑛𝑃𝑄 = 𝑃 ± 3√𝑛𝑃(1 − 𝑃). (9)

Case(i) Standards specified. If the value of P is known or specified as P’, i.e. P=P’, then the
control limits are
𝑈𝐶𝐿𝑑 = 𝑛𝑃′ + 3√𝑛𝑃′(1 − 𝑃′)
𝐶𝐿𝑑 = 𝑛𝑃′ (9a)
𝐿𝐶𝐿𝑑 = 𝑛𝑃′ − 3√𝑛𝑃′(1 − 𝑃′).
Case(ii) Standards not specified. Using 𝑝̅ given in (8b) as an unbiased estimate of P, the
control limits are
𝑈𝐶𝐿𝑑 = 𝑛𝑝̅ + 3√𝑛𝑝̅ (1 − 𝑝̅ )
𝐶𝐿𝑑 = 𝑛𝑝̅ (9b)
𝐿𝐶𝐿𝑑 = 𝑛𝑝̅ − 3√𝑛𝑝̅(1 − 𝑝̅ ).
Since p cannot be negative, if LCL given by any of the above formulae turns out to be
negative, then it should be taken as zero.
Remarks: 1. p and d charts for fixed sample size. If the sample size remains same or
constant for each sample, i.e. 𝑛1 = 𝑛2 = ⋯ = 𝑛𝑘 = 𝑛 (say), then using (8b) an estimate of
the population proportion P is given by
∑𝑘
𝑖=1 𝑑𝑖 ∑𝑘
𝑖=1 𝑛𝑝𝑖 1
𝑝̂ = 𝑝̅ = ∑𝑘
= = 𝑘 ∑𝑘𝑖=1 𝑝𝑖 . (9c)
𝑖=1 𝑛 𝑛𝑘
In this case the same set of control limits can be used for all samples inspected and it is
immaterial if one uses p-chart or d-chart.
2. p and d charts for variable sample size. Method 1. If the sample size (n) varies for
different samples, then for p- chart, separate control limits have to be computed for each
sample while the central line is invariant (or constant) whereas for d-chart, the control limits

10
as well as the central line have to be computed separately for each sample. In such situations,
p-chart is relatively simple and hence preferred compared to d-chart.
Method 2. Since 𝑆. 𝐸(𝑝)= √𝑃𝑄/𝑛, it should be noted that smaller the sample size
wider the control band and vice versa. If the sample size does not vary appreciably, then a
single set of control limits based on the average sample size (∑𝑘𝑖=1 𝑛𝑖 /𝑘) can be used. For
practical purposes, this holds good for situations where the largest sample size does not
exceed the smallest sample size by more than 20% of the smallest sample size.
Alternatively, for all sample sizes, two sets of control limits, one based on largest
sample size and another based on smallest sample size can be used. The largest sample size
gives the smallest control band which is called inner band and the smallest sample size gives
the largest control band which is called outer band. Points falling within inner band indicates
the process to be under control while points lying outside the outer band are indicative of
presence of assignable causes of variation which must be searched and rectified. For other
points, i.e.the points which lie in between the inner and outer band, action should be based on
exact control limits.
Method 3. Another procedure is to standardize the variate, i.e. instead of plotting p or
d ,corresponding standardized values given by
𝑝−𝑃′ 𝑝−𝑝̅
Z= or (10)
√𝑃′ 𝑄′ /𝑛 √𝑝̅ (1−𝑝̅ )
according as P is known or not, are plotted. This stabilises the variable and the resulting chart
is called stabilised p-chart or d-chart. In this case the control limits as well as the central line
for p and d-charts are invariant with n (i.e. they are constants independent of n) and are given
by
UCL =3
CL = 0 (10a)
LCL = -3.
Hence, the problem of variable control limits can be solved with little more computational
work.
Interpretation of p-chart.
1. If all the sample points fall within the control limits without exhibiting any specific
pattern, the process is said to be under control.
2. Points above UCL are called high spots. These suggest deterioration in the quality and
should be reported to the production engineers.
3. Points below LCL are called low spots. Such points indicate improvement in the
quality. However before taking this improvement for granted, it should be
investigated whether any slackness in investigation was there or not.
4. When a number of points fall outside the control limits, a revised estimate of P should
be obtained by eliminating allthe points that fall above UCL (it is assumed that points
which fall below LCL are not due to faulty inspection). The standard fraction
defective P should be revisedperiodically in this way.

Remark: The interpretation for number of defectives (d-chart) is same as that for p-chart.

11
Control Chart for Number of Defects per Unit (c – Chart).

Unlike d or np chart which applies to the number of defectives in a sample, c-chart

applies to the number of defects per unit. Sample size for c-chart may be a single unit like a
radio or a pen etc., or group of units or it may be fixed time, area etc. For example in case of
surface defects, area of the surface is the sample size; in case of casting defects a single part
(such as a base plate, side cover etc.) is the sample size. However, defined sample size should
be constant in the sense that different samples have essentially equal opportunity for the
occurrence of defects.

