13

Statistical Process Control a n d

Quality Control

Overview sional measurements of the precision needed

for SPC are generally done offline. Obtaining
Statistics is a branch of mathematics deal- the final dimensional stability needed to mea-
ing with the collection, analysis, interpreta- sure a part may take time. For example, amor-
tion, and presentation of masses of numerical phous injection molded plastic parts usually
data. The word statistics has two generally ac- require at least a half hour to stabilize.
cepted meanings: (1)a collection of quantita- The SPC system starts with the premise
tive analysis data (data collection) pertaining that the specifications for a product can be
to any subject or group, especially when the defined in terms of the product’s (customer’s)
data are systematicallygathered and collated, requirements or that a product is or has been
and (2) the science that deals with the collec- produced that will satisfy those needs. Gen-
tion, tabulation, analysis, interpretation, and erally a computer communicates with a series
presentation of quantitative data. of process sensors and/or controllers that op-
Statistical process control (SPC) is an im- erate in individual data loops (Chap. 7). The
portant real-time online method by which computer sends set points (built on which
a production process can be monitored and performance characteristics of the product
control plans can be initiated to keep quality are desired) to the process controller, which
standards within acceptable limits. Statistical constantly feeds back to the computer to sig-
quality control (SQC) provides offline anal- nal whether or not the set of points are in fact
ysis of the big picture, such as what was the maintained. The systems are programmed
impact of previous improvements. It is impor- to act when key variables affecting product
tant to understand how SPC operates. quality deviate beyond set limits (1, 13, 559,
There are two possible approaches for real- 596).
time SPC. The first, done online, involves the
rapid dimensional measurement of a specific
product characteristic or a nondimensional Combining Online SPC and OfJline SQC
bulk parameter such as weight which is often
the more practical method. In the second ap- Online SPC software excels at monitoring
proach, in contrast to weight, other dimen- production processes in real time to give you

1127
1128 13 Statistical Process Control and Quality Control

a close-up view of what is currently occurring. Prior to the widespread implementation

This important capability uses parameters of supervisory control and data acquisi-
that have a direct, understandable effect on tion (SCADA) and human-machine inter-
the process. However, SPC does not provide face (HMI) systems, most SPC and SQC was
any information about the overall operations. performed by quality-control departments as
It cannot detect differences over time, look an offline process. Data were collected from
at a complete process, or compare multiple test stations, laboratories, etc. and statistical
production lines. SPC’s classical use involves analysis was performed later. SCADA-HMI
keeping plant-floor operators from overad- systems, however, have made it feasible to
justing their machines. It also helps determine provide plant-floor SPC charts using data col-
when a significant shift in the process happens lected in real time directly from the process.
and whether it requires corrective action. Fabricators who want to standardize SPC and
Offline SQC is essential to detecting differ- SQC to increase their usefulness should (1)
ences over time such as shift changes, day-to- provide the plant floor with SPC charts and
week problems, differences between suppli- (2) make data collected by SCADA systems
ers of materials, differences between IMMs, available for offline analysis. A number of
and so on. SQC more easily permits combin- SPC and SQC software programs are avail-
ing plant-floor data with results from test sta- able to support these efforts. One should rec-
tions and the laboratory. Once the process is ognize that the bulk of SPC’s value is derived
in control, offline SQC provides the means from process improvements developed from
to make long-term process improvements. It offline SQC analysis (141).
can uncover relationships, monitor the results
of process changes, and provide various other
decision-support functions such as whether Statistical Material Selections: Reliabilities
the process can continually deliver products
within specifications. Virtually all classical design equations as-
To achieve SQC’sfullest potential, the pro- sume single-valued,real numbers. Such num-
cessor should analyze data the same way as bers can be multiplied, divided, or otherwise
those in the QC department. The biggest subjected to real-number operations to yield
challenge with offline software is data avail- a single-valued, real number solution. How-
ability. Because these systems collect vast ever, statistical materials selection, because
amounts of data, they provide challenges it deals with the statistical nature of property
and opportunities for improving the process values, relies on the algebra of random vari-
(Chap. 12). ables. Property values described by random
variables will have a mean value, represent-
ing the most typical value, and a standard de-
Improve Quality and Increase Profits viation, which represents the distribution of
values around the mean value.
To achieve better yields, higher quality, The mean values and standard deviations
and increased profits, fabricators should con- of particular property measurements must
sider the SPC and SQC techniques as stan- be treated according to a special set of laws
dard tools for understanding, validating, and for the algebra of random variables. Exten-
improving processes in all areas of manu- sive information can be found in any statis-
facture, including product distribution, trans- tics text. The algebra of random variables
portation, and accounting. Using online shares many elements of structure in com-
software,SPC provides the close-up view; us- mon with the algebra of real numbers, such
ing offline software, SQC detects differences as the associative and cumulative laws, and
over time. These two complementary tech- the uniqueness of sum and product. Distribu-
niques provide two different essential func- tive laws for addition and multiplication also
tions. hold.
 13 Statistical Process Control and Quality Control 1129

Statistical Material Selections: Uncertainties and the probability of a six is 1/6. Another
That A r e Nonstatistical example of probability is illustrated from a
deck of 52 cards. The probability of a spade
Limitations in processing plastics do ex- is 13/52, since there are 13 spades in a deck.
ist. Thus, some engineering random variables
carry with them a degree of uncertainty that
may be nonstatistical; that is, they cannot be Statistics and Commitments
described in terms of mean values and stan-
dard deviations. Unpredictable examples ex- Today “made in Japan” means something
ist. They include (1)material properties, such very different from what it meant to U.S. con-
as strength, may be influenced by time, cor- sumers before 1950. Once synonymous with
rosion, and fluctuating thermal environments inferior quality,Japanese precision and work-
that are not factored into the analysis; (2) fre- manship made a complete turnabout; today
quently a stress analysismay require simplify- Japan’s top quality products have changed
ing assumptions so that, as a result, uncertain- the world markets. How did Japan make such
ties are introduced of unknown magnitudes; a great stride in this area? The answer re-
and (3) uncertainties may arise from process- sides in both their people’s devotion and in
ing operations assumed to be constant, such their implementation of superior quality con-
as melt flow. trol methods. Many of these methods were
The statistical approach compels the ex- taught to top management and engineers by
perimenter to specify as accurately and com- W. Edwards Deming, who has accurately
pletely as possible those factors that influence been referred to as the American who re-
the properties under examination. Equally made “made in Japan.”
important, the technique requires that those Well before 1940, Deming had established
factors that cannot be specified accurately are a reputation for himself in the United States
recognized and considered in assessing prop- as a statistician. Via General D. MacArthur
erty values. in 1946, following professional duties in
India, he assisted Japanese statisticians in
their reconstruction by applying his know-
Statistical Probabilities and Quality Control ledge, which they put to work in their man-
ufacturing plants. He predicted that Japan
The term probability has a number of would invade worldwide markets with qual-
synonyms: likelihood, chance, tendency, and ity products within five years; they made it in
trend. To the layman probability is a well- four. Because of his work on improvement of
known term, which refers to the chance that quality, the Union of Japanese Science and
something will happen. It is possible to de- Engineering (JUSE) instituted the annual
fine probability with extreme mathematical Deming Prize. Later his work was accepted
precision via statistics. It can be defined from and used in the United States.
a practical viewpoint as it applies to quality
control as the likelihood of a molded prod-
uct being successful (or having a degree of Statistics and Injection Molding
success) based on the different well-defined
variables that exist in materials and during Statistical process control (SPC) is a statis-
processing. tical method of process monitoring to meet
If a coin is tossed, the probability of a head quality assurance during injection molding.
is 1/2 and the probability of a tail is 1/2. The Basically, it is concerned with information
dice used in games of chance are cubes with about the stability and reproducibility of a
six sides and spots on each side from one to process and concentrates on a specific se-
six. When a die is tossed on the table, the quence of key tests on the characteristics of
probability of a one is 1/6,that for a two is 1/6, the parts being manufactured. Such results,
1130 13 Statistical Process Control and Quality Control

presented graphically as a quality-control able control limits are exhibited. A process

chart, provide information on the progress is statistically capable when parts can be con-
of the process and the need for intervention sistently produced within specification. Both
(16, 30, 65, 188, 210, 211, 437, 493, 558, 559, conditions should be present for optimum
596,611). process operations. SPC analysis can prove
The proper use of control charts with plot- the balance and offer means for problem solv-
ted data and graphs will help continuously ing if special cause variations take a process
boost quality by tightening control limits. out of control or capability.
This chapter will review the powerful tools The first thing you want to learn is what
of controls to help you consistently and ac- the natural variation of the process is so that
curately fine-tune your processing plant. The you know when and when not to intervene.
terms used for these controls can be sum- If fabrication processes were more standard-
marized as statistical process/quality control ized in general, considerable variation would
(SP/QC), or they can be subdivided into sta- be eliminated before any measurements were
tistical process control (SPC), statistical good even taken. Thus, SPC should obviously be
manufacturing practices (SGMP), continu- applied when it is cost effective. Ideally, you
ous process control (CPC), and others. These have SPC in every phase of fabrication, but
different abbreviations will be used, with the you must consider the cost of prevention ver-
main emphasis on SPC. sus that of mistakes, especially for inexpen-
The goal of SPC is to decrease rework and sive molded parts.
increase the first-time yield of higher-quality For a part to be produced within an op-
cost-effective parts that meet specifications timum process window by means of injec-
and just-in-time (JIT) delivery. More specifi- tion molding, the people involved in the pro-
cations are requiring SPC, with the method- cessing of an order must act knowledgeably
ology proving its effectiveness even in short- and responsibly in all phases. As shown in
run production. SPC provides a feedback Fig. 13-1, quality assurance in an injection
loop for the manufacturing process. Typical molding plant comes into effect only late in
process parameters monitored include times the course of order processing. This means
(cycle, plastication, injection, mold open, and that the efforts taken to avoid defects at that
cure), temperatures (mold, nozzle, melt, bar- time are relatively great, whereas those nec-
rel zones, and dryer), pressures (first stage, essary at the beginning of order processing
second stage, and back injection pressures are relatively slight.
and clamp), ram position parameters (shot If, for instance, the incorrect shrinkage was
length, shot cushion, injection velocity, and used when designing the mold, excessively
plastication velocity), screw speed, and hop- high injection and holding pressure must now
per relative humidity. be used to produce dimensionally accurate
Specifications may state a process con- parts. Thus, the production department is ex-
dition, expected baseline performance, and pected to solve problems that do not fall
allowable range of variation, with a nomi- within its area of responsibility. This situa-
nal value in the midpoint of that range. A tion can result in different problems occur-
fabricator then applies past experience with ring. Although the dimensions of the molded
equipment capabilities. There is a degree of part may be within tolerance, the mechan-
negotiability between the specification and ical properties may be reduced because of
fabricator’s capability. It has been found that the higher holding pressure required. Also,
an acceptable range of natural variation can minor fluctuations of the processing param-
sometimes offer a higher standard than re- eters already result in rejects. Overall, more
quired by specifications, resulting in better defects occur during production and the pro-
process control and safer products. duction process is unreliable.
A process can be identified as being in sta- A process out of control does not neces-
tistical control when only random variations sarily mean that you are making bad parts,
(or common causes) falling within accept- but SPC analysis can indicate a variation that
 13 Statistical Process Control and Quality Control 1131

production scheduling

mold rework

materials procurenknt

mold design. mold construction

clarification of order with the customer

little great
to avoid defects __c

Fig. 13-1 Triangle (left) shows the different departments involved in quality assurance as they relate
to the final product curve (right) in order to avoid defects.

