5.

Control Charts for Variables

Although control chart is sometimes considered as one of the basic quality control
techniques, it is undoubtedly the most important statistical technique in quality
engineering and it has attracted attention from quality professionals in various industries.
We usually distinguish between control charts for variables and control charts for
attributes. The former refers to measurements that can take any value and the latter refers
to measurements in forms of count or percentage. In this chapter, control charts for
variables are introduced after a general discussion of the concept of process and product
variation.

5.1. The concept of variation

The variation is considered a law of nature as no two items can be exactly the same. The
variation may be large or small, but it is always there. There are basically three types of
variation that we usually talk about:

(1) Within-piece variation - it is the difference in quality at different places for the
same item;

(2) Piece-to-piece variation - it is the difference between any two items that are
considered similar;

(3) Time-to-time variation - it is the difference between the items produced at

different time as the quality might vary according to time.

The variation can be caused by many reasons. First, it can be due to the variation of the
equipment as they might be of different age, subject to different vibration and hydraulic
and electrical fluctuations, for example. Even supposedly identical machines will have
different capabilities. Second, the variation of product quality might be due to the
variation of materials which is someone else's final product and their products are not of
the same quality. It is clear that with materials of different quality, the finished products
will be of different quality. Third, the variation can be due to the environment. Factors
such as temperature, pressure, humidity, etc., can all contribute to the variation of product
quality. Finally, another major source of variation is of human nature. People with
Quality Engineering (SYE3102) 40
[Link]@[Link]

different knowledge and experience, for example, will perform differently and they will
make different decisions. That will introduce additional variation to product quality.

These factors are best summarized in a cause-effect diagram, as an application of that.

Equipment Material
Precision

Strength
Wear Ductility
etc

etc
Positioning Thickness
VARIATION

Education Humidity
etc

etc
Experience

Temperature

Health Pressure

People Environment

Figure 5.1. Common sources of variation that

make similar products different.

Because variation is always present, we then need to find a way to help us to judge
whether a certain variation is normal or abnormal. If the sources of variation fluctuate in a
natural manner, no corrective actions need to be taken. These causes of variation are
called chance causes and they are inevitable. The causes of variation that are large in
magnitude, and therefore readily identified are called assignable causes. When a process
shifted because of the assignable causes, actions have to be taken to identify the cause and
adjust the process to normal again.

Control charts can be used to monitor the process to see if there is any assignable cause.
Control charts show whether or not the process is in a stable state by adding statistically
Quality Engineering (SYE3102) 41
[Link]@[Link]

determined control limits to the run chart. The control chart method for variables is a
means of visualizing the variations that occur in the central tendency and dispersion of a
set of observations. It also serves as a graphical record for the quality of a particular
characteristic.

A control chart is usually characterized by a series of plotted points connected by solid

lines. In addition to the central line which indicates the average value, there are also two
other lines. One is the Upper Control Limit (UCL) and the other is the Lower Control
Limit (LCL). When the process is in control, the plotted points should be concentrated
close to the central line. Whenever there is a point that falls outside the control limits, the
process is considered to be out-of-control and action has to be taken to modify the
process.

25
Measurement value

Out-of-control
20
UCL 18.78

15 MEAN 15.27

LCL 11.75

10
0

10

20

30

sample number

Figure 5.2. Control chart can be used to help making judgment

regarding the state of process - when there is a significant shift, the
process is considered to be out-of-control.

The determination of control limits is based on sound statistical principles. The general
form of the control limits for normally distributed quantities can be given in the following
form
Quality Engineering (SYE3102) 42
[Link]@[Link]

UCL    z / 2  

and

LCL    z / 2  

where z/2 is the value that satisfies

 ( z / 2 )  1   / 2

In the above,  is the mean and  is the standard deviation of the quantity we are plotting.

