5

Section

WARWICK MANUFACTURING GROUP

Quality Management

Introduction to
Statistical
Process Control
 QUALITY MANAGEMENT

Introduction to Statistical Process

Control

© Warwick Manufacturing Group

School of Engineering
University of Warwick, Coventry, CV4 7AL, UK
Phone +44 (0)24 7652 4240 • Fax +44 (0)24 7652 4307
e-mail paul.roberts@warwick.ac.uk
Table of Contents
Background 1

Principles and benefits of SPC 2

Variation and the normal distribution 3

Process capability 5

Process control 7

The fallacy of first article inspection 11

I N T R O D U C T I O N T O S T A T I S T I C A L P R O C E S S C O N T R O L
I N T R O D U C T I O N T O S T A T I S T I C A L P R O C E S S C O N T R O L

5
Section

Introduction to Statistical Process

Control

Background

T he traditional approach to quality that prevailed (and to some extent still

prevails) in the industrialised West for many years was for the Production
Department to make the product and the Inspection Department (Quality
Control) to inspect and check it against specification.

Recently, the emphasis for achieving quality has moved to preventative techniques.
This has come about as a result of the realisation that detection is wasteful,
inefficient and costly. The figure below shows the generalised process diagram for a
process operating on a detection basis. There are three major problems that can be
identified with such a system:

Rework

Good
Input Output (To customer)
Operation I

Scrap

Figure 1. A process using 100% inspection.

100% inspection is not 100% effective. No matter how good the inspector, some
good products will always be rejected or sent for rework due to fatigue, boredom
or a dozen other factors. More significantly, bad product will get shipped to
customers.

The system is costly in terms of manpower; enough inspectors must be employed

to ensure that inspection does not become a bottleneck in the production system.

1
©Warwick Manufacturing Group
Revision date: October 06
I N T R O D U C T I O N T O S T A T I S T I C A L P R O C E S S C O N T R O L

Responsibility for quality devolves from the person making the item to the
inspector of the item whilst the control of quality remains where it always will
remain, with the person in control of the production process. Thus, the only one
with the ability to affect the final quality of the finished product has no incentive to
pursue such improvements.

The logical way to overcome the problems associated with this type of system is to
apply preventative techniques at the operation stage to ensure that the product is
produced to the required quality. Such a system is shown in schematic form in the
figure below, the approach is based on Statistical Process Control (SPC) which is a
statistical method of data collection and analysis that works in such a way as to
monitor the operation and control it to it's maximum potential. This enables the
operation to be carried out in confidence that the final product will be good.

Output
Input (Good)
Operation SPC

Figure 2. The application of Statistical Process Control.

The origins of SPC date back to the inter-war period, and are based on the work of
Walter Shewhart who, in 1927 identified the use of control charts to detect process
variation. The man who is seen to have most influenced the development of SPC
as a technique, and popularised its use is W. Edwards Deming. Deming was a
disciple of Shewhart and was sent to Japan at the end of World War Two to help
redevelop Japanese industry. Amongst other philosophies he propounded the
principles and practices of SPC, the Japanese listened, took up his teachings with
enthusiasm and the rest is, as they say, history.

The recent growth in popularity of SPC as a manufacturing discipline is due partly

to the success of Japanese companies, but more directly to the initiatives of the
Ford Motor Company who have seized upon quality, and SPC in particular, as the
major competitive weapon for the present and foreseeable future.