needs to be limited. This allows you to build orientation, etc.; and molded part proper-
quality, rather than inspect quality into the ties include dimensions, weight, strength, and
part. The application of SPC to short-run pro- impact behavior. In the injection molding
duction can cause problems when there is process, a large number of factors influence
not enough data to calculate control limits the quality of the part. A distinction is made
in a timely manner, or data exist but on many between parameters directly adjustable on
small lots of different parts run on the same the machine and those that are adjusted on
equipment. The competitive power of SPC the basis of machine parameters, disturbance
can be realized even by those who produce factors, and material characteristics, as well
complicated parts in low volume. as their interrelations.
In conventional SPC, parts are sam- The physical description of these relation-
pled during the production process. Rele- ships is highly complex. Today we have rel-
vant characteristics (usually dimensions) are atively precise ideas as to the qualitative re-
checked and the mean, as well as scatter, lationships. For practical purposes, however,
monitored with the aid of control charts. Sta- it is necessary to know the quantitative rela-
tistical criteria apply when assessing the con- tions. The quality of the molded part must be
trol charts. If they are not fulfilled, it is nec- capable of being described from the material
essary to act in order to regulate the process. properties and process parameters. This ulti-
How such an intervention is carried out is left mately necessitates, among other things, the
to the experience of the operators. Process calculation of the crystallization and shrink-
control by action limits within the tolerance age processes. No model of this type is known
range is intended to largely avoid the produc- to be perfect.
tion of parts outside the tolerance limits.
Physical process models describe the rela-
tions among process parameters, material Computers and Statistics
structure, and molded part properties. Pro-
cess parameters include injection speed, Computers make statistics a more flexi-
mold temperature, screw speed, and hold- ble tool and help prevent the “cookbook”
ing pressure; material structure includes in- approach (the blind application of the same
ternal stresses, degree of crystallization, filler standard techniques no matter what problem
1132 13 Statistical Process Control and Quality Control

Fig. 13-2 IBM-compatibleoperator interface used with Husky's injection molding machines.

exists). As shown in Fig. 13-2, an IBM- ers, is no dreary science of number-crunching

compatible operator interface is standard on drudgery.
Husky Injection Molding Systems' line of in- Statistical methods should be applied to
jection molding machines, which range from decision making at all stages of production,
225 to 3,650 tons (250 to 4,000 U.S. tons). from incoming materials to outgoing prod-
The operator interface consists of an industri- ucts. For example, statistics can help with
ally hardened computer and high-resolution forecasting, a problem managers face ev-
color monitor. The screens are logically or- ery day: Should raw materials be reordered;
ganized to allow access at any level with no should marketing and advertising techniques
more than two keystrokes. Data entry is sim- be changed? The data used to make these
plified by the use of four-direction cursor decisions represent random variation-white
movements. The PLC memory is backed up noise-as well as real changes, such as drops
on the hard disk, in the event of a power in sales or increases in production.
failure. Quality control is an area where manage-
A statistical perspective can lead to a ment strategy can be applied easily. In the
simple route to substantially increasing past, quality control simply meant throwing
productivity, quality, and profit. Statistics out bad products, and management regarded
is concerned with the design of efficient it as a tradeoff with productivity. That meant
experiments and transformation of data into quality control was being exercised too late.
information-in other words, with asking Quality control should mean learning about
good questions and receiving good answers. the variability of all aspects of production,
For most people, the word "statistics" including maintenance, purchasing, market-
conjures up endless tables of uninteresting ing, and design. Traditionally, quality control
numbers. But modern statistics has very has been the exclusive concern of engineers.
practical applications and, thanks to comput- But it should be the concern of all employees,
 13 Statistical Process Control and Quality Control 1133

and quality-control data should be displayed

prominently for workers, engineers, and
managers to examine and discuss.
Statistics is also concerned with designing
experiments. Poorly designed experiments
give no useful information no matter how so-
phisticated the statistical techniques used to
analyze their results. Most companies need to
run experiments to develop new processes,
but experimentation is expensive. Factorial
experimental design is the way to get the most
information for the least expenditure.
Many experimenters still believe that one
variable must be examined at a time. That
Fig. 13-3 z3 factorial experiment can produce an
variable is varied, while other conditions are impressive quantity of data. Here temperature,
held constant. Besides requiring an enor- concentration, and catalyst are examined for their
mous number of runs, this method of experi- effects on production.
mentation does not reflect nature.
If you vary one factor at a time, you as- creases in concentration, change in temper-
sume that nature behaves as if variables op- ature, and change in catalyst-can be found,
erate independently. They usually do not. with each main effect being discovered by
Raising the temperature may have one re- comparing the means of two sets of four trials.
sult at low mixing speeds and an opposite Production with catalyst A, for example, is
effect at high mixing speeds. Interactions of examined at high temperature, high concen-
several variables as well as the effects of tration (180°C, 40%); at high temperature,
changing a single variable can be examined low concentration (180"C, 20%); at low tem-
in factorial experiments. A simple example is perature, high concentration (160"C, 40%);
a 23 factorial experiment: three variables- and at low temperature, low concentration
temperature, concentration, and catalyst- (160"C, 20%). The mean of these production
are examined at two levels (+, -). All com- levels (the numbers circled in the front cor-
binations can be examined in only eight runs, ners of the cube) is compared with the mean
as shown in Table 13-1. of production levels when catalyst B is used
Production results for each set of condi- (the numbers circled in the back corners of
tions can be plotted as corners of a cube, as the cube) under the same conditions of tem-
shown in Fig. 13-3.Factorial experiments pro- perature and concentration.
duce an impressive quantity of information. Interactions of variables can also be de-
In the figure, three main effects-from in- tected. An increase in temperature affects
production differently when catalyst B re-
Table 13-1 Statistical analysis: interactions of places catalyst A. A third-order interaction-
several variables change in the two-factor interaction when the
third factor is varied-can also be obtained.
Run T C K T C K
It is suggested that sometimes it is better to
1 _ _ - 1 0 0 0 study four or five variables. Four variables re-
2 + - - t 1 O O quire 24 or 16 runs; five require Z5 or 32.
3 - + - c 0 1 0 Frequently in practice, and especially in
4 + + - t c 1 1 O early stages of process development, more
5 - - + k O O 1 than five factors must be examined. But a
6 + - + t k l O l full factorial experimental design with 2" runs
7 - + + ck 0 1 1 is usually unnecessary. Fractional factorial
8 + + + t c k 1 1 1
designs, in which only a carefully selected
1134 13 Statistical Process Control and Quality Control

charts plot measurements taken on a con-

=-
tinuous scale and include the X-bar, R (av-
I
FLOWCHARTS RUN CHARTS erage, range); X-bar, S (standard deviation);
and X-bar, MR (individual moving range).
Attribute control charts plot qualitative data
- - - -CHARTS
CONTROL --- (padfail, good/bad, go/no go).
CHECK SHEETS Cause-and-effect (or fishbone) diagrams
examine a process problem based on worker,
material, and machine experiences to iden-
CAUSE AND EFFECT tify possible causes. Scatter diagrams show
PARETO CHARTS DIAGRAMS
the relationship between two different mea-
surements and can verify the true cause of a
problem. True cause will show a strong rela-
HISTOGRAMS SCATTER DIAGRAMS tionship with process deviation.
Fig. 13-4 Statistical tools.

portion of the possible combinations of ex- Online Monitoring of Process Variables

perimental conditions is run, can still provide
an enormous amount of information. SPC began in the late 1920s with the work
Fractional factorial designs are especially of W. A. Shewhart and the subsequent use of
useful for finding the main factors that will statistical control charts in industrial appli-
affect production. They miss interactions of cations in the 1930s. Later work by Dodge
several variables, but these interactions are and Romig, Deming, Juran, Feigenbaum,
usually negligible. Ishikawa, Crosby, and many others advanced
the science of SPC to a very high plane.
There are three phases in the evolution of
Statistical Tools most quality-control systems:

Several statistical tools can be used to mon- Defect detection. An army of inspectors
itor process parameters and solve the varia- tries to identify defects.
tion problems. These are summarized graph- Defect prevention. The process is moni-
ically in Fig. 13-4. Flowcharts show the order tored, and statistical methods are used to
of steps in a process. Check sheets record in- control process variation, enabling adjust-
formation about a problem and can be used ments to the process to be made before de-
to construct Pareto charts, histograms, and fects are produced.
other charts. Pareto charts set the priorities Total quality control. It is finally recognized
by graphing problems in order of worst to that quality must extend throughout all
best. Histograms are bar charts showing the functions and it is management’s respon-
variation among measurements, including sibility to integrate and lead the various
centering, range, and frequency distribution. functions toward the goals of commitment
Problems are charted in a “should/actual” to quality and customer-first orientation.
comparison. It is well known that there are two major
Run charts plot a process measurement problems in using the defect-detection ap-
in sequence and can identify trends. Control proach to quality control:
charts graph measurement variations over
time and range, indicating the upper and 1. Inspection does nothing to improve the
lower control limits of the normal range of process, and is not very good at sorting good
variation. If the centerline corresponds with from bad.
the quality target, then the chart can identify 2. Sampling plans developed to support an
when deviation occurs and whether a process acceptable quality level (AQL) of 5 % , for ex-
is under statistical control. Variable control ample, say that a company is content to ship
 13 Statistical Process Control and Quality Control 1135