This implies that when the process is in control we will observe an out-of-control signal
with probability (which is /2+/2, one for each side of the control limit). This is the
probability of false alarm and usually this is set to be very low. In that case, when there is
a point that falls outside of control limits, we can with very high certainty to say that the
process has shifted.

In practice, control charts are made very easy as the z/2 value is usually selected to be
three. Hence, we use the mean plus/minus three times standard deviation to calculate the
control limits. This will be described in the following.
Quality Engineering (SYE3102) 43
[Link]@[Link]

5.2. Control charts for the central tendency - X -chart

The most commonly used control chart is the X -chart. An example of a X -chart is given
in Figure 5.3. using the following data set:

Output from the five plants (P1 to P5) of the same capacity.

Day P1 P2 P3 P4 P5 Mean Range

1 9.5 10.6 9.2 9.1 8.5 9.4 2.1
2 9.1 9.3 10.7 8.7 9.8 9.5 2
3 11.2 9.8 8.9 9.6 10.5 10 2.3
4 10.7 8.9 9.2 8.6 8.6 9.2 1.1
5 9.3 8.8 11.4 10.2 10.7 10.1 2.6
6 8 10.5 9.4 10.9 9.7 9.7 2.9
7 9.8 12 10.6 9.3 11.3 10.6 2.7
8 11 11.5 9.5 10.5 11.1 10.7 2
9 10.9 9.5 8.8 8.2 11.3 9.7 3.1
10 11.5 10.2 9.6 10.4 9.1 10.2 2.4
11 11.1 10.6 10.8 9.3 11.5 10.7 2.2
12 9.6 11.4 9.3 11.1 10.5 10.4 2.1
13 9 9.2 9.1 9.9 10.8 9.6 1.8
14 10.7 10.7 8.2 10.2 7.6 9.5 3.1
15 10.8 11.1 10.8 9.8 13.1 11.1 3.3
16 10.9 10.3 9.1 8.5 7.4 9.2 3.5
17 11.3 12.3 10.5 10.7 9.9 10.9 2.4
18 9.5 8.9 10.1 9.5 10.1 9.6 1.2
19 9.3 7.9 9.8 9.1 10 9.2 2.1
20 9.2 10.3 10.5 8.7 9.4 9.6 1.8
21 9.4 10.8 9.6 10.4 10 10 1.4
22 9.6 11.7 9.1 9.5 8.7 9.7 3
23 10.7 9.1 10.4 14.1 9.9 10.8 5
24 12.1 9.9 11.7 10 10.4 10.8 2.2
25 10.2 9 9.1 11.3 10.8 10.1 2.3

The formulas commonly used for the calculation of the control limits are

UCL X  X  A2 R

and

LCL X  X  A2 R
Quality Engineering (SYE3102) 44
[Link]@[Link]

where X is the mean of all observations and R is the average of the ranges. The values of
A2 can be taken from standard tables.

12

UCL 11.4

11
Production

10 MEAN 10.0

LCL 8.6

8
0

10

15

20

25

30

Row Numbers

Figure 5.3. A x-bar chart for average daily production volume.

The control limits we have used is the 3-sigma limits. That is, it is the mean plus/minus
three times standard deviation as we described before. First, we can see that an estimate of
the mean is X .

An estimate of the standard deviation of the population, , is R/d2 and the quantity plotted
is the average of n measurement values, where n is called subgroup size. Hence the
standard deviation for the measurement value is

x  / n

Instead of doing the calculation, A2 is introduced and tabulated. It is defined as

A2  3 / (d2 n)
Quality Engineering (SYE3102) 45
[Link]@[Link]

This implies that

A2 R  3 x

In this example the subgroup size is 5 which is the most commonly used subgroup size
although other number can be used as well. The selection of this rational subgroup varies
in practice. It has been used to justify the use of standard control limits as they are based
on normal distribution. However, when the process characteristic is fairly normally
distributed, even a subgroup of size one can be used and the average is the actual
measurement value. Sometimes it is recommended to plot individual measurement
because by subgrouping, some information is lost.