Control limits forc-chart. In many manufacturing or inspection situations, the sample size n,
i.e.the area of opportunity is very large (since the opportunities for defects to occur are
numerous) and the probability p of the occurrence of a defect in any one spot is very small
such that np is finite. In such situation, the pattern of variation of the data can be best
represented by Poisson distribution and hence 3-𝜎 control limits based on Poisson
distribution are used for c-chart. Thus, if c, the number of defects per unit is assumed to
follow a Poisson distribution with parameter 𝜆, then

𝐸(𝑐) = 𝜆 and 𝑉𝑎𝑟(𝑐) = 𝜆.

Thus 3-𝜎 control limits for c-chart are given by

𝑈𝐶𝐿𝑐 = 𝐸(𝑐) + 3√𝑉𝑎𝑟(𝑐) = 𝜆 + 3√𝜆

𝐶𝐿𝑐 = 𝐸(𝑐) = 𝜆 (11)

𝐿𝐶𝐿𝑐 = 𝐸(𝑐) − 3√𝑉𝑎𝑟(𝑐) = 𝜆 − 3√𝜆.

Case (i) Standards specified. If 𝜆′ is the specified value of 𝜆, then

𝑈𝐶𝐿𝑐 = 𝜆′ + 3√𝜆′

𝐶𝐿𝑐 = 𝜆′ (12)

𝐿𝐶𝐿𝑐 = 𝜆′ − 3√𝜆′.

Case (ii) Standards not specified. If the value of 𝜆 is not known, then it is estimated by the
mean number of defects per unit. Thus if 𝑐𝑖 is the number of defects observed on the ith
(𝑖 = 1,2, ⋯ , 𝑘) unit, then an estimate of 𝜆 is given by
1
𝜆̂ = 𝑐̅ = 𝑘 ∑𝑘𝑖=1 𝑐𝑖 . (12a)

It can be easily seen that 𝑐̅ is an unbiased estimate of 𝜆. The control limits are then given by

𝑈𝐶𝐿𝑐 = 𝑐̅ + 3√𝑐̅

𝐶𝐿𝑐 = 𝑐̅ (12b)

12
𝐿𝐶𝐿𝑐 = 𝑐̅ − 3√𝑐̅.

If LCL given by above formulae turns out to be negative, then it should taken as zero as c
cannot be negative.

The central line is drawn at 𝑐̅ and UCL and LCL are drawn accordingly. The observed
numbers of defects on the inspected units are then plotted on the control chart. The
interpretations of c-chart are same as that of p-chart.

c-chart for variable sample size or u-chart.

Instead of inspecting just one unit at regular interval of time, a sample of size n was observed,
the average number of defects in all the units included in the sample is considered. Thus if c
is the total number of defects present in the n observed units, then u =c/n is plotted. If the
sample size varies from time to time, say 𝑛𝑖 is the sample size of the units observed at the ith
instance and 𝑐𝑖 the total number of defects observed in it, then
𝑐𝑖
𝑢𝑖 = (𝑖 = 1,2, ⋯ , 𝑘) (13)
𝑛𝑖

gives the average number of defects per unit of the ith sample.

In this case, an estimate of 𝜆, the mean number of defects per unit in the lot, based on
all the k samples is given by
1
𝜆̂ = 𝑢̅ = 𝑘 ∑𝑘𝑖=1 𝑢𝑖 . (13a)

As it is well known that if 𝑋̅ is the mean of n variables each with variance 𝜎 2 , then the
variance of 𝑋̅ will be 𝜎 2 /𝑛 and standard error (S.E) = 𝜎/√𝑛. Here, u being the mean of n
values of c, Var(u) = Var(c)/n = 𝜆/𝑛. Hence,

S.E(u) = √𝜆/𝑛. (13b)

The 3- 𝜎 control limits for u-chart (or c- chart for variable sample size) are given by

𝑈𝐶𝐿𝑢 = 𝑢̅ + 3√𝑢̅/𝑛

𝐶𝐿𝑢 = 𝑢̅ (13c)

𝐿𝐶𝐿𝑢 = 𝑢̅ − 3√𝑢̅/𝑛.

Even though the central line will remain constant, but the lower and upper control limits will
vary for each sample. The interpretations of these charts are same as that of p-chart.

Applications of c-Chart. Some of the applications of c-chart are as follows.

1. Number of imperfections observed in a bale of cloth or a roll of paper etc.

2. Number of defects of all types observed in aircraft assembly.
3. Number of accidents on a highway or number of epidemiology etc.

13
 Product control
In any manufacturing process, after production of the manufactured goods, the
producer has to ensure that the manufactured goods are according to the specifications of the
customer. This is achieved through product control.

Natural Tolerance Limits: If 𝜇 and 𝜎 are the process average and process standard deviation
respectively, then the limits 𝜇 ± 3𝜎 are called Natural Tolerance Limits. The probability of
an observation lying outside these limits is 0.0027. The width ‘6𝜎’ is the inherent variability
of the process and is known as Natural Tolerance. If 𝜇 and 𝜎 are not known then 𝜇̂ ± 3𝜎̂ are
the estimates of the natural tolerance limits, where

𝜇̂ = 𝑋̅ and 𝜎̂ = 𝑅̅ /𝑑2 or 𝜎̂ = 𝑠̅/𝐶2 .

It might happen that even though the process is in statistical control as exhibited by
control charts, but the customer may not be satisfied by the product. This happens when the
process does not conform to specification limits.