5% defectives. How many customers would 2. Information about performance. To be

accept 5% defectives these days? able to improve the process, you need infor-
mation. Useful information can be derived
As a point of interest, it has been shown from studying the final output-detection in
that 100% inspection is at most 80% effective, the old sense. But studying what is usually
and 200% inspection is usually less reliable called “intermediate outputs” is just as im-
than 100% inspection. portant and helps with prevention. Infor-
You may have heard “quality if free.” This mation on the operating state of the pro-
can never be true if your approach to quality cess, such as temperatures, cycle time, or part
improvement is to add more inspectors. weight-provided that it has been properly
gathered and interpreted-can show whether
Defect detection forces quality control
and where action is required to correct the
and production departments to be adver-
process and/or the most recent output. This
saries. The old role, which casts the quality-
is commonly done with statistical methods.
control engineer as the policeman, must be
But if timely and appropriate actions are
replaced with the philosophy that quality
not taken, any information-gathering effort
control and production are members of the
is wasted.
same team. The quality department must
advise and serve the production depart- 3. Action on theprocess. This describes the
ment relative to control of the process and action you take, based on the information
reduction of process variability. you have gathered, to prevent the produc-
e Many items cannot be inspected at all. Ex- tion of out-of-specification products. This ac-
amples are products that must be destruc- tion could range from operator training, using
tively tested, require very costly and time- different raw materials, or changing process
consuming performance testing, or are so conditions, to buying more up-to-date man-
numerous that testing to meet an AQL re- ufacturing equipment or redesigning a tool.
quires an enormous sample size. The effect of such actions should be moni-
tored, and further analysis and action should
Rather than detect problems after they be taken, if necessary.
have occurred, statistics preaches the virtue
4. Action on the output. This involves de-
of prevention. And statistical methods are
tecting out-of-spec output already produced.
little more than a way of institutionalizing
When that happens, you have to go back
prevention of a “Do-it-right-the-first-time”
to the old time-consuming methods of sort-
mindset. This attitude should extend to all
ing, scrapping, or reworking. This will con-
aspects of manufacturing, from product con-
tinue until corrective action on the process
ception, resin formulation, and compounding
has been taken.
to final processing, assembly, and shipping.
Preventive maintenance of equipment is just
as much part of a successful SP/QC program
Gathering and Analyzing Data
as the online monitoring of process variables.
Practitioners and teachers of SP/QC fre-
The heart of SP/QC is “action on the pro-
quently call what happens with statistical
cess.” This involves proper statistical gather-
methods a “feedback system.” That system
ing and analysis of information on the pro-
is divided into four basic parts:
cess and the ability to draw conclusions for
the proper reaction. The key terms here are
1. Process. This is the combination of peo-
common and special causes, local actions and
ple, equipment, raw materials, methods, and
actions on the systems, process control and
environment that work together to produce
process capability, and control charts. Here
a product. How well the process performs in
are the details:
terms of the quality of output and productiv-
ity depends on the way the process has been Common and special causes. Teachers of
designed and is operated. statistical methods stress again and again
1136 13 Statistical Process Control and Quality Control

Pieces vary from each other:

Size + Size + Size+ Size-

But they form a pattern that, if stable. is called a distribution

Size‘ Size Size+

Distributionscan differ in:

Location Soread

Size+ Size 4 Size +

...Or any combination of these.
Fig. 13-5 Variations resulting from common causes and special causes.

that quality problems are the result of vari- acterized by location (typical value), spread
ation, such as in raw materials or process (amount by which the smaller values differ
conditions, and statistical analysis will help from the larger ones), and shape (the graph of
explain the causes of this variation. Two the variation-whether symmetrical, peaked,
types of causes of variation have been iden- etc.). Common causes are often defined as
tified: common and special causes. a source of variation that affects all individ-
ual values of the process output being stud-
Any process contains a multitude of vari-
ied. The causes of variation are random, like
ables. For instance, the dimensions of a
throws of the dice. But if the process is left
molded product can be affected by changes in
to produce parts continually without change,
the resin’s specific gravity or flowability due
the variation will remain. The causes cannot
to batch-to-batch inconsistency, inconsistent
be altered without changing the process it-
regrind ratio, operator inattention, tool wear,
self. Statistics provides us with ways of recog-
pressure changes, mold surface temperature,
nizing variation due to common causes (see
clogged dies or nozzles, or outside temper-
Figs. 13-5 and 13-6).
ature and humidity changes, to name only
some common variables (see Chaps. 7 and 8).
SP/QC will help you trace the problem that
led to the part being out of spec. The first
step is to distinguish between “common” and
“special causes.”
For example, Ford Motor Company’s
SP/QC manual defines common causes as the
many sources of variations within a process
/
that is in statistical control. Collectively they Size-b
behave like a constant system of chance. Al- Fig. 13-6 Example of common causes. If only
though individual measured values are all dif- common causes of variations are present, the out-
ferent, as a group, they tend to form a pat- put of a process forms a distribution that is stable
tern. This pattern of distribution can be char- over time and predictable.
 13 Statistical Process Control and Quality Control 1137

normal dice, you keep on playing until

your luck (a common cause) changes; with
loaded dice (a special cause), you take ap-
propriate action against the crook (pro-
vided that you can analyze the play results
sufficiently to identify the special cause).
Local actions are usually required to elimi-
nate special causes of variation, for example,
s i 2 4
training an operator, getting a new tool, or
Fig. 13-7 Example of special causes. If only spe- changing the resin. Local actions can usually
cial causes of variations are present, the process be taken by people close to the process, such
output is not stable over time and is unpredictable. as machine operators. And, as SP/QC experi-
ence consistently proves, local action can cor-
rect about 15% of all process problems.
Special causes (often called assignable
Action on the system is usually required to
causes) are factors causing variation that lies
reduce the variation due to common causes,
outside the normal, consistent distribution
which, like special causes, can be determined
of the process output. “Unless all the spe-
by simple statistical methods. But finding out
cial causes of variation are identified and cor-
what those common causes are often requires
rected, they will continue to affect the process
a good deal of analysis. In almost all cases,
output in unpredictable ways,” the Ford man-
management action is required to correct
ual states. On a control chart (defined below),
common causes, since they are inherent in the
special causes are signaled by a point beyond
process itself as presently constituted.
the control limits. A special cause could be
Buying new machinery or adding process
a change in resin, excessive tool wear, or a
controls to existing equipment would be ex-
new operator. A special cause would disturb
amples, as would be any fundamental change
a process to such an extent that it draws atten-
in shop practices aimed at improving qual-
tion. A pair of loaded dice (a special cause)
ity. Manufacturing experience has shown that
would disturb the standard variations (com-
about 85% of all process problems must be
mon causes) in a dice-throwing game (see
corrected through management action. Ex-
Figs. 13-7 and 13-8).
perience also shows that the people clos-
0 Local action and action on the system. est to the process-machine operators-are
There is an important connection between often best suited to identify the nature of
the type of cause of a variation and the type the problem. Thus, good communication be-
of action required to counteract it. To re- tween management and line personnel is
turn to the dice-throwing example: With crucial.

In Control
(SpecialCauses Eliminated)

Out of Control
(SpecialCauses Present)

Fig. 13-8 Process control provides the means of reporting “in-control”and “out-of-control” products.
1138 13 Statistical Process Control and Quality Control

Process Control and Process Capability will keep a manufacturing process within pre-
determined limits of variation. Variation it-
The frequently used terms “process con- self cannot be eliminated. Even with the best
trol” and “process capability” must be thor- machine setup, resin, and equipment, you will
oughly understood to succeed with statistical continue to have minute variations from part
methods. to part. But with the elimination of special
Process control, in the context of SP/QC, causes, variation can be brought within ac-
means maintaining the performance of a pro- ceptable limits.
cess at its best level. Process control involves Process capability, according to the Ford
a range of activities such as sampling the pro- manual, is often thought of as the propor-
cess output (the product), charting process tion of product output that is within speci-
performance, determining the causes of any fications. Since a process in statistical control
problems, and correcting them. can be described by a predictable distribu-
Process capability is the level of product tion, the portion of out-of-spec parts can be
uniformity that a process is capable of yield- estimated from this distribution. As long as
ing. For instance, a certain blown film line the process remains in statistical control, it
may be capable of producing products with a will continue to produce the same proportion
thickness variation of f0.001 in. (0.0254 mm) of in-specification parts. Management must
under optimum conditions. Process capabil- take action to reduce the variation due to
ity must be expressed by the percentage of common causes, in order to change the dis-
defective products or the range or standard tribution and improve the process’ ability to
deviation of some product dimensions or meet specification.
weight. Process capability is usually deter- “Once you are operating in control,” says
mined by performing measurements on some one molder, “the real challenge starts. You
or all the products produced by the process keep on tightening control limits bit by bit.
(Fig. 13-9). You start out, for example, with a tolerance
“Aprocess is said to be operating in statisti- of 10.2 mil thickness. You tighten your limits
cal control when the only source of variation to 10.09 or less. It is a way of continuously
is common causes,” the Ford manual says. improving quality, and it never ends.”
Deming, in a 1975 technical paper, stated
further: “But a state of statistical control is
not a natural state for a manufacturing pro- Control Charts
cess. It is instead an achievement, arrived at
by elimination, one by one, by a determined Control charts are a powerful tool used by
effort, of special causes of excessive varia- all SP/QC practitioners. A control chart can
tion.” In other words, good process control be a simple piece of paper filled in at the

In Control and Capable

(Variationfrom Common Clauses
Reduced)

Size-
In Control
Control, but not Capable
(Variationfrom Common Causes
Excessive)

Fig. 13-9 Examples of process capability.

 13 Statistical Process Control and Quality Control 1139

molding machine with a pencil stub, or it can mine critical variables and potential rational
be a sophisticated CRT display adjusted with subgrouping. During implementation, moti-
a few keystrokes. vational aspects should be considered and
Control charts are used to (1) gather in- can often be accomplished by using a team
formation; (2) calculate control limits; and approach that involves operators and fore-
(3) calculate the process capability or the best men as much as possible. To sustain interest,
that the process can do. Construction of the charts must be changed over the life of the
chart, frequency of updating, and type of data application. Eventually, of course, when con-
entered obviously vary from process to pro- tinued control is assured, the charts should
cess, product to product. But in any single be withdrawn in favor of spot checks as ap-
case, the progression is the same. propriate. This is seen in Table 13-2.
You have to collect data and plot them In initiating control charts, certain con-
to see what is happening. The type of data siderations are paramount, including ratio-
you collect has to be carefully determined nal subgrouping, the type of chart, charting
ahead of time. You then analyze the data to frequency, and the type of study being con-
determine what the natural variation of the ducted. A check sequence for implementing
process is; this will tell you the amount of vari- control charts is shown in Fig. 13-10.
ations that should be expected if only varia- Control charts are not a cure-all. Using
tions from common causes were present. You them properly requires a great deal of time
can also determine if corrective action taken and effort. Moreover, they are not appropri-
to eliminate special causes actually works. ate in every situation to which statistical qual-
You then can determine process capability- ity control can be applied. A small firm with
the best you can do with present common numerous small job-shop vendors would be
causes-and compare actual results with that hard put to insist on process control as the
optimum capability. These three steps-data source for acceptance of products, since few
collection, data analysis, and quantification pieces are made and purchased at any given
of common causes-are repeated for contin- time. In this case, acceptance sampling would
uous process improvement. be the method of choice. On the other hand, a
Control charts show when action should be big firm that receives a large amount of prod-
taken and, equally important, when no action uct from only a few vendors would be well
should be taken. In other words, when the advised to work with the vendors to institute
process is consistent, it should be left alone. process control at the source, thus relieving
After you have achieved consistency and the necessity of extensive incoming product
the process is thereby in what is called “sta- testing.
tistical control,” you can begin to tighten con-
trol limits by eliminating common causes.The
result can be charted again on your control Defect Prevention
charts.
“Control charts,” states the Ford manual, The difficulties in detecting defects have
“provide a common language for communi- driven many companies to statistical process
cations about the performance of a process- control-defect prevention. Here, the pro-
between the two or three shifts that oper- cess is monitored and statistical methods are
ate a process, between the line production used to control process variation, enabling
(operator, supervisor) and support activities adjustments to the process before defects are
(maintenance, material control, process en- produced.
gineering, quality control), between different This, however, does not mean that you no
stations in the process, between suppliers and longer inspect finalproduct. It does mean that
user, between manufacturing/assembly and the objective of inspection has changed from
the designlengineering activity.” sorting good from bad to providing assurance
There is a life cycle in the application of that process control activities are effective.
control charts. In the preparatory stage, an in- A process in control means that the pro-
vestigation of the process is required to deter- cess is being impacted only by random (or
 1140 13 Statistical Process Control and Quality Control