Principles and benefits of SPC

T he core principle of SPC is the belief in the need to understand the variation
in a process and manage it on that basis. The long-term aim of SPC is to
minimise variation in processes so that customer requirements are more
closely met than before. There are three key elements in achieving this aim:

2
©Warwick Manufacturing Group
Revision date: October 06
I N T R O D U C T I O N T O S T A T I S T I C A L P R O C E S S C O N T R O L

ƒ Evaluating capability

ƒ Providing control systems

ƒ Providing guidance towards continuous improvement

Variation and the normal distribution

V ariation is part of our everyday lives. Both at work and in our private lives
we make allowances for its effects from the process of getting to work in
the morning to the output of a complex manufacturing system. However,
whilst a seat-of-the-pants approach to deciding how long we allow ourselves to get
to work may be perfectly adequate, a similarly haphazard approach to managing
processes at work is not desirable. We need to get a quantitative feel for the
variation in our processes. There are two basic elements to this variation: the
central tendency and the spread. We need to have a handle on both these since they
are vital to a successful process. It's no good being the right temperature on average
if, to achieve this, you've got one foot in the fire and one in the fridge! In addition
to these facts we need to understand the shape of the distribution. The two
processes below may have the same average and spread but they would elicit very
different decisions.

Figure 3. Different distributions-different decisions.

At this stage it is important to note the four potential causes of variation that can
affect a process:

Common (unassignable) cause variation. This is variation that is inherent in the

process and thus requires fundamental action to reduce it. In the process of getting
to work this will mean things like waiting time at fixed traffic lights; only
fundamental action on the process like changing route or removing the traffic
lights will remove the cause of the variation.

Special (assignable) cause variation. This is variation due to transient causes

outside the process and can usually be traced back to the cause. In the journey to
work example this would include road works, breakdowns etc. In most cases action
can be taken to achieve a reduction in the future effect of these ‘transient
problems’, eg better maintenance to avoid breakdowns, which does not
fundamentally change the process.

3
©Warwick Manufacturing Group
Revision date: October 06
I N T R O D U C T I O N T O S T A T I S T I C A L P R O C E S S C O N T R O L

Cyclical variation. As the name implies this refers to causes of variation that
follow a non-random repeating pattern. In the case of going to work this might
include summer to winter weather variations etc. This form of variation is
frequently difficult or expensive to control.

Tampering. This is a very serious cause and one that often goes unnoticed. The
principal cause of this type of variation is treating a common cause as a special
cause and reacting to it in the wrong way. The result is the introduction of greater
variation into the system. An example of this might be changing your route every
time you see a red traffic light ahead. Below we can see the effect of tampering in
an industrial situation where action is taken (resetting in this case) on arbitrary limits
which are inside the common cause variation pattern of the process and, as such
can be exceeded in normal operation without any special causes being present. This
is by no means the only form of tampering that may occur. For example, using
previous parts as a master causes problems too. What we must assess is whether we
are managing the process based on its natural variation and only treating special
causes as one offs to be eradicated.

Adjustment
50 Distribution Arbitrary Natural
Average limits process
spread

Process reset based on arbitrary action limit (tolerance based).

Arbitrary
limits
Natural
35 Distribution process
Average spread

The resulting mismatch of requirements will lead to more and opposing

adjustments, introducing more and more variation into an originally
stable situation.

Figure 4. The effect of tampering.

In most industrial processes the distribution can be said to approximate to the

normal distribution (the symmetrical, bell-shaped curve on the reader's left above).
This largely obviates the need for worrying about the shape of the distribution and
so emphasis shifts to the distribution central tendency (usually represented by the
average) and the spread (dealt with in terms of the standard deviation). The figure
below is a representation of the normal distribution indicating the relationship
between the standard deviation and the proportion of the population contained
within limits set at multiples of the standard deviation from the central point
(average) of the distribution.

4
©Warwick Manufacturing Group
Revision date: October 06
I N T R O D U C T I O N T O S T A T I S T I C A L P R O C E S S C O N T R O L

99.73%

95.44%

68.26%

Average
-3sd -2sd -1sd + 1sd + 2sd + 3sd
Figure 5. The Normal distribution.