Specification Limits: These are the limits as desired or fixed by the customer for that item.
These specification limits are generally given in terms of upper and lower limits. Thus 𝑋𝑚𝑎𝑥
may denote the upper specification limit (USL) and 𝑋𝑚𝑖𝑛 the lower specification limit (LSL)
for the characteristic under study.

Comparison of natural tolerance limits with specification limits

A decision, whether a process needs adjustment or not, can be made by comparing

natural tolerance limits with specification limits.

(a) Natural tolerance is considerably smaller than specified tolerance, i.e.,

𝑋𝑚𝑎𝑥 − 𝑋𝑚𝑖𝑛 > 6𝜎.

Interpretations.(i) In such a case almost all the manufactured goods will conform to
specifications as long as the process is in statistical control and is appropriately centred as in
positions A, B or C shown in the following figure1.

Figure1: Natural Tolerance smaller than specified tolerance.

𝑋𝑚𝑎𝑥

B UCL

CL
A

LCL
C

𝑋𝑚𝑖𝑛

14
 If the process is operating under one of these conditions, then 𝑋̅ may be permitted to
go out of control, provided it does not go too far, i.e. the distribution of 𝑋̅ may allowed to
fluctuate between positions B and C. This will save money and time for frequent checking of
assignable causes of variation which may cause unsatisfactory product.

(ii) In such situations the time interval between taking successive samples can be appreciably
increased as considerable shifts in the level of working may not result in items falling outside
specification limits.

(iii) The larger the ratio of (𝑋𝑚𝑎𝑥 − 𝑋𝑚𝑖𝑛 ) to natural tolerance 6𝜎, the greater is the
likelihood of getting product without assistance of any control chart. This implies that the
process is too good for the product and it will be economical to examine in conditions of
production like less costly experiment or material or processing etc.

(b) Specification limits coincide with tolerance limits, i.e.

𝑋𝑚𝑎𝑥 − 𝑋𝑚𝑖𝑛 = 6𝜎.
Interpretations. This is an ideal situation and in this case a process in statistical control
obviously implies that the product is meeting the specifications. Here careful centering of the
process at the specification mean is very important so that no item is rejected. Any departure
from the centering would result in some of the product going outside specification limits. As
soon as a control chart detects such departure, immediate action should be taken to maintain
the centering of the process.
(c) Natural tolerance is greater than specified tolerance, i.e.,
𝑋𝑚𝑎𝑥 − 𝑋𝑚𝑖𝑛 < 6𝜎.

Interpretations. (i) If natural tolerances are not included within the specification limits then
even with process under control and process average perfectly centred at the specification
mean, the production of several articles not conforming to the specifications are unavoidable.
In such a situation, a re-adjustment of the process is advisable with respect to process mean
process dispersion or both.

(ii) It would also be worthwhile to investigate the possibility of relaxing the specified
tolerances to the extent of natural tolerances.

If 100% inspection is possible, then defective articles can be removed but in many
situations 100% inspection is not possible (example, if testing is destructive like crackers,
shells, bulbs etc.) then it may not be possible that all the product conform to the
specifications. In such cases the only option is to relax specification limits to tolerance limits.

Acceptance Sampling Inspection Plans

Acceptance sampling plans refer to the use of sampling inspection by a purchaser to

decide whether to accept or reject a lot of a given product. This is known as product control
in statistical quality control terminology. In case of acceptance sampling by attributes, the
decision for accepting or rejecting a given lot is taken on the basis of the fact that the sampled

15
items possess a particular attribute or not. In other words, acceptance sampling by attributes
is an ‘attribute based inspection’ that merely grades the product as defective or non-defective.
If the product is found defective, then it is rejected otherwise accepted.

The following are various definitions involved in product control.

Acceptable Quality Level (AQL). This is the quality level of a good lot. It is the defective
percentage that can be considered satisfactory as a process average, and represents the level
of quality which the producer wants accepted with high probability of acceptance.

A lot with relatively small fraction defective say 𝑝1 that it is not desired to be rejected
more often than a small proportion of times is referred to as a good quality. Usually,

Pr(Rejecting a lot of quality 𝑝1) = 0.05

⇒ 𝑃𝑎 = Pr(Accepting a lot of quality 𝑝1) = 1 - 0.05 = 0.95.

This small fraction defective 𝑝1 is then called as the ‘Acceptance Quality Level’ or AQL and a
lot of this quality is considered as satisfactory by the consumer.

Lot Tolerance Proportion or Percentage Defective (LTPD). The lot tolerance proportion
defective, usually denoted by 𝑝𝑡 is the lot quality which is considered to be bad by the
consumer. This quality level 𝑝𝑡 is the quality level which the consumer regards as rejectable
and 100𝑝𝑡 is called as the Lot Tolerance Percentage Defective. This quality level 𝑝𝑡 is also
known as Rejecting Quality Level (R.Q.L).

Process Average Fraction Defective (𝒑 ̅). 𝑝̅ represents the quality turned out by the
manufacturing process over a long period of time. In industrial production, the quality of any
process tends to settle down to more or less the same for any particular machine. Hence, if
this level is maintained, then it can be concluded that the process is out of any assignable
cause of variation. The process average of any manufactured product is obtained by finding
the percentage of defectives in the product over a fairly long time.