When only chance causes are present, the process is considered to be in a state of
statistical control and it is stable and predicable. However, when an assignable cause of
variation is also present, the variation will be excessive and the process is considered as
out of control.

It should be pointed out that when these control limits are used, the associated probability
of false alarm

P(out of control signal / processisincontrol )

 2  [1  P( z  3)]  2  0.00135  0.0027

This is regarded as a good level economically as it is important in practice not to have too
many false signals. The 3-sigma limits are also easy to compute and hence suitable for
practical applications.

We usually use  to denote the probability of false alarm which is the probability of Type
I error is

  P(Type I Error )  P( False Alarm)

 P (out of control signal / processisincontrol )

The general formula for determining the control limits when there are n observations in
each sample is
Quality Engineering (SYE3102) 46
[Link]@[Link]

UCL    z / 2   / n

and

LCL    z / 2   / n

Note that the factor 1/ n is because of the n observations in the sample and we are
monitoring the average of the observations. The sample standard deviation for the sample
mean is *(1/ n ) where  is the standard deviation for the individual value.

The above formula can be used when the mean and standard deviation are known. In
practice, they have to be estimated based on sample observations and the values can be
inserted in the formula. However, the sample size must be large enough for the estimates
to be accurate.

5.3. Control charts for the range

The most common type of control charts for variables is the X -chart. This type of control
chart is usually associated with another chart for the variation of the process, R-chart.
Suppose that the sample size is n, then based on the average measurement, we cannot only
plot a X -chart, but also plot a R-chart by taking the difference between the maximum and
minimum of the measurement in the sample.

The formulas for the calculation of the control limits of the R-chart are

UCLR  D4 R

and

LCL R  D3 R

where the values for D3 and D4 can be taken from standard tables.

The statistical properties for the subgroup range is difficult to derive analytically. It can be
noted that it does not follow normal distribution and it is not symmetrical. We can see it as
the range is always positive. That is why numerical tables are always used.
Quality Engineering (SYE3102) 47
[Link]@[Link]

Today, many statistical quality control software packages are making the implementation
and interpretation much easier. Most of the SPC software produce control charts based on
the standard formulas.

UCL 5.21
5

4
Column 3

MEAN 2.46
2

0
0

10

15

20

25

30

Row Numbers

Figure 5.4. The range chart for the daily

production data in the previous section.

Usually X-bar and R charts are used together. A typical x-R chart is shown in Figure 5.5.
Quality Engineering (SYE3102) 48
[Link]@[Link]

25
UCL 23.8

20
Average

MEAN 17.4

15

LCL 11.0
10

30
0

10

20

30
UCL 23.5
20
Range

MEAN 11.1
10

0
0

10

20

30

Sample number

Figure 5.5. A x-bar together with an R chart and plotting can be

continued for both chart when new data are obtained.

R-chart is a type of chart that is useful to monitor within-sample variation of process

characteristic as the range is an indicator of the variation. However, some people prefer to
use the estimate of standard deviation and by charting the sample standard deviation, s,
another type of chart can be constructed.

The formulas for the computation of the control limits are

UCL X  X  A3 s, LCL X  X  A3 s,
Quality Engineering (SYE3102) 49
[Link]@[Link]

for x-bar chart based on s and

UCLs  B4 s, LCL s  B3 s

for s-chart. In the above formulas

X  i 1 X i , s  i 1 si
n n

and si is the standard deviation of the sample i and n is the number of samples. The values
A3, B3 and B4 can be obtained from standard quality tables.

5.4. Control chart signals

A point outside of the control limits is the most common out-of-control signal. However,
the process might be considered to be out-of-control for some other reasons. By out-of-
control we generally mean that the process is affected by an assignable cause which makes
the process significantly different from a normal behaviour.