Table 13-2 Life cycle of control chart applications

~~~~ ~

Stage Step Method

Preparatory State purpose of investigation Relate to quality system
Determine state of control Attributes chart
Determine critical variables Fishbone
Determine candidates for control Pareto
Choose appropriate type of chart Depends on data and purpose
Decide how to sample Rational subgroups
Choose subgroup size and Sensitivity desired
frequency
Initiation Ensure cooperation Team approach
Train user Log actions
Analyze results Look for patterns
Operational Assess effectiveness Periodically check usage and relevance
Keep up interest Change chart, involve users
Modify chart Keep frequency and nature of chart
current with results
Phase-out Eliminate chart after purpose Do spot checks, perform periodic
is accomplished sample inspection, overall p, c charts

common) causes and all assignable (or spe- Lack of process capability
cial) causes have been found and eliminated. 0 Lack of management commitment to qual-
Random (or common) causes are due to ity
many small influences that affect a measure- 0 Lack of proper training
ment. Assignable (or special) causes are due These shortcomings have resulted in many
to one or more large influences that must be
companies moving toward phase 111: total
eliminated. Assignable causes are those that
quality control or companywide quality con-
show up on a control chart as a point out
trol.
of control or nonrandom sequence of points.
This means that quality is everyone’s busi-
The major advantages of a process in control
ness, and it is clearly management’s responsi-
(a stable process) include:
bility to integrate and lead the various func-
The process is free from all assignable tions within an organization toward the goals
causes. This means the process is predic- of customer-first orientation and commit-
table and statistical methodology can be ment to quality. Vital to total quality control
used to make decisions concerning the pro- is the recognition that each individual is both
cess (e.g., what percentage of production a customer of a preceding operation and a
will be out of specification). supplier to a subsequent operation, and must
Since the process is predictable, one can be trained and motivated to serve these cus-
reduce inspection and increase the confi- tomers.
dence in certification of product properties.
0 The capabilities of the process are defined.
Information is generated on how to im-
prove the process. Understanding Modern Methods of Control
Process stability should increase sales and
decrease customer complaints. The greatest obstacle to the use of mod-
ern methods of control is the mistaken idea
Using SPC to achieve process stability is
that they are too difficult for the average
not a cure-all. SPC cannot deal effectively
person to understand. Now we will admit that
with the following situations:
the theory on which modern quality control is
Basic design errors based does involve some high-powered math-
Incomplete or improper specifications ematics, but you d o not have to be a graduate
 I 3 Statistical Process Control and Quality Control 1141

,- ,, . --.,,~
I . . . - . . L - . .

Fig. 13-10 Check sequence for control chart implementation.

mathematician to use these modern methods Let us go back in the kitchen where the
or understand them. customers will not be too curious and weigh
Let us start by taking a piece of pie. Just as 100 pieces of pie. Of course, you were right
expected, you reached for the biggest piece. about there being big and little pieces. But
It shows that you understand the first rule do you observe that the number of pieces in
of statistics: There is an inherent variability each gram step varies from the smallest to
in even a very good product, if you have a the largest piece in a fairly regular and sym-
means of measurement sensitive enough to metrical pattern? In fact, when you examine
detect such a variation. How do you know Fig. 13-11 you see that an evenly balanced
that’s the biggest piece? Maybe by looking at distribution curve exists. A smooth curve has
them, you can classify them as big and little been drawn that results in an area under the
pieces. It would be difficult for you to arrange curve that very closely fits its particular dis-
the whole counter in order of size. tribution (bell-shaped pattern).
1142 13 Statistical Process Control and Quality Control

Fig. 13-12 Normal curve has a change in direc-

175 180 GRAMS 185 190
tion; curve stops curving downward and starts
Fig. 13-11 Distribution curve. curving upward.

What would happen if we measured an- Standard Deviations

other lot of 100 pieces? We would get a very
similar pattern in any lot. This is a pattern that From a technical viewpoint, the standard
repeats itself endlessly, not only for pieces of deviation is the distance from the center to
pie but for most manufactured articles, and where the curve stops curving downward and
even in nature. But there is always an inher- starts curving upward. For most purposes,
ent variability, provided we have a measur- however, we are interested in only the points
ing instrument sensitive enough to find these at one, two, and three standard deviations,
variations. And these variations usually fol- measured from the center. If we measure one
low the same bell-shaped pattern, called the standard deviation on each side of the cen-
“normal curve.” ter of this curve, 68% of the area will be be-
Here is an example that can be easily un- tween the lines drawn through these points
derstood. The height of men in any organi- (Fig. 13-13).
zation or army averages 67.7 in. (172 cm), The table tells you that the standard
but individuals vary all the way from 60 to 76 deviation of men’s height is 2.6 in. (6.6 cm)
in. Sixty-eight percent are between 65.1 and so by simple subtraction and addition, you
70.3 in. (165 and 179 cm). Ninety-five per- know that 68% of the men are between 65.1
cent measure between 62.5 and 72.9 in. (159 and 70.3 in. (165 to 179 cm) (67.7 - 2.6 = 65.1
and 185 cm). (Different height ranges exist +
and 67.7 2.6 = 70.3) (Fig. 13-14).
for women in the army.) These data are ob- Two standard deviations would be 5.2 in.,
tained from Table 13-3, which reports men’s so 95% of the men would be between 62.5
heights. and 72.9 in. (cr = 2.6 x 2 = 5.2; thus, 67.7 -
The table lists X (pronounced “ex-bar”), 5.2=62.5and67.7+5.2=72.9(Fig. 13-14).
which gives the average height of 67.7 in. Three standard deviations are 7.8 in.
(172 cm). It is the middle of the curve (Fig. 99.73% of the area of the normal curve is
13-12) where the largest percentage of men between -3 sigma and +3 sigma, so we
exist. The table also lists the Greek letter cr can say that practically all the men will be
(sigma), known as the standard deviation. between 59.9 and 75.5 in. (a = 2.6 x 3 = 7.8;

Table 13-3 Human proportions (in inches)

Men Women
x m x u

Height (standing) 67.7 2.6 62.5 2.4

1 ’ 1 1
Height (sitting) 36.0 33.9 1.2 I l l
Length of foot 10.1
Span
Forearm
69.9
18.3
3.1
1.0
-
16.3
,-.,
-,
I+ 68%4

Fig. 13-13 One standard deviation on each side

of the center.
 13 Statistical Process Control and Quality Control 1143

crometer. What you recorded is shown in Fig.

13-15 and is referred to as a frequency distri-
bution. As you see, the parts vary in the same
bell-shaped pattern as the pieces of pie and
height of soldiers.
The average is 2.00 in. (5.08 cm), but, indi-
vidually, the pieces vary from 1.91 to 2.09 in.
(4.85 to 5.31 cm). This variation is acceptable
to our customers since their specificationsre-
quire an average of 2.00 in., with a tolerance
59.9 62.5 65.1 67.7 70.3 72.9 75.5 of f0.15 in. (0.38 cm). That means they will
Fig. 13-14 One, two, and three standard devia- accept anything between 1.85 and 2.15 in. (4.7
tions (Gaussian distribution). to 5.5 cm). Our problem is to keep the cutting
machines at the center of 2 in. and not let the
thus, 67.7-7.8=59.9 and 67.7+7.8=75.5) individual pieces vary much more than they
(Fig. 13-14). did in this lot.
Of course, there are some men taller than All these parts were made at one time, on
75.5 in. and some shorter than 59.9 in., but the same machine, by the same operator, and
they amount to only about 15 each in 10,000. from the same lot of material, so the pattern
Next we review probability. We walk down of their variation would almost exactly fit the
the street and measure the height of every usual normal curve.
male soldier that passes. What odds would We would also know that the greatest indi-
you give that he would be 67.7 in. tall? Of vidual variations would be quite close to -3
course, you would have to give big odds. Ob- and +3 standard deviations. That is a spread
viously, the soldier would probably be from of six standard deviations. The difference be-
59.9 to 75.5 in. (Fig. 13-14). tween the longest at 2.09 and the shortest at
But suppose I estimate that the next sol- 1.91 is 0.18 in. (0.46 cm). One-sixth of 0.18 is
dier would be between 65.1 and 70.3 in. How 0.03 in. (0.08 cm), so the standard deviation
would you figure the odds? I have already is 0.03 in. (0.08 cm).
given you the answer to that question when Next pick out a random sample of five parts
saying that 68% of all the soldiers are be- from a bucket. While I measure these five
tween 65.1 and 70.3 in. (Fig. 13-14). There- pieces, get me four more samples using the
fore, there are 68 chances in 100 that the
next soldier would be between those limits. -
I would have to give you odds of 68 to 32, or =.B=
about 2 to 1. E $r
With 95% of all soldiers between 62.5 and
s e rr,
ZZ1H'f
72.9 in. (159 and 185 cm), the odds are 95 to
5, or 19 to 1, that the next man will not be
T+B%'f -
PZ€Z€€Z
shorter than 62.5 in., nor taller than 72.9 in. ZZITiEZ2
There are only 2; chances in 100 that he will 5 s z z rzzr h
be taller than 72.9 in., and 13; chances in Z ~ ~ 3 %~ ~ Z 3
,EZIZ€%BZ€=
10,000,or 1 in 740, that he will be taller than ZS~Z3+ZHZEZ
72.5 in. (184 cm). Zt%ZZZHZEB~Z~
- = 31% 3 € Z z z z z Z f % Z Z Z= -
1.91 1.93 1.95 1.97 1.99 2.01 2.03 2.05 2.07 2.09
Frequency Distribution 1.92 1.94 1.96 1.98 2.00 2.02 2.04 2.06 2.08
LENGTH IN INCHES
Next consider cutting a lot of blow-molded Fig. 13-15 Frequency distribution of 500 parts
parts exactly 2 in. long. You have already cut targeted to all be cut to 2 in. (5 cm); they vary
a lot of 500 and measured them with a mi- from 1.91 to 2.09 in. (4.85 to 5.31 cm).
1144 13 Statistical Process Control and Quality Control

Table 13-4 Sampling parts

Sample 1 Sample 2 Sample 3 Sample 4 Sample 5
2.016 2.025 2.002" 1.973 2.033
2.027" 1.963' 1.988 2.046" 2.033
1.994 2.015 1.999 1.941 2.037"
1.954' 2.059" 1.996 2.001 1.968'
1.985 1.990 1.978' 2.009 2.013
Average 1.995 2.010 1.993 1.994 2.011
Range 0.073 0.096 0.024 0.105 0.069
Grand average 2.001
Range 0.0734
a Largest measurement in sample.
'Smallest measurement in sample.
same random sample approach. The results important characteristics about the distribu-
of all these measurements are given in Table tion of the measurements of any product: the
13-4. average X and standard deviation (T.Instead
These five sample averages vary quite a bit of stumbling over the word "standard devia-
from the actual average of 2.00 in. (5 cm), tion," we will call it by its Greek name, sigma.
but the grand average is 2.001. We would not The sigma of our parts is 0.03 in. (0.08 cm)
always come quite so close to the actual av- and the average is 2.00 in. (5 cm). Knowing
erage in five samples, but if we have enough these two values, we can now put some per-
samples, say 20 or 25, the average of a num- centage figures on our old friend, the normal
ber of samples will be a very good estimate curve (Fig. 13-16).If we are doing a good job
of the actual average of the lot. of cutting our parts to the specified average
Now, you will notice that I have advanced of 2.00 in., 68% should be between 2.00 in. -1
another figure for each sample. I have labeled sigma and 2.00 in. +1 sigma, or between 1.97
it the "range." It is the difference between and 2.03 in. Also, 95% should be between -2
the largest and smallest measurement in each sigma and 1-2 sigma, or 1.94and 2.06 in. Thus,
sample. These range figures also vary for the 99.73% should fall between -3 sigma and +3
several samples. The average range is 0.0734. sigma, or 1.91 in. and 2.09 in.