Consumer’s Risk. Any sampling scheme would involve certain risk on the part of the
consumer – in the sense that the consumer has to accept certain percentage of undesirably bad
lots, i.e. lot with fraction defective 𝑝𝑡 or greater fraction defective. More precisely, the
probability of accepting a lot with fraction defective 𝑝𝑡 is termed as consumer’s risk and is
written as 𝑃𝑐 . This is usually denoted by 𝛽. Thus,

Consumer’s risk = 𝑃𝑐 = P[accepting a lot of quality 𝑝𝑡 ] = 𝛽. (14)

Producer’s Risk. The producer has also to face the situation that some good lots are rejected.
The probability of rejecting a lot of quality 𝑝̅ is termed as producer’s risk and is written as 𝑃𝑝 .
This is usually denoted by 𝛼. Thus,

Producer’s risk = 𝑃𝑝 = P[rejecting a lot of quality 𝑝̅ ] = 𝛼. (15)

16
Rectifying Inspection Plans. Harold F. Dodge and Harry G. Romig of Bell Telephone
Laboratories introduced an inspection plan before World War II which is known as
‘Rectifying Inspection Plans’. In this, the rejected lots are inspected and the defective pieces
are replaced by the good ones. These plans enable the manufacturer to have an idea about the
average quality of the product that is likely to result at a given stage of manufacture through
the combination of production, sampling inspection and rectification of rejected lots.

Average Outgoing Quality (AOQ). The average quality of the product after sampling and
100% inspection of the rejected is called Average Outgoing Quality (AOQ).

Average Total Inspection (ATI). The average amount of inspection required for rectifying
inspection plan is called Average Total Inspection (ATI).

Average Outgoing Quality Limit (AOQL). Sometimes the consumer is guaranteed a certain
quality level after inspection – regardless of what quality level is being maintained by the
producer. Let the producer’ fraction defective, i.e., the quality of the lot before inspection be
‘p’, which is known as the ‘incoming quality’ of the lot. The fraction defective of the lot after
inspection is known as ‘outgoing quality’ of the lot. The expected fraction defective
remaining in the lot after application of the sampling inspection plans is called as Average
Outgoing Quality (AOQ) and is denoted by 𝑝̅ . Obviously it is a function of the incoming
probability. The maximum value of 𝑝̅ subject to variations in the incoming lot quality ‘p’ is
called Average Outgoing Quality Limit (AOQL) and is denoted by 𝑝̅𝐿 .

Remark: For rectifying inspection single sampling plan which calls for 100% inspection of
the rejected lot, AOQ values are given by
𝑝(𝑁−𝑛)𝑃𝑎
𝑝̅ = AOQ = , (16)
𝑁

where N is lot size, n is sample size and 𝑃𝑎 is probability of acceptance of the lot.

OC Curve. Operating Characteristic (OC) curve of a sampling plan is a graphic

representation of the relationship between the probability of acceptance 𝑃𝑎 (𝑝) or generally
denoted by 𝐿(𝑝) for variations in the incoming lot quality ‘p’.

Average Sample Number (ASN) and Average Amount of Total Inspection (ATI). The
average sample number (ASN) is the expected value of the sample size required for coming
to a decision about acceptance or rejection of the lot in a sampling plan. Obviously it is a
function of the incoming lot quality p.

The average amount of total inspection (ATI) is the expected number of items inspected per
lot to arrive at a decision in a sampling inspection plan calling for 100% inspection of the
rejected lots. ATI is also a function of the incoming quality p.

It is to be noted that

ATI = ASN + Average size of the remainder in the rejected lots. (17)

17
 Thus, if the lot is accepted on the basis of sampling inspection plan, then ATI =ASN,
otherwise ATI >ASN. In other words ASN gives the average number of units inspected per
accepted lot.

Sampling Inspection Plans for attributes

The commonly used sampling inspection plans for attributes and count of defects are:

(i) Single sampling plan, (ii) Double sampling plan and (iii) Sequential sampling plan.
(i) Single Sampling Plan.
If the decision about accepting or rejecting a lot is taken on the basis of one sample
only, the acceptance plan is described as single sampling plan. It is completely
specified by three numbers N,n and c, where
N is the lot size,
n is the sample size, and
c is the acceptance number, i.e.maximum allowable number of defectives in the
sample.
The single sampling plan is described as follows:
1. Select a random sample of size n from a lot of size N.
2. Inspect all the articles included in the sample. Let d be the number of defectives in
the sample.
3. If 𝑑 ≤ 𝑐, accept the lot, replacing defective pieces found in the sample by non-
defective (standard) ones.
4. If 𝑑 > 𝑐, reject the lot. In this case, the entire lot is inspected and all the defective
items are replaced by standard items.
Flow Chart of single sampling Plan
Inspect a random sample of size n

If 𝑑 ≤ 𝑐 If 𝑑 > 𝑐
Compare

d with c

Accept the lot replacing all the Reject the lot. Resort to
defectives in the sample, if any by 100% inspection of the lot
non-defectives. and replace all defectives
by non-defectives.