Some commonly used out-of-control signals are

(1) Two out of three points in a row in zone A on the same side of the central line

When the process is in control, the chance that a point will fall in Zone A is less than 5%.
That there will be two points out of three in a row in Zone A on the same side of the
central line will be very small. When this happens, we can confidently say that the process
has shifted and an investigation is needed.
Quality Engineering (SYE3102) 50
[Link]@[Link]

25

out-of-control
UCL 21.24
20
Measurement value

UWL 19.28

15 MEAN 15.4

LWL 11.44
10
LCL 9.48

5
0

10

20

30
sample number

Figure 5.6. When there are two points out of latest three
the process can be considered out-of-control.

(2) Six points in a row that are increasing/decreasing

When there are six or more points what are steadily increasing or decreasing, there is an
obvious trend and it can be suspected that the process id drifting. Early action could be
taken to actually prevent the process to deteriorate further.

(3) Seven points in a row on one side of the central line

When there are seven or more points in a row that are on the same side of the central line,
we can say that the process is clearly changed although the change might be small.
Sometimes this may be due to wrongly set control limits and some investigation is
necessary.

There are also some other commonly used "run-rules" and the application of those run
rules varies depending on the need for the identification of other type of changes. The
basic one-point outside the control limits is always the basic rule and it is useful to detect
sudden process changes. Other type of changes such as slow increasing/decreasing trend,
cyclic behaviour, mixture of populations, etc, can also be of importance in decision
making.
Quality Engineering (SYE3102) 51
[Link]@[Link]

25

out-of-control
UCL 21.26
20
Measurement value

UWL 19.30

15 MEAN 15.38

LWL 11.46
10
LCL 9.49

5
0

10

20

30
sample number

Figure 5.7. When there are six points in a row that are
decreasing/increasing the process can be considered out-of-control.

5.5. Revising the control limits

When the process is subject to permanent change, there is a need to revise the control
limits. New control limits should be based on the data after the latest major modification
of the process. This could be the case when the process had deteriorated and the action
taken to restore it to its original level has led to significant change of the process.

For a new process, initial data points are used to derive a trial control limits which have to
be revised by removing all out-of-control points. Revising control limits is necessary at
the beginning of the implementation of a new control chart or a control chart for a new
type of measurement. Usually the first control limits are based on few measurement values
which may lead to many out-of-control signals. When the process has been stabilized,
false alarm should rarely occur.
Quality Engineering (SYE3102) 52
[Link]@[Link]

6. Control Charts for Attributes

In this chapter, we introduce some commonly used control charts for attributes. Attribute
data in general are those related to the counting of the number of nonconforming items or
the number of nonconformities. First the p-chart for percentage of nonconforming in a
sample is studied. An associated chart, np-chart for the number of nonconforming items in
a sample is also presented. This is followed by the c-chart which is useful for monitoring
the number of nonconformities in a sample. Finally, we discuss some common problems
when standard formulas are used blindly.

6.1. The p- and np-chart

The information about the product conformance is usually given as count numbers or
percentages. The p chart is usually used to measure the percentage nonconforming in a
sample.

The control limits for p chart can be derived in the following manner. It is well-known
that the number of nonconforming items in a sample of size n follows binomial
distribution with parameter p. The mean is np and the standard deviation is np(1-p). Hence
the control chart for p can be obtained as

UCL  p  3 p (1  p ) / n
CL  p

LCL  p  3 p (1  p ) / n

Usually an estimate of p has to be used because it is seldom known. An estimate that is

commonly used is the total number of nonconforming items divided by the total number of
items.

It happens frequently that LCL<0. As this has no practical meaning, LCL is set to be zero
in this case. We can also say that there is no lower control limit because a percentage
nonconforming cannot be negative.
Quality Engineering (SYE3102) 53
[Link]@[Link]

A typical p-chart can be found in Figure 6.1.

0.55
UCL 0.51
Fraction nonconforming items

0.45

0.35
MEAN 0.32

0.25

0.15
LCL 0.12

0.05
0

10

15

20

Sample number

Figure 6.1. A typical p-chart.