In our lot of 500 parts, we divided the range Even if something very unusual happened
by six to obtain the standard deviation, but we to give me an occasional part that was shorter
cannot do this in samples of only 5. In the big than 1.91 in. or longer than 2.09 in., as long
lot, we had 500 chances of acquiring some of as all of my part-cutting machines are prop-
the extreme values. In a sample of 5 , there erly adjusted to make an average of 2.00 in.,
is much less chance of getting these extreme I could give you odds of 997 to 3 that you
values, so the divisor is much smaller for small
samples.
This divisor is called the d2 (d two) factor,
and we can find the appropriate factor for any
sample size in a table that is very familiar to
any quality-control engineer.
For samples of 5, the d2 factor is 2.326.
Dividing the average range, 0.0734, by 2.326
gives us a standard deviation of 0.0315, which
does not differ much from the figure of 0.030
that we got by measuring 500 pieces.
This demonstrates that, by taking relatively 1.91 1.94 1.97 2.00 2.03 2.06 2.09

few samples, we can discover these two very Fig. 13-16 Distribution curve of parts.
 13 Statistical Process Control and Quality Control 1145

about the center of 2.04 in. was the same

as in any good lot, except that the bell-
shaped curve was shoved over 0.04 in. Sigmas
were still the same, 0.03 in. The three-sigma
limits of his distribution were now changed
to 1.95 and 2.13 in. They are still within
the customer's specification of f0.15, but the
I I *

1.95 1.98 2.01 2.04 2.07 2.10 2.13 customer would probably squawk about the
higher average (Fig. 13-17).
Fig. 13-17 Standard deviations for parts.
We have already seen that a good lot can
be expected to vary between 1.91 and 2.09
could not walk out in the factory and pick up
in. Five percent of the parts Bill was making
one at random that would be outside of these
were over the limit of 2.09 in. (Fig. 13-18).You
limits.
would have to use a statistical table to figure
Both you and I know that machines, ma-
that out. If we were to pick up single pieces
terials, and operators do not always do what
from Bill's machine, there are only 5 chances
they are supposed to do. Take Bill Jones on
in 100, or 1 chance in 20, that we would find
the number 2 machine as an example. Bill's
one of the oversized pieces.
stock market investments have experienced
Even if we were lucky enough to find one of
price volatility; and Bill has been feeling
these long pieces, Bill would probably claim
rather frustrated. His mind is simply not fo-
it was just one of those freak accidents that do
cused on cutting parts. Yesterday his machine
sometimes happen. But we are able to pin the
setting was off about four-hundredths of an
problem on him by showing him his control
inch (1 mm).
chart (Fig. 13-19).
How did we find that out? Well, you do not
find four-hundredths of an inch by looking
at the measurement of just a few individual
Control Chart
parts.
For the sake of argument, let us assume
What? You do not know what a control
that his average was 2.04 in. The pattern
chart is? Go back to the Army! Do you re-
of the variation of individual measurements
member that I was willing to give you odds of
740 to 1that the next male soldier to pass our
Bill's
Distribution door would not be taller than 75; in. (191.8
cm). I was just betting that he would not ex-
ceed the 3-sigma limit.
Let us change our bet to the average of the
next two soldiers. Did you ever notice that an
unusually tall chap usually has a buddy who
is on the short side? It is a safe bet that the
average of the two will be nearer the grand
1.91 2.00 2.04 average of 67.7 in. than the extreme height of
Fig. 13-18 Good lot versus Bill's lot. 75.5 in.

7 8 9 10 11

Fig. 13-19 Control chart of parts; sample size of five.

1146 13 Statistical Process Control and Quality Control

For the same odds of 740 to 1, I would GOOD BILL‘S

DISTRIBUTION DISTRIBUTION
have to use narrower 3-sigma limits. If we
bet on the average of the next four soldiers,
would we have to use still narrower limits?
No! These limits are not reduced in propor-
tion to the number of soldiers.
Limits for the average of two soldiers
would be about two-thirds those for individu-
1.96 2.00 2.04 2.08
als. For the average of four soldiers, the limits
would be one-half those for individuals; and Fig. 13-20 Averages; samples of five.
for averages of ten soldiers, the limits would
be about one-third those for individuals. ter, or between 1.96 and 2.04 in. As long as
How does one obtain such figures? Ele- our sample averages vary between these lim-
mentary! Not only do measurements of in- its, we can be reasonably sure that the actual
dividuals vary in the bell-shaped pattern of average of the machine is close to the speci-
the normal curve, but sample averages vary fied average of 2.00 in. It is a 740 to 1bet that
in the same pattern, about the same center. no sample average will be higher than 2.04 in.
However, the standard deviation of sample (5.18 cm).
averages is the sigma of the individual, di- But if Bill’s machine setting is 2.04 in., av-
vided by the square root of the sample size. erages of samples of five will vary 0.04 in. to
In the case of our soldiers, the sigma of indi- either side of this higher center. There is now
vidual heights is 2.6 in. For each sample size, a 50-50 chance that any sample average will
we divide 2.6 in. by the square root of the be above 2.04 in., instead of the 1 chance in
sample size. You have forgotten how to de- 20 that an individual part will be longer than
termine the square root? So have I. We will 2.09 in. (Fig. 13-20).
just look it up in a table. Of course, there is the same 50-50 chance
Now, let us apply this principle to measur- that the sample average will be below 2.04 in.,
ing the parts from Bill Jones’ machine: We but with these odds in our favor, we are prac-
take samples of five parts from his machine tically certain to obtain a high average in the
every 15 min and measure them very care- first two or three samples.
fully with a dial micrometer. We have seen Well, that is exactly what happened with
that the standard deviation of this measure- Bill’s machine. The Chinese have a proverb:
ment is 0.03 in. (0.08 cm). Dividing this by Confucius says “One picture is worth ten
the square root of five (2.236), we get a stan- thousand words.” Thus, Fig. 13-19, with its
dard deviation of 0.0134 in. for averages of control chart, is the “picture.”
five parts. Three standard deviations would Its centerline says that the parts should av-
be 0.0402 in. (Table 13-5). erage 2.00 in. The upper and lower dotted
If Bill’s machine is set right, for an average lines say that if this specification is met, av-
of 2.00 in., averages of five should vary not erages of samples of five should be not less
more than 0.04 in. above and below this cen- than 1.96 in., nor more than 2.04 in.

Table 13-5 Standard deviation versus sampling size

3-Sigma Limits
Square Root of 2.6 Divided by 3 Times Sigma Shortest Longest
Sample Size Sample Size Square Root of Averages (677 - 30) (677 + 3a)
2 1.414 1.84 5.52 62.18 73.22
4 2.000 1.30 3.90 63.80 71.60
5 2.236 1.16 3.48 64.22 71.18
10 3.162 0.82 2.46 65.24 70.16
 13 Statistical Process Control and Quality Control 1147

From seven in the morning until eleven centerline for most of the control charts for
o’clock, the sample averages zig-zagged averages. [As previously reviewed, the stan-
between these limits in a pattern that we usu- dard deviation is a measure of the dispersion
ally expect. At eleven o’clock, the cutting (scatter) about the mean (average). It is de-
tool broke, and Bill put in a new one. Bill fined as the square root of the mean sum of
was thinking about the snappy comeback he squares of deviations about the average. For a
should have made in last night’s argument normal distribution, f 3 standard deviations
with his mother-in-law and forgot to check from the mean includes approximately 99.7%
the setting carefully. The new setting was ac- of the population.]
tually at 2.04 in. PPG has adopted the standard deviation
At eleven o’clock, Nick Rosato, the qual- since range is a poor estimator of variation
ity inspector, measured five parts from Bill’s for sample sizes greater than 10. It uses the
machine. The average seemed fairly high, but nominal value because it is usually an easy
it was inside the upper control limit, so Nick task to control to a target value. The real chal-
plotted it on the chart and let it pass. lenge is to control variation (Fig. 13-22).One
At eleven-fifteen, he took another sample. statistic that PPG finds very useful is known
It was also inside the upper limit, but Nick as process capability. This is calculated as
knew that the odds of two samples being so
close to the upper limit were a lot slimmer 6x3
for short-term capability
than people figure. So he did not wait an- Tolerance
other 15 min. He took another sample right and
away, and sure enough, that one was over the
6xS
3-sigma limit. Just to be sure, he took one for long-term capability
more. That was also over the 2.04-in. limit. Tolerance
Even Bill could not argue against the evi- where = the variation of the product
dence of the four samples,particularly the last property over a short period of
two. He shut down the machine and corrected time, usually the average
the cut. Nick stuck around and took some standard deviation between
more samples to be sure they were within ac- individual specimens within the
ceptable limits and then went back to the reg- sample selected for plotting on
ular 15-min schedule of sampling. process control charts
The control chart (Fig. 13-21) shows that S = the standard deviation between
Bill is back in the groove again. individuals taken over a long
period of time
Standard Deviation versus Range For example, specifications for the binder
content of a particular PPG roving are
PPG Industries handles two things a lit- 2.10 f 0.20 and total tolerance is, therefore,
tle differently than most companies by using 0.40. (An example of PPG’s glass fiber rov-
the standard deviation, instead of range, to ing plastic reinforcement is not generally in-
control variation and the nominal (or target) volved in blow molding to date but is used to
value rather than the process average as the provide an approach in considering SPC used

Fig. 13-21 Control chart back under control; sample size of five.
1148 13 Statistical Process Control and Quality Control

PERCENT BINDER
2.40

2.25

A
V
E
R
A
G
E

S D 0.25-
T E 0.20-
A V
0 . la-
N I
D A 0 , (0 ...................................... _____________ _________-__----
~ ~ -----------.
~

A T
R I
D O
N

Fig. 13-22 Typical PPG roving reinforcement using 3-sigma control limits for percent binder ( N = 6).

slightly differently.) Short-term capability is tual value of a product property (e.g.,molding

calculated to be weight) from the mean or set value is called
6(0.05) straggling.The actual values are said to strag-
= 0.75 or 75% gle or vary around the mean value.
0.40
The causes for straggling are either acci-
For this product, the standard deviation dental or systemic. Systemic deviations can
over a 2-month period was 0.065, so long- be attributed to causes that act in a definite
term capability is way and can always be removed or offset
fj(0.065) (e.g., changed raw-material properties, ma-
= 0.975 or 98% chine faults, or operator errors). Accidental
0.40 deviations are usually caused by very many
A word of caution is in order: Process ca- different factors. These deviations cannot be
pability only has meaning if the process is in completely removed.
control. The value of this statistic is that it
gives management one number that tells if
the process is capable of producing a product Mean Value, Range, and Standard Deviation
within specifications.If the process is not ca-
pable, then management must take action to The arithmetic mean value is calculated as
decrease variability or have the specification
changed. Reduction of variation usually re- xq =
XI x2 . ' . x, + + +
n
quires a fundamental change to the process,
since the process should be in control before - xy=1
-- xi
the PC is calculated and is, therefore, doing n
the best it can. where x q = arithmetic mean value
xi = measured individual value
Basic Statistical Concepts n = size of sample