Single sampling plan is very easy to understand, design and carry out. The basic
problem in administering a single sampling is the choice of n (sample size) and c (acceptance
number) which have to be determined in advance.
Determination of n and c. The lot size N is generally known. The two unknown quantities
that need to be determined are n and c.

18
 If p is the incoming quality of a lot, then the number of defective pieces in the lot is
Np and non-defectives N – Np =N(1- p). The probability of getting exactly x number of
defectives in a sample of size n will be given by the mass function of a Hyper-geometric
distribution
𝑁𝑝 𝑁−𝑁𝑝
( )( )
𝑥 𝑛−𝑥
𝑔(𝑥, 𝑝) = 𝑁 .
( )
𝑛
Then, probability of accepting a lot of quality p is given by
𝑁𝑝 𝑁−𝑁𝑝
( )( )
𝑃𝑎 (𝑝) = ∑𝑐𝑥=0 𝑔(𝑥, 𝑝) = ∑𝑐𝑥=0 𝑥
𝑁
𝑛−𝑥
. (18)
( )
𝑛
Hence, the consumer’s risk is given by
𝑃𝑐 = P[accepting a lot of quality 𝑝𝑡 ]
𝑁𝑝 𝑁−𝑁𝑝𝑡
( 𝑡 )( )
= ∑𝑐𝑥=0 𝑔(𝑥, 𝑝𝑡 ) = ∑𝑐𝑥=0 𝑥
𝑁
𝑛−𝑥
. (18a)
( )
𝑛
To protect himself against poor quality, the consumer usually demands a small value of 𝑃𝑐 for
given p.
The producer’s risk is given by
𝑃𝑝 = P[rejecting a lot of quality 𝑝̅ ]
= 1 – P[aceepting a lot of quality 𝑝̅]
= 1 − ∑𝑐𝑥=0 𝑔(𝑥, 𝑝̅)
𝑁𝑝̅ 𝑁−𝑁𝑝̅
( )( )
=1 − ∑𝑐𝑥=0 𝑥
𝑁
𝑛−𝑥
. (18b)
( )
𝑛
If the process average fraction defective is 𝑝̅ as claimed by the producer, then the average
amount of total inspection per lot is given by
ATI = 𝑛 + (𝑁 − 𝑛)𝑃𝑝 , (18c)
since n items have to be inspected in each case and the remaining (𝑁 − 𝑛) items will be
inspected only if 𝑑 > 𝑐, i.e. the lot is rejected when the lot quality is 𝑝̅ and the probability for
this is 𝑃𝑝 .
The computation of hyper-geometric probabilities in (18a) and (18b) is extremely
difficult and in practice, the binomial approximation to hyper-geometric distribution is used.

[Binomial approximation to Hyper-geometric distribution. Consider the hyper- geometric

mass function
𝑚 𝑁−𝑚
( )( )
𝑥 𝑛−𝑥
𝑝(𝑥) = 𝑁 . (i)
( )
𝑛

If N and n are large but m or 𝑁 − 𝑚 is relatively small, then hyper-geometric distribution (i)
can be approximated by binomial distribution with parameters m and p= n/N. Under these
conditions,
𝑚 𝑁−𝑚
( )(
𝑥 𝑛−𝑥
) 𝑚 𝑛 𝑥 𝑛 𝑚−𝑥
𝑁 → ( ) (𝑁) (1 − 𝑁) (*).
( ) 𝑥
𝑛

19
If N,m and 𝑁 − 𝑚 are relatively large as compared to n and x, then (i) can be approximated
by binomial distribution with parameters m and p= m/N. Then
𝑚 𝑁−𝑚
( )(
𝑥 𝑛−𝑥
) 𝑛 𝑚 𝑥 𝑚 𝑛−𝑥
𝑁 → ( ) ( 𝑁 ) (1 − 𝑁 ) (**)]
( ) 𝑥
𝑛

Thus, using (*),

𝑁𝑝𝑡 𝑁−𝑁𝑝𝑡
( )( )
𝑃𝑐 = ∑𝑐𝑥=0 𝑥
𝑁
𝑛−𝑥
( )
𝑛

𝑁𝑝𝑡 𝑛 𝑥 𝑛 𝑁𝑝𝑡 −𝑥
= ∑𝑐𝑥=0 ( ) (𝑁) (1 − 𝑁)
𝑥
(𝑁𝑝 )! 𝑛 𝑥 𝑛 𝑁𝑝𝑡 −𝑥
= ∑𝑐𝑥=0 𝑥!(𝑁𝑝 𝑡−𝑥)! (𝑁) (1 − 𝑁) . (18d)
𝑡

Using (**),
𝑁𝑝̅ 𝑁−𝑁𝑝̅
( )( )
𝑃𝑝 = 1 − ∑𝑐𝑥=0 𝑥
𝑁
𝑛−𝑥
( )
𝑛

𝑛 𝑁𝑝̅ 𝑥 𝑁𝑝̅ 𝑛−𝑥

= 1 − ∑𝑐𝑥=0 ( ) ( 𝑁 ) (1 − 𝑁 )
𝑥
𝑛
= 1 − ∑𝑐𝑥=0 ( ) (𝑝̅ )𝑥 (1 − 𝑝̅ )𝑛−𝑥
𝑥
𝑛!
= 1 − ∑𝑐𝑥=0 𝑥!(𝑛−𝑥)! (𝑝̅ )𝑥 (1 − 𝑝̅ )𝑛−𝑥 . (18e)

In most of the practical problems 𝑝̅ is likely to be less than 0.10 and n is likely to be
sufficiently large so that Poisson approximation to Binomial distribution can be used. Thus,
(18e) can be further approximated as
̅
(𝑛𝑝̅ )𝑥 𝑒 −𝑛𝑝
𝑃𝑝 = 1 − ∑𝑐𝑥=0 . (18f)
𝑥!