Since the normal distribution never reaches the x-axis, there is no point at which
100% of the distribution will be contained within any two boundaries. It can be
seen, however, that when these limits are set at +/- 3sd either side of the centre
99.73% of the distribution are bounded. This means that only 1 in 1000 will fall
outside teach boundary and is taken, in industrial terms, to be a good
representation of the total spread. For more accuracy we can go further (+/- 4sd
gives approximately 3 in 100,000 outside each limit) but within the constraints of
the assumption of normality the value of this exercise is limited.

The precise shape of the normal distribution will vary from process to process but,
provided we have no good reason to believe that significant non-normality exists,
we can characterise a process by use of the mean and standard deviation of that
process. These factors then provide us with the basis for capability assessment and
control of the process.

Process capability

C apability evaluation is the method by which we determine whether a

process is up to the job of meeting the specifications set for it. It is
important, before attempting to establish the capability of a process to
ensure that the process is stable. The concept of stability is effectively the same as
the concept of a process being "in control". The key issue is that if a process is not
stable the capability will be constantly changing due to the transient effects of
special causes.

5
©Warwick Manufacturing Group
Revision date: October 06
I N T R O D U C T I O N T O S T A T I S T I C A L P R O C E S S C O N T R O L

The importance of understanding process capability cannot be overstated. If we are

to attempt at any level to design for manufacture we need to understand not only
the requirements on the process (effectively our specifications) but also what the
process is able to achieve (capability). Without both sides of the equation we are
not able to make sensible decisions about how to manage our processes at an
appropriate stage in the product lifecycle and we doom ourselves to fixing and fire
fighting when we actually try to make the product.

Process A Process B

Process C Process D

Figure 6. The output of four processes.

Consider the four processes shown above with the specification limits. Clearly,
process A is producing many components both above and below tolerance;
process D is offset and is, as a result, producing components below the bottom
tolerance limit; process C is producing almost all components within tolerance and
process B is operating well within the tolerance limits.

Given the information provided in the above diagram we can act upon the process
(resetting process D, for example, or attempting to reduce the spread of process
A), without such information we would be making such decisions in the dark.

6
©Warwick Manufacturing Group
Revision date: October 06
I N T R O D U C T I O N T O S T A T I S T I C A L P R O C E S S C O N T R O L

Process control

Figure 7. Means chart for a bottle filling process.

T he figure above shows the output of a process that is used to fill miniature
spirit bottles with a nominal 50 cl of spirit.. It can be seen that the output of
this process fluctuates about the 50 cl nominal, but since no sample
breaches the Upper or Lower Control Limits, nor is there an identifiable trend, the
process is said to be ‘in statistical control’. Note that in order to produce the above
plot, no knowledge of the specification is necessary. The distance between the
upper and lower control limits represents the spread of the process due to
common cause variation, and no inference can be made on the suitability of this
process to meet the needs of the situation. Thus, although the process appears to
be in control, it may be incapable of meeting specification. Conversely, a process
that is inherently capable of meeting specification may be running out of control.
The important point being made here is that capability and control are not linked,
thus knowledge of one tells us nothing of the other.

Statistical Process Control is all about predictability. If we have a process the spread
of the key product characteristics from which we know approximates to a normal
distribution, provided that no special or assignable causes of variation are
present, we can have a good degree of confidence that the output will fall between
two limits given by +/- 3 standard deviations from the process average. The very
nature of statistics means that we cannot predict any individual outcome but the
comfort afforded by knowing the upper and lower bounds of production is far
preferable to the lack of predictability currently seen in many factories.

Predictability of outcome is only possible if no special causes of variation are

present. When special causes are operating the assumption of a normally
distributed population is no longer valid and confidence in the bounds of the
output is lost. The purpose of running SPC charts is to predict when an assignable
cause is likely to be operating to allow appropriate action to be taken on the
process. In effect, what we are attempting to do with these charts is simply to
distinguish between the effects of unassignable causes (ie when the process is
merely exhibiting the variation expected under normal circumstances) and the
effects of assignable causes (extraneous variation which takes the process outside

7
©Warwick Manufacturing Group
Revision date: October 06
I N T R O D U C T I O N T O S T A T I S T I C A L P R O C E S S C O N T R O L

normal expectations). When put in these terms it can be seen that there is no direct
link between process capability and process control. However, it is obviously less
useful to run control charts on processes which are not capable since they will still
produce out of tolerance items and thus require 100% inspection in an attempt to
prevent these reaching the customer.
Chart Lower Chart Upper
Control Limit Control Limit

Population
Distribution

Sample
Average
Distribution

Common
centre-line

Figure 8. Relationship between population and sample distributions.