To reduce testing for quality-related prod-

uct properties, statistical methods can be
used. They are employed to assess samples
of the product to make statements about the
total output (159). The deviation of the ac-
 13 Statistical Process Control and Quality Control 1149
measured value

Xl
X
XI-xq
v X
xq
X X
X

1 measured value no.

Fig. 13-23 Standard deviation.

where s = standard deviation Distribution

xi = measured individual value
xq = arithmetic mean value For each process, there is a scatter diagram.
n = size of sample One method of displaying the scatter dia-
gram is the bar chart or histogram. The his-
Range is defined as the difference be- togram shows the distribution of the abso-
tween the maximum and minimum individual lute and relative frequencies of values (Fig.
values found within the sample: 13-24). In other words, it indicates how often
a certain measured value occurred in a sam-
R = x m a x - Xmin ple (e.g., how often the value of the melt cush-
where R = range ion was2.3 mm). The more samples are taken,
xmaX= maximum value within the smoother and steadier the resulting curve.
a sample In nature and engineering, the bell-shaped
xmin= minimum value within Gaussian distribution curve is most frequent.
a sample The standard deviation can be found as
the point of inflection of the bell curve
Standard deviation is a measure of the (Fig. 13-25). Once the mean value and stan-
straggling of a process. The wider the strag- dard deviation of a normally distributed sam-
gling, the higher the standard deviation. It is ple have been calculated, the share of total
the mean value of the deviation of the individ- production that lies between two limits can be
ual values from their mean value (Fig. 13-23). deduced. If, for example, the samples lie in a

frequency

measured value
Fig. 13-24 Histogram.
1150 13 Statistical Process Control and Quality Control

4 99.994% t' Standard deviation

4 99.13% c UCLS = B4 x sq
I+ 95.44% d
LCLS = B3 x sq
+68 26% --D
with

n
as the mean value of n sample standard devi-
7 - 4 s
ations. The factors A3, B3, and B4 from the
factor table of Ford Q l O l depend on the size
Fig. 13-25 Gaussian distribution. of the sample.

range of xq = f 2 x s, this means that 95.44% Machine Capability

of total production only deviates 2 x the stan-
dard deviation from the mean value. Critical machine capability is expressed as

Zrit U S L - LSL
Process Control Chart c, = - cm =
3 ' 6xs

The process control chart is a two- The machine capability index cm includes
dimensional coordinate system. On the x axis, only straggling. The index Cmk also includes
the time of sampling or sample number is en- the position of the mean value relative to
tered. The y axis shows the mean value, range, specification limits. Zritstands for the dis-
or standard deviation of this sample. For vari- tance from the nearest specification limit ex-
able properties, there are xqlRor xqls process pressed in standard deviations:
control charts. On an xqlR chart, the range R
is entered in the lower part. The standard de- U S L - xq xq - LSL
viation s says more about the straggling of Grit = or
S S
the process. Because s is more difficult to cal-
culate manually, s charts are typically used The minimum requirement for machine ca-
when the data are recorded by computer. pability is that straggling must be xq f 4 x s
The control limits are calculated as follows within the specification, that is, 99.994% of
(UCL = upper control limit, LCL = lower the parts manufactured are expected to lie
control limit): within the tolerance limits.
Mean value The index Cmk must be at least 1.33. If this
UCLxq = xqq A3 x sq + condition is fulfilled,the process is under sta-
tistical control, that is, there are no more sys-
LCLXq = Xqq - A3 x S q temic influences.
with
Process Capability

The determination of process capability is

used to find accidental influences on the pro-
cess. A condition for this is that the process
as the mean value of n mean values = the be under statistical control. The indices of ca-
process mean value. pability are defined as follows:
 13 Statistical Process Control and Quality Control 1151

LSL specification pro, USL

average ave

Fig. 13-26 Process off mean value.

Process capability this, these process parameters do not show

USL - LSL a Gaussian distribution. Suitable parameters
cp =
6x6 are, for example,
with the estimated value of standard devia- Maximum interior mold pressure
tion Melt cushion
o = -sq
L

Injection time
c4
Dosing time
critical process capability (Fig. 13-26)

cpc = -
3 Importance of Control Charts
USL - xqq xqq - LSL
Zrit = or
o Classical SPC, based on the assessment of
C4 again depends on the size of the sample the whole process by means of samples, can
and can be found in the Ford QlOl factor not be used sensibly for process parameters.
table. In the injection molding process the contin-
Before process capability can be calcu- uous recording of actual values, which is al-
lated, the number of samples, size of samples, ready available in the machine control sys-
and distance of samples must be defined. As tem, proves to be more suitable. In this way,
with machine capability, the upper and lower all process data can be used for evaluation.
specification limits must be given. This approach may be called continuous pro-
The minimum requirement for process ca- cess control (CPC). The control charts ob-
pability is that straggling must be xq f 3 x s tained in this way are termed process param-
within the specifications, that is, 99.73% of eter control charts.
all parts produced are expected to lie within These control charts can be helpful in eval-
the tolerance limits. The index Cpk must be at uating the process. Unusual curves or trends
least 1.0. in the process parameter control chart-even
Only those parameters that are subject to within the control limits-can provide the
accidental influences and do not depend on first indication of an unfavorable process de-
systemic influences are suitable for statistical velopment, which should be corrected even
process control. This means that all closed- before points outside the control limits occur
or open-loop controlled parameters are un- (Fig. 13-27).
suitable for SPC (e.g., a wrong controller set- So-called runs indicate that the process has
ting could be a systemic error). Apart from shifted, for example, when seven subsequent
1152 13 Statistical Process Control and Quality Control

long sequence above the long sequence below the

average average

I - \/ - \

---- I v
LcL i
Fig. 13-27 Trends in the process parameter control chart.

points lie on the same side of the mean operation and gives no objective measure for
value or seven intervals rise or fall in a row further assessment of the process.
(Fig. 13-28).
Of course, it is indispensable to prove that
evaluation based on the process parameters Practical Example
can actually be considered relevant quality
information. Assume the property relevant for quality
Determining machine or process capability is weight. Weight is called x. The sample size
via process parameters is not very sensible.To is set to be 50, also with a view to machine ca-
calculate machine and process capability, the pability (Fig. 13-29). Because of a maximum
upper and lower specificationlimits of a prop- weight
erty must be set. Setting the tolerances for the
xmax= 415.5 g
actual values of an injection molding machine
requires a smooth touch and exact knowl- that was found, although only once, the range
edge of the process. The tolerance range for is
process parameters is not set by the end cus-
R = 415.5 - 413 = 2.5 g
tomer, but by the machine operator. Thus, a
statement about the machine and process ca- The cause for this "outlier" is a short cy-

"cL
pability is of no value. It can only be used cle interruption. The resulting standard devi-
for monitoring one's own injection molding ation as a measure of straggling around the

long slope of the curve

LCL

Fig. 13-28 Runs in the process parameter control chart.

 13 Statistical Process Control and Quality Control 1153

R = 2,s g s refer to xq: 0,11 %

418

2 414 .-:. ..' _....*... . ..... .. - .. . . . . ' xq

M

*
1
. :. . ...
412

410 1 I 1 1 1 k 1 1 1

mean value of and

xq = 413.66 g 416 - 413.66
Zcrit = = 5.06
0.462
is
s = 0.462 g
This means that the process is under statis-
That is, on the average, the measured values
tical control (c,k > 1.33).
deviate by this value from the mean value.
If range R is related to the mean value,
the resulting deviation is 0.60%. If, however,
standard deviation is related to the mean Process Capability
value, the deviation is 0.11%. If the outlier
is disregarded, these values are as follows: The same molding of the same weight is
considered here. To determine process capa-
Range bility based on classical SPC, five moldings in
R = 1.6 g 2 0.38% a row are weighed after each 20 shots, that is,
the sample size is
Standard deviation
s = 0.382 0.09% n=5

After 300 shots, for example, process ca-

pability is calculated as follows (Fig. 13-30).
Machine Capability Based on the samples taken, the following
values result:
According to specifications, the minimum
weight must be 410 g and the maximum xqq = 413.55 g
weight 416 g, that is, s q = 0.282 g
USL = 416 g and from the QlOl factor table for n = 5,
LSL = 410 g
C4 = 0.94
with
The resulting d is
s = 0.462
6 ~ 0.282
c, = = 2.16 c=- = 0.3
6 x 0.462 0.94
1154 13 Statistical Process Control and Quality Control

4143
cp = 2.2
[g] cpk= 2.7

5 414 --I .OEG

W
m

llxqq

a WEG
I 1 I 1 I I
413 ' I

With USL = 416 g and LSL = 410 g, A Successful SPC System

6 Statistical process control must be used by

c -- = 3.33
'-
6 x 0.3 management to decrease process variability.
That process can be producing a molding,
and with
writing a purchase order, or filling out an
416 - 413.55 expense report. A successful SPC system re-
= 8.2
Zcrit = quires:
8.2
Cpk = - = 2.7 Top management commitment. Not just in-
3 terest, sympathy or support, but active in-
Thus, the minimum requirement of cpk= 1is volvement and commitment to quality.
fulfilled. Total workforce participation. Full involve-
ment of all employees, perhaps through
quality circles and profit improvement
Learns; but most important, through daily
Control Limits for the Process Control Chart
hands-on activity in the quality arena.
Continuous education. Both in terms of
For N = 5 , the factors have the following
concepts (variability, control, etc.) and
values:
functional (full knowledge of job require-
A3 = 1.427 ments and techniques).
Use of statistical methods. Not only control
B3 = 0
charts, but more advanced statistical meth-
B4 = 2.089 ods such as experimental design, multivari-
ate analysis, and nonparametric methods.
The resulting control limits for the mean
Close vendor and customer relationships.
value are thus
Clear understanding of all requirements
+
UCL,, = 413.55 1.427 x 0.282 = 414 g and utilization of the Deming helix (de-
sign, produce and test internally, field test,
LCL,, = 413.55 - 1.427 x 0.282 = 413.1 g redesign) in the manufacturing and mar-
and for the standard deviation keting chain.
The key to this way of life are the beliefs
UCL, = 2.089 x 0.282 = 0.58 that clear requirements must be established
LCL, = 0 and enforced: Errors must be prevented and,
 13 Statistical Process Control and Quality Control 1155

when errors occur, corrective action must be tion, and final product. About 80% of the
taken to prevent the error from recurring; testing effortis usually spent on the control of
the goal must be to achieve 100% confor- process parameters that impact final product
mance to requirements while striving to elim- properties.
inate all defects; and areas of cost created Control charts are usually developed dur-
by nonconformance must be identified and ing the standardization phase of product de-
corrected. velopment. During this phase, 60 to 90 days of
In this move toward total quality control, product are analyzed to set tentative 3-sigma
you are attacking three primary areas: the de- control limits for the average and standard
sign and development of process and prod- deviation of process parameters and prod-
ucts that will meet customer requirements, uct properties. By this stage, the process has