Consequently,
̅
(𝑛𝑝̅ )𝑥 𝑒 −𝑛𝑝
ATI = 𝑛 + (𝑁 − 𝑛) [1 − ∑𝑐𝑥=0 ]. (18g)
𝑥!

Consumer’s requirement fixes 𝑃𝑐 and 𝑝𝑡 . N is always fixed. For given values of 𝑃𝑐

and 𝑝𝑡 the equation (18a) which involves two unknowns n and c is satisfied by a large
number of pairs of n and c. To safeguard producer’s interest also, out of these possible pairs
the one involving minimum amount of inspection as given by (18c) is chosen. Though
theoretical computations are cumbersome and time consuming, Dodge and Romig have
prepared extensive tables for minimising values of n and c for 𝑃𝑐 = 0.10 and different values
of 𝑝̅.

20
AOQL. If p is the incoming quality, there will be no defective left in a lot of size N if the
sample contains more than c defective, i.e. if x >c. On the other hand, if 𝑥 ≤ 𝑐, the number of
defectives in a lot of size N is Np – x. Thus, the mean value of the number of defectives after
sampling inspection is given by

m =∑𝑐𝑥=0(𝑁𝑝 − 𝑥) 𝑔(𝑥, 𝑝) + ∑𝑁
𝑥=𝑐+1 0. 𝑔(𝑥, 𝑝)

𝑁𝑝 𝑁−𝑁𝑝
( )( )
= ∑𝑐𝑥=0(𝑁𝑝 − 𝑥) 𝑥
𝑁
𝑛−𝑥
.
( )
𝑛

The mean value of the fraction defective after inspection, i.e. AOQ will be
𝑁𝑝 𝑁−𝑁𝑝
𝑚 1 ( )( )
AOQ = 𝑝̃ = = ∑𝑐𝑥=0(𝑁𝑝 − 𝑥) 𝑥
𝑁
𝑛−𝑥
𝑁 𝑁 ( )
𝑛

𝑁𝑝 𝑁−𝑁𝑝
𝑥 ( )( )
= ∑𝑐𝑥=0 (𝑝 − 𝑁) 𝑥
𝑁
𝑛−𝑥
. (19)
( )
𝑛

Subject to variation in p, 𝑝̃ has a maximum value , say 𝑝̃𝐿 which is termed as AOQL.

AOQL takes care of consumer’s interest.

Given N and 𝑝̃𝐿 (AOQL) , equation (18a) contains two unknowns n and c. It is possible
to select several pairs of values of n and c that will give 𝑝̃ as defined in (19) having
approximately the same value of 𝑝̃𝐿 . As a safeguard to producer’s interest, it is necessary to
select a pair that minimizes ATI as defined in (18c). Dodge and Romig have prepared
extensive AOQL tables for minimising values of n and c.

OC Curve. The OC curve for the incoming probability ‘p’ is given by

𝑁𝑝 𝑁−𝑁𝑝
( )( )
𝑃𝑎 (𝑝) = 𝐿(𝑝)= ∑𝑐𝑥=0 𝑔(𝑥, 𝑝) = ∑𝑐𝑥=0 𝑥
𝑁
𝑛−𝑥
. (20)
( )
𝑛

When 𝑝 < 0.10, a good approximation to (20) is given by the first (𝑐 + 1) terms of
𝑛 𝑛 𝑁𝑝
the binomial expansion of [(1 − 𝑁) + 𝑁] , i.e.

𝑁𝑝 𝑛 𝑥 𝑛 𝑁𝑝−𝑥
𝐿(𝑝) = ∑𝑐𝑥=0 ( ) (𝑁) (1 − 𝑁) . (20a)
𝑥
𝑛
When 𝑝 < 0.10 and also 𝑁 < 0.10, a good approximation to (20) is given by Poisson
distribution, i.e.

𝐿(𝑝) = ∑𝑐𝑥=0[𝑒 −𝑛𝑝 (𝑛𝑝)𝑥 /𝑥!]. (20b)

(ii) Double Sampling Plan. In this method, a second sample is permitted if the first
sample fails, i.e. if the data from the first sample is not conclusive on either side
(about accepting or rejecting the lot), then a definite decision is taken on the basis

21
 of the second sample. Such rectifying double sampling inspection plan is as
described below. The following notations are used.

N =Lot size from which samples are taken; 𝑛1 = Size of sample 1; 𝑛2 = Size of sample 2.

𝑐1 = Acceptance number for first sample, i.e. maximum permissible number of defectives
in first sample if the lot is to be accepted without taking another sample.