Note that although we have assumed that all samples have the same size, it is not
necessary that this has to be the case. When samples are not of the same size, we can use
the actual sample size to determine the control limits valid for that sample size. When
sample sizes vary, we have control chart using variable sample size. The consequence is
that the control limits will not be straight line. A possibility to overcome this problem is to
use an average sample size. This is especially good when the sample sizes are almost the
same.

When we are interested in the number of nonconforming items in a sample of size n, we

can make use of np chart. The control limits for np-chart are

UCL  np  3 np (1  p )
CL  np

LCL  np  3 np (1  p )
Quality Engineering (SYE3102) 54
[Link]@[Link]

As for p-chart, we can use an estimate of p and we can also use different n when the
sample sizes are not the same for all samples.

10
UCL 9.5

8
# nonconforming items

4 MEAN 4.08

0
0

10

20

30

40

50
Sample number

Figure 6.2. A typical np-chart. Note that

the LCL<0, so it is not included.

6.2. The c-and u-chart

When dealing with the number of nonconformities in a sample, we can use the c-chart.
The number of nonconformities is different from the number of nonconforming items, The
former can be greater than the sample size while the latter cannot.

For c-chart, the traditional approach is using the Poisson model as it is well-known that
the number of nonconformities in a sample follows Poisson distribution with certain
parameter c. The mean and standard deviation are given as c and c, respectively. Hence
the control limits for c-chart are

UCL  c  3 c
CL  c
Quality Engineering (SYE3102) 55
[Link]@[Link]

LCL  c  3 c

A typical c-chart is displayed in Figure 6.3.

40

UCL 33.0
30
Number of nonconformities

20 MEAN 19.7

10

LCL 6.4

0
0

10

15

20

Sample number

Figure 6.3. A typical c-chart.

A related chart is the so-called u-chart when the quantity of interest is the number of
nonconformities per unit. The control limits for c-chart are

UCL  u  3 u / n
CL  u
LCL  u  3 u / n

The differences between the c-chart and u-chart are that u-chart can be used when sample
size varies and the average number of defects per unit in the sample is plotted. A typical u-
chart is shown in Figure 6.4.
Quality Engineering (SYE3102) 56
[Link]@[Link]

2.5
UCL 2.37

2
# nonconformities/sample

1.5

1 MEAN 1.016

0.5

0
0

10

20

30

40

50
Sample number

Figure 6.4. A typical u-chart. Note that in this

case LCL<0 and hence it is not included.

6.3. Potential problems when using standard formulas.

Firstly, the traditional np-chart and the control limits are based on normal approximation.
When p is small and the sample size is not large enough, the probability of false alarm is
very high. The table below is generated by

( False alarm)  1  P( LCL  Z  UCL)

Prob

assuming that Z follows binomial distribution with parameter n and p.

Exact false alarm probability for np-chart with

3-sigma limits when p is small.
p n=5 n=10 n=20 n=50 n=100
0.01 0.9520 0.9086 0.8189 0.6066 0.3695
0.02 0.9078 0.8179 0.6747 0.3674 0.1367
0.03 0.8672 0.7402 0.5465 0.2218 0.0508
0.04 0.8160 0.6710 0.4494 0.1335 0.0191
0.05 0.7749 0.5998 0.3611 0.0801 0.0074
Quality Engineering (SYE3102) 57
[Link]@[Link]

0.06 0.7359 0.5406 0.2957 0.0480 0.0047

0.07 0.6988 0.4876 0.2362 0.0287 0.0023
0.08 0.6636 0.4402 0.1925 0.0172 0.0026
0.09 0.6304 0.3904 0.1529 0.0103 0.0015
0.1 0.5910 0.3503 0.1240 0.0084 0.0023

Secondly, the probability of failure in detecting a change of process characteristic will be

high. Because of the actual distribution of p is far from normal, the control limits based on
normal approximation can no longer describe the real state of the process; neither can it
detect process changes accurately. A related problem is that, because of the problem of
high false alarm probability, true signal may also be overlooked.