raw materials control to assure suitability for demonstrated some degree of stability and
use in your manufacturing operations, and sampling frequency, and test procedures have
statistical process control to prevent the pro- been established. These limits are then used
duction of defects. to subsequently track the process and are re-
All three of the above basics must be done viewed periodically to determine if a reduc-
well if customer expectations are to be met. tion in variation has dictated a change in con-
An important subset of the quality system is trol limits.
the intensive use of SPC techniques. Statis- After much thought and a few false starts,
tical process control is a means of achieving you decided that a process will be judged
process consistence and conformance to es- out of control if a data point lies outside
tablished quality standards through the use the control limits or, as an example, more
of statistical methods in all stages of a manu- than six consecutive data points are on the
facturing operation. The goal you are striving same side of the centerline. Either of these
for is the continuous reduction of variability conditions indicates that an assignable cause
in the final product. must be found and corrective action be
The workhorse of our SPC system is the taken.
control chart for average and standard de- Statistical process control (SPC) is used
viation, but control charts of the process throughout the plastics-processing industry.
are only a start and you must move from con- The key components for injection molding
trol of the process via process testing to con- process control have been identified, as well
trol of the process via process understanding. as the general techniques. This was accom-
plished through a series of five papers deliv-
ered at the “Statistical Process Control” por-
Production Controls tion of the 1985 SPE ANTEC in Washington,
D.C. These papers described how several
Recognize that production controls the companies went through various lengthy, for-
process. Quality control exists to advise and mal (costly!) investigations. The significant
serve production in the areas of assuring outcome was that they all came to basi-
that raw materials are suitable for use, ap- cally the same conclusions. This review of
plying proper statistical and computer anal- the Washington presentations has two basic
ysis, and providing certain data that require, objectives:
for example, X-ray or IR analysis. Any sys-
tem in which quality control alone tests and 1. To derive a skeletal SPC program that
adjusts the process is a barrier to the sense can ultimately be developed into a general-
of ownership that production workers must ized uniform approach to SPC for the injec-
feel toward their job for that job to be done tion molding process.
well. 2. To stimulate more publications in this
You test and maintain control charts at all area and thus advance the state of the art.
strategic points in the production chain: raw This topic is woefully lacking in examples of
materials batch mixing and melting, fabrica- specific applications.
1156 13 Statistical Process Control and Quality Control

Directions The first result of the Wash- SPC Step Three: Injection Molding
ington conference was the identification and
definition of the three basic process areas di- Weight Part (or shot) weight was found
rectly affecting the injection molding process: to be the only practical parameter that can be
used for SPC. The equipment required is min-
1. Raw materials imal, relatively inexpensive, and quick and
2. Internal materials handling (drying, easy for the operator to use. Although weight
blending, etc.) does not characterize a molded part, it is the
3. Injection molding preferred “control” measurement. Weight
was selected for this purpose by several mold-
The most significant characteristic of these ers working independently (Fig. 13-31).
basic process areas is the sequential depen-
dence of the three steps. SPC for injection Operator The operator (the person who
molding (step 3 ) is impossible if either step 1 physically turns the knobs on the machine)
or 2 is not under SPC. Each of these areas, must be an integral part of the SPC proce-
as well as the nature of the implementation dure. SPC must operate at a real-time rate;
required, will be discussed in detail. otherwise, SQC must be done when too much
scrap has already been produced. The only
way control can be accomplished with the
SPC Step One: Raw Material timeliness appropriate for SPC is to build the
procedure around the operator.
Single test measurement A simple, rapid
single point test was needed; the melt in- Primary problem After the material
dex was found to meet this requirement. Al- problems are resolved (generally with diffi-
though the melt index does not completely culty), the process is then capable of being
characterize a material, for the purposes of analyzed for “assignable cause” effects. This
SPC, it does not need to! usually starts with mold- or machine-related
causes. Once these are resolved, the primary
Time-dependent sampling The simplest operating problem that emerges is operator
method was to have the supplier sequen- overcontrol. The usual range of changes that
tially number each box in the run. A sample an operator makes to “improve” the process
from each box was then measured and the re- will push the process out of statistical control.
sults plotted using the standard control chart This is not his or her fault. It is a result of his
format. or her inability to quantitatively determine
the results of an action in a timely manner.
By putting the operator in the “process loop,”
SPC Step Two: Materials Handling that is, having that person weigh the parts and
plot the weight, the operator receives instant
Drying The importance of this element feedback as to the effect, direction, and mag-
varies considerably with the nature of the nitude of his or her effort. Adjustment strate-
resin. Improper dryer control or procedures gies and their amplitude soon become more
will make SPC impossible (see Chap. 6). appropriate to the process (see Chap. 7).

Blending A “minor” blend ratio shift (vir-

girdadditive or concentrate) is easily dis- SPC Implementation: Summary
cernible on the weight-response control of Experience
charts. “Regrind” should not be combined
with virgin materials except under rigorously Management support Complete support
measured (controlled) conditions since the and unwavering commitment from manage-
process mean (weight) is generally signifi- ment are an absolute requirement. (Through-
cantly different between virgin and regrind out this chapter and also in others, we have
material (see Chap. 10). to repeat management support.)
 13 Statistical Process Control and Quality Control 1157

Part/Shot
Weight

I1-4 \
Control response - a
response that is influenced
by the process Variabi/ity
Definitions:
Response - an
observed parameter
change
Single response - a
response that is influenced
by the process Means

Characterize current Experimentation Designed

process (exploratory) Experiment

Analysis
identify assignable causes and mine correlation coefficients
allow them to be corrected. or a Taguchi Signal Factor
Numerical comparisons accom-
plished by Taguchi TYPE N
SignaVNoise Ratio as the re-
sponse variable

Fig. 13-31 Use of weight as a response for injection molding quality products.

Resources SPC initially requires signifi- simply not available. Generally, only the most
cant resources. Dedicated time, personnel, important products can be considered. New
and money are needed to establish the pro- products being prepared for production are
cedures and develop the required database. often excellent candidates for SPC since the
This represents a team effort using person- resources typically needed for SPC are al-
nel from up to a dozen different disciplines. ready committed and only an adjustment in
A pilot program for a fairly straightforward internal procedures will be required.
product is needed to develop the organiza-
tional structure and operational procedures Discussion Let us assume the following
appropriate for a specific manufacturing working definition of SPC: “Statistical pro-
unit. cess control seeks to more closely control the
manufacturing process and permits the man-
Patience SPC requires lots of time! A ufacture of tighter tolerance parts by indicat-
properly developed plan is neither simple nor ing when the manufacturing process is start-
obvious. A historical database must be de- ing to drift away from the ideal set point.”
veloped and evaluated. This activity cannot It becomes obvious that the key concept for
be done quickly. Often, the data-gathering SPC is timeliness. The procedure selected for
process is slowed by the discovery of sig- SPC must operate on a time scale appropriate
nificant “assignable causes” whose solutions for the process.
take away resources from the original goal. There are, in fact, only two possible ap-
proaches for real-time process control: either
Priority SPC may not be practical for ev- through rapid dimensional measurement of a
ery product because of the high cost in time specific product characteristic or via measur-
and personnel needed to prepare a product ing a “bulk” (nondimensional) characteris-
or process for SPC; sufficient resources are tic. Weight is such a “bulk” parameter. The
1158 13 Statistical Process Control and Quality Control

identification of weight as the most practical SPC:

parameter for injection molding SPC was one Immediate. Measurement can be made as
of the most important elements that emerged fast as the part can be put on the scale.
from these studies. Robust. Results are insensitive to proce-
In contrast to weight, dimensional mea- dures: Algorithms built into modern scales
surements of the precision needed for SPC compensate for procedures and the envi-
are generally done offline. This results in a ronment.
response that is simply too slow for effective
Low cost. Scalesfor this purpose are cheap!
SPC. Using dimensional measurements for
0 Simple analysis. Data can go directly into
SPC has another inherent time-related prob- operator-station statistical display or even
lem. Typically, amorphous materials require simply be manually plotted.
30 to 60 min to cool before an “approximate”
dimensional measurement can be made. Figure 13-33visually summarizes the basic
Figure 13-32shows an example of this char- thrust or direction developed by several pro-
acteristic. The figure shows the postmolding cessors to achieve SPC with their processes.
shrinkage of a typical, small part molded from
HIPS. It is obvious from this figure that a fair Weight The variability of weight, how-
amount of time must elapse before the part ever, has a multifaceted nature. Confusion
stabilizes sufficiently for precise dimensional between these various natures of weight re-
analysis. This time frame is unacceptable for sponse can lead to the rejection of weight as
SPC. This example is a thin-walled part; if it a reliable measure of quality. What is starting
were thicker or molded of a crystalline mate- to emerge are the following uses for weight
rial, a much longer stabilization time would data:
be expected.
An additional problem arises with the use 1. SPC. Statistical process control can now
of dimensional measurements; they lack suf- be readily achieved in a rapid and low-cost
ficient resolution [typically f O . O O 1 in. (0.003 manner. The procedures are readily avail-
cm)] for “control” purposes. Sufficient expe- able, widely accepted, and quite well under-
rience has been obtained to show that weight stood.
data have an inherent resolution that is at 2. Process analysis. The superb resolution
least an order of magnitude better than stan- of process variability exposed by the use of
dard dimensional measurements. weight provides a powerful tool to expose
One can easily summarize the impor- “assignable causes” and other process char-
tant characteristics of the use of weight for acteristics.

15185 -;30 sec

15180

1 5175

,,5170
t{b;9 rnm

Curue Equation I
e IC = 1 .SI9 - .005*eA(-1/t)!

h
1.51 5 5

15150

._.-_ .
0 10 20 30 40 50 60 70 80 90
Time(Min1
Fig. 13-32 Postmolding shrinkage time example for a typical small high-impact polystyrene part.
 13 Statistical Process Control and Quality Control 1159