We noted earlier that the preferred form for control of processes to take is
predictive rather than reactive. To achieve this we use a system of ongoing
sampling rather than 100% inspection. Control limits (i.e. the points at which we
do something) are based upon the distribution of the sample averages, which is
taken to be representative of the population average. This assumption is valid
because of the Central Limit Theorem, which states that the distribution of
sample averages will tend to be a normal distribution (the tendency is more
pronounced for sample sizes of 4 and above) even when the parent distribution is
significantly non-normal. Where the parent distribution is normal and the sample
size is n, the two distributions will be linked by the formula shown below.

σ
s=
n

where, σ = population standard deviation, s = sample average standard deviation

The control limits applied to the sample average distribution are at +/- 3s. This
means that only one in one thousand sample average readings would fall outside
these limits if no special causes were present. This gives a low probability of taking
action when an assignable cause is not present. There is, clearly, a complementary
risk of failing to take action when a special cause is present. This latter risk could be
reduced by using a limit of +/- 2s. There would, however, be an increased risk of
tampering. Due to the disastrous effects of tampering (as noted earlier) it has been

8
©Warwick Manufacturing Group
Revision date: October 06
I N T R O D U C T I O N T O S T A T I S T I C A L P R O C E S S C O N T R O L

found in practice that the best balance of the two risks (known as alpha and beta
risks respectively) is given by setting control limits at +/- 3s.

There are several basic types of control charts that can be run for variable control.
These are all based upon the principle that two elements of the distribution need to
be controlled for satisfactory process performance; these are the sample average
and the spread of the sample. The former can be charted as sample average or
sample median and the latter as sample range or standard deviation. Any
combination of these two sets of charts may be acceptable. The most common
chart is a combination of average and range, an example of this type of control
chart is shown overleaf. Average is preferred to median because it is a more
accurate measure although median may be used if the workforce are unhappy with
calculating averages or if the environment is particularly dirty or harsh preventing
the use of calculating aids. Standard deviation is more accurate than range but not
greatly at the normal sample sizes used and is much more complex to calculate with
a far greater possibility of error. The view is generally taken that the range is an
adequate, unbiased estimator of spread.

Several issues arise when using variable process control charts (also known as
Shewhart charts). The decisions that are made are principally the accommodation
of statistical niceties and practical considerations. For this reason it is of vital
importance that SPC is not seen as an alternative to engineering or management
judgement but as an aid to it when considering the following aspects:

ƒ Where do we use variable control?

In short, on any process where the output can be measured. However,

it makes sense to concentrate on areas of most immediate benefit
taking into account things like customer complaints, build problems,
high scrap/rework, high quality costs etc.

ƒ What characteristics do we control?

The characteristics to control using SPC are the same ones to which
priority is given for any form of control. Those that are important to
the customer and those with which we are presently experiencing
difficulties.

ƒ What subgroup size is best employed?

A compromise between time/cost to measure and level of control is

the key element of this decision. The best compromise statistically
speaking is 5. It is possible to go down to a subgroup size of 3 if
necessary for pragmatic reasons and even lower can be used with
caution but the smaller the subgroup size the less confidence we can
have in the control applied.

9
©Warwick Manufacturing Group
Revision date: October 06
I N T R O D U C T I O N T O S T A T I S T I C A L P R O C E S S C O N T R O L

ƒ How frequently should samples be taken?