Thirdly, in a near-ZD environment, UCL will usually be less than zero, making it
meaningless and when counting the number of nonconforming items, process
improvement cannot be detected unless some run rules are applied. This will generally
make the control chart more cumbersome to use for practicing engineers and even if it can
done by computer, the run rule will make the interpretation of the chart difficult to
explain.

Last, but not least, a more serious problem is that when the fraction nonconforming is very
small, even the UCL can be smaller than one, unless the sample size is very large. In this
case, every sample containing only one nonconforming item will cause a false alarm,
which is not useful to the chart user at all and makes the control chart itself useless.

The same problems are also associated with c-chart. If c is small, c≤0.9, say, the upper
control limit can be less than one. Since one is the minimum number of nonconformity for
an alarm, it is usually taken as the upper control limit. Note that since the number of
nonconformities in a sample, C, is a Poisson distributed random variable, P(C≥1) can be
calculated as follows

P (C  1)  1  P(C  0)  1  e  c

where c is the expected value of C, usually called the average number of nonconformities.
Quality Engineering (SYE3102) 58
[Link]@[Link]

For really small values of c, say c<0.01, the approach will not be sensitive because of the
low probability of the occurrence of nonconformities. Also, even though it is thought
conservative when we round the upper control limit up to one when the calculated value
of UCLc is lower, it can give a much higher probability of false alarm than desired for a
wide range of c. In the table below, we tabulate, for some values of c, the probability that
one or more nonconformities will be detected.

Probability of false alarm when the average number

of nonconformities is lower than 0.92, for which the 3-sigma
UCLc has to be set to one.
____________________________________
c Prob(C≥1), in %
____________________________________

.01 1.00
.02 1.98
.03 2.96
.04 3.92
.05 4.88
.06 5.82
.07 6.76
.08 7.69
.09 8.61
___________________________________

To highlight with a concrete case, we use the example in Xie and Goh (1993). In Messina
(1987), p.144, a control chart is given for average process proportion nonconforming
p=0.0234. The conventional 3-sigma UCL and LCL are calculated as 0.0426 and 0.0043,
respectively. If the LCL is used, then for a sample size of n=563 used in the original
example, the process is out of control when two or less nonconforming items are found.

In that case, we can calculate the probability of the number of nonconforming items, Z, as
P ( Z  0)  (1  p) n  0.000002

P ( Z  1)  np (1  p) n 1  0.000020
n(n  1) 2
P ( Z  2)  p (1  p) n  2  0.00015
2
Thus
Quality Engineering (SYE3102) 59
[Link]@[Link]

P ( Z  2)  0.000002  0.000020  0.00015  0.00017.

This is far from the assumed level 0.00135 for a process control chart.

In this case, the lower limit should be set at least 3. The probability is then 0.00106 which
is fairly close to the desirable level. This means that when there are only 3 nonconforming
items in a sample (of size 563), the probability that the process is better than before is 1-
0.001=99.9%.

One possible way to solve the problem is to use the probability limits instead of the
traditional 3-sigma limits. It is an easy task to calculate the exact probability limits when
the underlying distribution for process characteristics is known. In the case of the charts
we have described, it is well-known that the number of conforming items in the sample, or
the number of nonconformities in the sample follow binomial and Poisson distribution,
respectively.

Given an allowed level of false alarm probability , the steps for setting up a control chart
based on probability limits can be summarised as follows.

(1) Obtain a series of process characteristics though observation or calculation

(2) Select a suitable model and estimate the parameters

(3) Calculate the mean and use it as the centre line (CL)

(4) Calculate the upper control limit (UCL) and the lower control limit (LCL) as the
upper 1-% and lower % percentile, respectively

(5) Plot the process characteristics on the chart and connect the consecutive points

(6) If there are points that fall out of the limits, check the reason

(7) Continue the plotting whenever a new measurement is obtained