This componcnt is primarily

under control of the materials
manufacturcr
~ ~~

Matcrial drying
Material blending
Use of rcground

r----l
Injection molding

*
Dimension Weight
I
I
Well-known and Very high sensitivity -
cstablishcd proccdurcs will pick up process
Measuring equipment Sclcct rcsponsc
changes before dimen-
usually on hand sional techniques

Slow - uwally donc o f f line

Not real time
Rcsolution finer than O.OOI in
I
or
Fast - real time
rcsponse possible
Measurements and

i-
N o t rcadily achicvcd analysis available
immediately at the machine
Little training required by
operator in use of balance
Equipment cost is modest

Not suitable for SPC --

too slow
Best potential for SPC -
close to real time I
Fig. 13-33 Synopsis.

3. Product “aimpoint” control, Some char- proach. A uniform approach that could be
acteristic of the molded product is expected accepted industrywide is needed. Published
to correlate highly with the absolute value of examples of successful applications are also
the part weight. Not all products will have a needed.
useful correlation; special experimental pro-
cedures are required to develop these corre-
lations. How to Succeed with SPC
Although these characteristics of weight
You should never stop improving, never
as a process response for injection-molded
stop tightening the control limits. To succeed
parts provide the engineer with a rich source with SP/QC, you should observe these key
of tools for process control (see Chap. 7),
points, based on what theorists and practi-
troubleshooting (see Chap. ll),and product tioners say:
control, additional work is needed to formal-
ize the methodology applied in each of these Have patience. Results will not come
cases. Although molders have been using var- overnight. It can take years in some cases.
ious aspects of weight for product control for And you are never at a point where you
many years, this usage lacks a uniform ap- can say, “We have improved enough.”
1160 13 Statistical Process Control and Quality Control

Have confidence in your people. Your of ongoing production will become more and
workforce is the biggest asset you have, more important. In many cases, pure pro-
provided your workers are properly cess data recording, which can be standard
trained and motivated. SP/QC provides today on all injection molding machines, will
you with a common language that cuts not be sufficient as proof of quality. In addi-
across all levels of management and thus tion, statistical measures for evaluation will
enhances the quality of communications. be required. The use of classical SPC on in-
Your machine operators can do much for jection molding should be viewed critically.
your company. So ask the people in your Whether the use of continuous process con-
plant what they think needs to be done, trol (CPC) will be accepted for the evaluation
and you will find out just how much they of product properties depends on whether a
know. You will be very surprised. connection between process parameters and
Train properly and continuously. The first quality-related product properties is found.
step toward the proper implementation This would be a first step toward real closed-
of SP/QC is the thorough training of ev- loop process control.
erybody involved in manufacturing. And
training must be ongoing, sharing new data
and new ideas. Terminology
Break down walls. SP/QC users say that
eliminating layers of management helps Analyze data Proper statistical gathering
enhance results. Having designers talk di- must be followed with analysis of the process-
rectly to manufacturing engineers is one ing information and properly drawn conclu-
typical recommendation, and talking di- sions. The key terms that apply are common
rectly to machine operators is another. and special causes, local actions and actions
on the systems, process control and process
In summary,the key to product consistency capability,and control charts. Teachers of sta-
process consistency. It is mostly manage- tistical methods stress again and again that
ment that must act to assure process stability quality problems are the result of variation
and a companywide commitment to quality. such as in materials and equipment process-
The benefits include: ing conditions. Statistical analysis will help
People. Processors who complete the all- explain the causes of these variations.
important training of plant personnel find
that worker motivation increases, turnover Statistical assessment A fabricator’s as-
is reduced, and productivity grows. sessment of its use of SPC tools should con-
Scrap. Drops in scrap rates from 50 to 80% sist of a formal, documented examination of
are the most frequently mentioned current statistical practices and procedures
changes. These are savings that go straight as well as an evaluation of future plans for
to the bottom line. improvements of the company’s QC. To be
Productivity. A molder cutting scrap rate useful, this assessment should go beyond the
on his twenty machines by just 10% has, in compliance-oriented approach that is com-
effect, increased available machine capac- monly seen in quality audits. Fabricators
ity by two machines and at no cost. Finding should set up key objectives for their assess-
and realizing this “hidden capacity” is one ment of good statistical practices. They could
of the most immediate payoffs of SP/QC. include: (1) determine your current state of
compliance regulations; (2) determine im-
pediments to compliance; (3) raise awareness
Outlook of regulations; (4) measure improvements
over time; ( 5 ) discover the best statistical
This review has highlighted various aspects practices in use throughout the company and
of quality in injection molding. Apart from share them with the rest of the company; and
quality-assurance measures around the injec- (6) provide advice on incorporating statistical
tion molding process, the quality monitoring tools into the quality improvement system.
 13 Statistical Process Control and Quality Control 1161

Statistical benefits Using statistical meth- Statistical F-tests A standard statistical

ods in the design of experiments and data test, applied to the ratio of two estimates of
analysis allows designers, compound formu- variance to determine whether there is a sta-
lators, processors, etc. to attain benefits that tistically significant difference between the
would otherwise be considered unachievable. variances of the distributions from which the
Benefits include a 20 to 70% reduction in estimates are made.
problem-solving time; a minimum 50% re-
duction in costs due to testing, machine pro- Statistical mechanics Describes systems
cessing time, labor, and materials; and a that have many degrees of freedom and
200 to 300% increase in value, quality, and a wide range of possible states. An exact
reliability of the information generated. classical or quantum mechanical description
of the full system is usually impossible,
Statistical data collections Data may be but a great deal can be understood about
collected directly by observation or indirectly the average properties of these systems by
through written and/or verbal questions. The using the concepts and methods of statistical
latter technique is used extensively by market mechanics. Plastic systems are naturally
research personnel and public opinion poll- adapted to be studied by statistical mechan-
sters. Data that are collected for quality con- ics and may be used for readily illustrating
trol purposes are obtained by direct observa- its general principles.
tion and are classified as either variables or
attributes. Variables are those quality charac- Statistical median The middle value in
teristics that are measurable, such as a weight an array arrangement in sequence. Thus, 1,
measurement in grams. Attributes are those 5,9,13, and 17 results in a median of 9.
quality characteristics that are classified as ei-
ther conforming or not conforming to speci- Statistical methods Concerned with
fications. In other words, attributes are either deriving information from a given set of
good or bad, while variables indicate degree data (analysis) to meet product performance
of goodness or badness. requirements and to solve problems. Statis-
tical methods minimize the amount of data
needed to derive specific information.
Statistical effects The response of the
process to a change in factor level from low Statistical modes A frequent value or
to high. could be several in a set such as bimodel,
trimodel, etc.
Statistical equivalent loading system The
St. Venant’s principle states that the stress Statistical normal curves Although there
and deflection of a part (handle, gear, etc.) are as many different statistical universes
at points sufficiently distant from points of as there are conditions, distributions can
load application may be determined on the usually be described by as normal or Gaus-
basis of a statistical loading system. sian. The normal curve or normal universal
distribution is a symmetrical, unimodal,
bell-shaped distribution with the mean,
Statistical estimations A procedure for median, and mode having the same value. A
making a statistical inference about the nu- universal curve or distribution is developed
merical values of one or more unknown pop- from a frequency histogram. Much of the
ulation parameters from the observed values variation in industry and in nature follows the
in a sample. frequency distribution of the normal curve.
The normal curve is such a good description
Statistical factors A process or recipe of the variations that occur to most quality
variable that can be controlled indepen- characteristics in industry that it is the basis
dently, such as temperature and the ratio of for many quality control techniques. The
filler to plastic. area under the bell-shaped curve is equal to
1162 13 Statistical Process Control and Quality Control

1.00 (when using the formula for the normal Statistical quality control (SQC) Mea-
curve) or 100% and therefore can be easily sures product quality and provides a track-
used for probability calculations. ing mechanism to reveal any shifts in level of
quality. It is a derivative practice based on the
Statistical phases (reasoning) The des- results of SPC. Conceptually, SQC can reject
criptive or deductive statistics technique parts that do not conform to the approved
used to describe and analyze a subject group. standard sample. In practice, parts are phys-
Inductive statistics endeavors to determine ically rejected and diverted into reclamation
from a limited amount of data (sample) an or recycling systems. Alarms are provided at
important conclusion about a much larger the machine and at the central computer to
amount of data (universe). Since these inform people (workers, management, etc.)
conclusions or inferences cannot be stated of the rejects. SQC is a scientific method of
with absolute certainty, the language of analyzing data and using the analysis to solve
probability is often used. practical problems.

Statistical populations Infinite in size and Statistical randomization A method of

containing all those things of interest that sequencing experiments by using a random
have one thing in common. number table so that each experiment in the
proposed plan has an equal chance of being
Statisticalpopulation parameters A fixed the first, second, or last experiment. This is an
value characterizing a certain aspect of a sta- extremely effective technique for minimizing
tistical population. An estimate of the value errors.
of the population parameter, derived from
a sample, is called a “sample estimate” or
Statistical ranges Measured spread of
“statistic.”
data around the control value; the difference
between the highest and lowest values of the
Statistical precision The range limits
variables in the sample.
within which the estimated or obtained
value lies. For example, tensile strength of
25,000psi (172 MPa) and f2,000 psi (14 MPa) Statistical R charts Charts revealing sta-
has a 50% probability. tistical variation over time.

Statistical probability, 6-sigma Based on Statistical regression methods Statistical

the statistical normal curve evaluation, 99.9% procedures dealing with the study of the asso-
of all manufactured products will fall within ciation or relationship between two or more
a 6-sigma specification. variables.