This is very process dependent. You should take into account the rate
of change of the process (is it stable like press tools or fast-changing
like some machining processes). The faster the process changes the
more frequent should be the sampling, this must be balanced against
the additional effort required to take samples. Another factor is the
number/value of items produced between samples as this is the
quantity at risk (and which needs to be inspected if an 'out of control'
signal is given). A common compromise is one sample per hour.
Always err on the side of too frequent sampling in the early stages and
relax this as control is demonstrated by long periods of stability. An
important point to note here is that in SPC we deal with random
rational subgroups this means that subgroups must be randomly
selected from the population and the samples forming the subgroups
must be consecutively produced. If our sampling pattern is too regular
we run the risk of adversely affecting the randomness of our samples
(by aligning with an unknown cyclical factor such as tea breaks etc.)

ƒ When should control limits be calculated?

Control limits are calculated using subgroup data and it is conventional

to wait until 20 subgroups have been generated before performing the
calculation. This can be done as soon as only 10 subgroups into the
chart but the limits are somewhat questionable and should, in any case,
be recalculated once 20 subgroups have been produced. The formulae
below are those used in calculating control limits for average and range
charts:

Control limits should also be recalculated when a chart is completed to

take account of any changes over time as well as when a significant
change to the process has taken place.

ƒ When should action be taken?

There are several indicators of an 'out of control' process (i.e. when

there are assignable causes present). These include: a range point
outside the control limit, an average point outside one of the control
limits, a run of 7 points going up or down (indicates a trend which may
be rapid tool wear), a run of 7 points above or below the mean
(indicates a shift in mean), a run of 8 points in the middle third of the
chart (apparently an improvement in the process spread). Examples of
these are shown overleaf. In addition any clearly non-random pattern
should be acted upon. Other action points can be defined of a
more complex nature but these may confuse the operator and result in
worse rather than better control.

10
©Warwick Manufacturing Group
Revision date: October 06
I N T R O D U C T I O N T O S T A T I S T I C A L P R O C E S S C O N T R O L

ƒ Should signals always be acted upon?

Yes. The chart is the process talking to you, if it asks for action it is
because it is necessary. It is not acceptable to use the excuse that 1 time
in 1000 the signal may be false. Once a signal has occurred you are no
longer sure of your process and must act accordingly. Ideally this
means stopping the process but, if this is not possible all items
produced should be inspected until the problem is solved. In the case
of a positive signal (e.g. the middle third rule) the process should not
be stopped but the control limits should be recalculated as soon as
possible to reflect the change in the process and investigation
undertaken to understand the reason for the improvement so that the
gain can be held.

The fallacy of first article inspection

Setting adjustment

USL

X
X X

Nominal x

LSL

Figure 9. The fallacy of first article inspection.

T he diagram above illustrates a common fallacy that if the first product from
a production process is measured and found to be within specification, all
that follow will also lie within specification. Any organisation that conducts
this practice does not understand the nature of variation inherent in every process.
From above, working from left to right, the first x lies above nominal. If the true
(but unknown) distribution from which the sample comes is the lightly shaded one,
an adjustment of setting towards nominal will result in succeeding product to come
from the darker shaded distribution, clearly a poor situation. In the second
example, if the adjustment is made, the improvement will not be sufficient to
prevent defective product. Finally, the third example shows that even if nominal is
recorded, if the process is incapable a percentage of defective product is inevitable.

11
©Warwick Manufacturing Group
Revision date: October 06
I N T R O D U C T I O N T O S T A T I S T I C A L P R O C E S S C O N T R O L

It is hoped that these introductory notes on Statistical Process Control give the
reader an insight into an extremely powerful yet inherently simple technique. It is
important to note that for the application of the technique to be successful,
understanding of variation and its effects on processes is essential, as is
management commitment to support those controlling and improving their
processes.

µµµ

12
©Warwick Manufacturing Group
Revision date: October 06