𝑐2 = Acceptance number for samples 1 and 2 combined, i.e. maximum permissible

number of defectives in combined samples if lot is to be accepted.

𝑑1 = Number of defectives in sample 1;

𝑑2 = Number of defectives in sample 2.

Procedure:

(i) Take a sample of size 𝑛1 from the lot of size N.

(ii) If 𝑑1 ≤ 𝑐1 , accept the lot replacing the defectives found in the sample by
nondefectives.
(iii) If 𝑑1 > 𝑐2 , reject the whole lot. Inspect the lot 100%, replacing all the bad items
by good ones.
(iv) If 𝑐1 + 1 ≤ 𝑑1 ≤ 𝑐2 , take a second sample of size 𝑛2 from the remaining lot.
(v) If 𝑑1 + 𝑑2 ≤ 𝑐2 , accept the lot, replacing defective items by standard ones.
(vi) If 𝑑1 + 𝑑2 > 𝑐2 , reject the lot. Inspect the rejected lot 100%, replacing all the
defective items by good ones.
Flow Chart of double sampling Plan

Inspect 𝑛1 items of
the first sample

If 𝑑1 ≤ 𝑐1 Compare If 𝑑1 > 𝑐2

𝑑1 𝑐1 , 𝑐2

If 𝑐1 + 1 ≤ 𝑑1 ≤ 𝑐2
Accept the lot, replacing the Inspect 𝑛2 items of the Reject the lot
defective items in the second sample
sample by non-defectives
D=𝑑1 + 𝑑2

Compare

𝐷 𝑐2
If 𝐷 > 𝑐2
If 𝐷 ≤ 𝑐2

22
O.C. Curve of Double Sampling Plan. The lot will be accepted under the following mutually
exclusive ways:
(i) 0 ≤ 𝑑1 ≤ 𝑐1; (ii) 𝑑1 = 𝑐1 + 1, 𝑑2 ≤ 𝑐2 − 𝑐1 − 1;(iii) 𝑑1 = 𝑐1 + 2, 𝑑2 ≤ 𝑐2 − 𝑐1 − 2 ;
⋮ ⋮
( .) 𝑑1 = 𝑐2 , 𝑑2 = 0.
Hence, by addition theorem of probability, the probability of acceptance for a lot of
incoming quality ‘p’ is given by:
𝑐1 𝑐2
𝑃𝑎 (𝑝) = ∑𝑥=0 𝑔(𝑥, 𝑝) + ∑𝑥=𝑐 1 +1
𝑔(𝑥, 𝑝) . ℎ(𝑦, 𝑝|𝑥),
where 𝑔(𝑥, 𝑝) is the probability of finding x defectives in the first sample and ℎ(𝑦, 𝑝|𝑥) is
the conditional probability of finding y defectives in the second sample under the condition
that x defectives have already appeared in the first sample. Thus,
𝑁𝑝 𝑁−𝑁𝑝
( )( )
𝑛 −𝑥
𝑔(𝑥, 𝑝) = 𝑥 𝑁 1 and
( )
𝑛1
𝑁𝑝−𝑥 𝑁−𝑛1 −(𝑁𝑝−𝑥)
( )( )
𝑦 𝑛2 −𝑦
ℎ(𝑦, 𝑝|𝑥) = 𝑁−𝑛1 .
( )
𝑛2
𝑁𝑝 𝑁−𝑁𝑝 𝑁𝑝 𝑁−𝑁𝑝 𝑁𝑝−𝑥 𝑁−𝑛1 −(𝑁𝑝−𝑥)
( )( ) ( )( ) ( 𝑦 )( )
𝑐1 𝑥 𝑛1 −𝑥 𝑐2 −𝑥 𝑐2 𝑥 𝑛1 −𝑥 𝑛2 −𝑦
∴ 𝑃𝑎 (𝑝) = ∑𝑥=0 𝑁 + ∑ 𝑦=0 ∑ 𝑥=𝑐1 +1 𝑁 . 𝑁−𝑛1 (21)
( ) ( ) ( )
𝑛1 𝑛1 𝑛2
= 𝑃𝑎1 (𝑝) + 𝑃𝑎2 (𝑝), (say),
where 𝑃𝑎1 (𝑝) and 𝑃𝑎2 (𝑝) are the probabilities of acceptance on the basis of first and second
samples respectively.
Consumer’s and Producer’s Risks. The consumer’s risk is given by:

𝑃𝑐 = P[accepting a lot of quality 𝑝𝑡 ] = 𝑃𝑎 (𝑝𝑡 )

𝑁𝑝𝑡 𝑁−𝑁𝑝𝑡 𝑁𝑝 𝑁−𝑁𝑝𝑡 𝑁𝑝𝑡 −𝑥 𝑁−𝑛1 −(𝑁𝑝𝑡 −𝑥)
( )( ) ( 𝑡 )( ) ( 𝑦 )( )
𝑛1 −𝑥 𝑛1 −𝑥 𝑛2 −𝑦
= ∑𝑐𝑥=0
1 𝑥
𝑁 + ∑ 𝑐2 −𝑥 𝑐2
𝑦=0 ∑ 𝑥=𝑐1 +1
𝑥
𝑁 . 𝑁−𝑛1 ; (21a)
( ) ( ) ( )
𝑛1 𝑛1 𝑛2
and producer’s risk is given by
𝑃𝑝 = P[rejecting a lot of quality 𝑝̅ ]
= 1 – P[aceepting a lot of quality 𝑝̅]
𝑁𝑝̅ 𝑁−𝑁𝑝̅ 𝑁𝑝̅ 𝑁−𝑁𝑝̅ (𝑁𝑝̅ −𝑥)(𝑁−𝑛1 −(𝑁𝑝̅ −𝑥))
( )( ) ( )( ) 𝑦
𝑛1 −𝑥 𝑛1 −𝑥 𝑛2 −𝑦
=1– ∑𝑐𝑥=0
1 𝑥
𝑁
𝑐2 −𝑥 𝑐2
− ∑𝑦=0 ∑𝑥=𝑐1 +1 𝑥
𝑁 . 𝑁−𝑛1 . (21b)
( ) ( ) ( )
𝑛1 𝑛1 𝑛2
ASN and ATI of Double Sampling Plan. In a double sampling plan, the number of items
inspected for a lot is either 𝑛1 , when the lot is accepted or rejected on the basis of the first
sample or (𝑛1 + 𝑛2 ) when a second sample of size 𝑛2 is drawn. Thus the expected sample
size for a decision is given by:
ASN = 𝑛1 𝑃1 + (𝑛1 + 𝑛2 )(1 − 𝑃1 ) = 𝑛1 𝑃1 + 𝑛1 − 𝑛1 𝑃1 + 𝑛2 (1 − 𝑃1 )
= 𝑛1 + 𝑛2 (1 − 𝑃1 ),
where 𝑃1 is the probability of a decision (acceptance or rejection of the lot) is taken on the
basis of the first sample.
However, in a double sampling plan, in which the rejected lots are inspected 100%,
the average total inspection (ATI) per lot is given by

23
 ATI = 𝑛1 𝑃𝑎1 + (𝑛1 + 𝑛2 )𝑃𝑎2 + 𝑁(1 − 𝑃𝑎 ), (22)

since (i) only 𝑛1 items would be inspected if the lot is accepted on the basis of the first
sample and its probability is 𝑃𝑎1 (𝑝),

(ii) (𝑛1 + 𝑛2 ) items would be inspected if the lot is accepted on the basis of the second
sample and its probability is 𝑃𝑎2 (𝑝), and

(iii) the entire lot of N items would be inspected if the lot is rejected and the
probability is 1 − 𝑃𝑎 (𝑝).

Also since, 𝑃𝑎 = 𝑃𝑎1 + 𝑃𝑎2 ⇒ 𝑃𝑎2 = 𝑃𝑎 − 𝑃𝑎1 , from (22),

ATI = 𝑛1 𝑃𝑎1 + (𝑛1 + 𝑛2 )(𝑃𝑎 − 𝑃𝑎1 ) + 𝑁(1 − 𝑃𝑎 )

= 𝑛1 𝑃𝑎1 + (𝑛1 + 𝑛2 )(𝑃𝑎 + 1 − 1 − 𝑃𝑎1 ) + 𝑁(1 − 𝑃𝑎 )

= 𝑛1 𝑃𝑎1 + (𝑛1 + 𝑛2 )[(1 − 𝑃𝑎1 ) − (1 − 𝑃𝑎 )] + 𝑁(1 − 𝑃𝑎 )

= 𝑛1 𝑃𝑎1 + 𝑛1 (1 − 𝑃𝑎1 ) + 𝑛2 (1 − 𝑃𝑎1 ) + (𝑁 − 𝑛1 − 𝑛2 )(1 − 𝑃𝑎 )

= 𝑛1 + 𝑛2 (1 − 𝑃𝑎1 ) + (𝑁 − 𝑛1 − 𝑛2 )(1 − 𝑃𝑎 ). (22a)

Single Sampling vs. Double Sampling Plans.

1. Single sampling plans are simple, easy to design and administer, and since each
sample can be plotted on a control chart, maximum information concerning the lot can
be obtained.
2. A very important advantage of double sampling over single sampling seems to be
psychological. To a layman, it seems unfair to reject the lot on the basis of one sample
alone and appears more convincing to say that the lot was rejected after inspecting
two samples. Moreover, in double sampling no lot can be rejected without finding at
least two defectives in the sample taken from it – thus lots of marginal quality always
get a second chance of being accepted.
3. The conditions under which the sampling schemes are operated, it has been found that
on the average the double sampling scheme involves less amount of inspection as
compared to single sampling plan for the same quality assurance. Usually double
sampling requires 25% to 33% less inspection on the average.
The general reduction in the amount of reduction afforded by double sampling plan is
one of the strongest advantages. This does not mean that double sampling could be
less costly as compared to single sampling plan. The double sampling plan being
more complicated and the necessity of inspecting second sample being unpredictable,
the unit cost of inspection for a double sampling plan may be higher than that of a
single sampling plan.

24
4. The operating characteristic(OC) curves of double sampling plan are generally steeper
than that of single sampling plan, i.e.the discriminatory power of double sampling
plan is a bit higher than that of a single sampling plan.

25