QUALITY PROGRESS

I
JUNE 2006
I
25
Use SPC for
Everyday Work
Processes
by Greg Gruska and Chad Kymal
tatistical quality control (SQC), also
known as control charting, started with
Walter Shewharts work at the Western
Electric plant outside Chicago in the 1920s.
Since then, SQC has been reintroduced into
industry every couple of decades or so and has
evolved into statistical process control (SPC) to
reflect the move away from product control to a
systems focus.
But why must SPC be periodically revitalized?
If it is all people say it was and is, shouldnt it be
self-sustaining? Partly the problem is that if times
are good, management focuses not on economic
control but on volume control. So we see many
organizations embracing SPC only during times of
trouble. When times are good, the attitude is We
dont have time for such luxuries.
Even organizations that implement SPC as part of
their continual improvement efforts fail to sustain
its use, sometimes because the results of applying
SPC to processes have a variation model different
from the one shown in most books. It is a case of
using the right toolbox but the wrong tool.
To help organizations use SPC tools the right
way, the Automotive Industry Action Groups
supplier requirements task force, representing
General Motors Corp., Ford Motor Co. and
DaimlerChrysler, recently released a second edi-
tion of its SPC Manual.
1
The entire first chapter
explains the philosophy and use of SPC, main-
taining it should not be applied to processes but
integrated into an organizations continual
improvement activities.
S
In 50 Words
Or Less
Theoretically, statistical process control (SPC)
is viewed as useful for economically producing
consistently acceptable products and services.
In practice, however, SPC isnt being used for
continually improving unique industrial
processes.
The right tools, such as advanced charts, can
make SPC effective in these situations.
STATISTICS
Use SPC for
Everyday Work
Processes
The edition discusses a useful three-stage
improvement cycle for integration (see Figure 1):
1. Analyze the process.
2. Maintain (control) it.
3. Improve it.
Deployment Shortcomings
Despite the advantages of SPC, why have many
organizational implementation efforts not been
successful or self-sustaining?
Many of the contributing causes have nothing to
do with the underlying methodology but with the
organization and deployment. Some examples fol-
low.
Constant change. SPC assumes the process con-
trols maintain the common cause variation system.
All too often this is not possible because there are
ongoing changes to the process resulting from:
Special causes of variation.
Physical changes to the processwith the
intent of improving it.
Administrative changes to the control activi-
ties for logisticalor whimsicalreasons.
Changes in management direction regarding
what is desired or needed.
Change is a necessary element of continual
improvement, but it must be within a plan-do-
study-act cycle, not haphazardly applied without
an understanding of its impact.
Right idea/wrong tool, or not understanding
the physics of the product and process. Applying
SPC without understanding the physics of the
product or process and the dominant sources of
variation will lead to frustration among both oper-
ators and management. Much of this happens
because most people have been exposed to only
basic SPC control charts.
Although the four basic variable charts and four
basic attribute charts are applicable to a wide vari-
ety of processes, advanced charts are better suited
to many processes. The term advanced does not
necessarily imply the use of more sophisticated sta-
tistics. Often, these charts are a modification of
basic charts for specific conditions of the process to
optimize the detection of special causes.
2
If they use the wrong charts, the operators will
not see any benefit from the extra work necessary
for the SPC implementation, and management will
still see inconsistencies in the process output.
Limited understanding. SPCs application is
often limited to processes similar to the examples
provided in an SPC class. But SPC can be useful in
a wide variety of sectors
outside manufacturing.
Within healthcare, many
organizations such as the
Joint Commission on Ac-
creditation of Healthcare
Organizations and the
Institute for Healthcare
Improvement have recog-
nized the need to under-
stand common and special
causes of variation and the
use of SPC in process analy-
sis. This goes beyond ad-
ministrative processes to
also include clinical pro-
cesses and improvement
actions.
3
Lack of patience. Even
when the SPC deployment
is the right idea using the
right tool, management and
workers seem to expect
26
I
JUNE 2006
I
www.asq.org
STATISTICS
SPC Improvement Cycle FIGURE 1
1. Analyze the process
Determine what the process
should be doing.
Determine what can go wrong.
Determine what the process is doing.
Achieve a state of statistical control.
Determine capability.
2. Maintain the process
Monitor process performance.
Detect special cause
variation and act on it.
3. Improve the process
Change the process to better
understand common cause variation.
Reduce common cause variation.
Plan
Do
Study Act
1
Plan
Do
Study Act
2
Plan Do
Study Act
3
QUALITY PROGRESS
I
JUNE 2006
I
27
instant gratification. When they dont immediately
observe consistency and improvement, manage-
ment may withdraw support or workers may not
follow through.
Assumptions
W. Edwards Deming taught us the role of man-
agement is to make predictions.
4
The purpose of
SPC activities is to enable management to predict
the future state of a process by identifying and
ameliorating special causes of variation. For SPC to
be implemented effectively, some assumptions
hold:
Variation and interdependencies exist in all
things.
Few systems and processes are constantly sta-
ble.
When applying the basic control charts, real-
ize the process being analyzed, monitored or
controlled must be a purely random (or white
noise stochastic) process.
A random process satisfies:
A common deviation from the standard assump-
tions lies in processes with outputs correlated with
each other. Some include stamping, machining that
is tool wear dominant, chemical processing, the
stock market and an individuals medical readings
(for example, temperature, blood sugar level and
blood pressure).
These processes are called stationary processes
and satisfy:
5
This is also called an autoregressive process or a
process with autocorrelated data.
The Shewhart chart control limits and the stan-
dard calculations for capability indexes depend on
the assumption of a white noise process. Stationary
E [
t
] = x
t
=
x
for all t.
t
k t 1
with the correlation between x
and x equals .
t
x

for all t. Var ( ) =
2
t
=
2
E [
t
] =
t
=

for all t.
t

t k
is uncorrelated with for all k.

for all t. V (t) =

2
t
=
2
ar
is not enough. It sometimes can be difficult to dis-
tinguish between a white noise and a merely sta-
tionary processin fact, the white noise process is
often called a weakly stationary process.
Several charts can monitor and control a station-
ary process:
Autoregressive charts. These include the autore-
gressive, and autoregressive and moving average
models. This approach seeks to model the underly-
ing relationships among the process output values
and use this knowledge to better identify other
special causes of variation.
Cumulative sum (CUSUM) and exponentially
weighted moving average charts (EWMA).
Although the CUSUM and EWMA were devel-
oped to detect small shifts in the mean in random
processes, they are robust enough to handle
processes with minor autocorrelation.
Individuals (I) and moving range (MR) charts.
If the within subgroup variation is less than or
equal to the discrimination of the measurement
system appropriate for the process, an I and MR
chart may be a suitable way to control the process
variation. However, very strong autocorrelation
may still display itself in a nonrandom pattern.
Structured samples. If the source of the autocor-
relation is a consistent and predictable special
cause, the selection of the sampling quantity and
frequency should reflect this dominant source of
variation. For example, if the process is material
Upper specification limit
Lower specification limit
Time
Rapidly Drifting Process Center FIGURE 2
28
I
JUNE 2006
I
www.asq.org
dominant, the sampling should occur whenever
the material changesfor example, with the
change of coils.
Structured charting. If the source (special cause)
of the autocorrelation is predictable, it is possible
to control the process by segregating the within
subgroup variation from the between subgroup
variation on separate charts. The between-within
chart uses an I and MR chart approach and the typ-
ical range chart:
The I chart plots the subgroup averages as
individuals against the control limits based on
the moving ranges.
The MR chart plots the between subgroup
variation using the moving ranges based on
the subgroup averages.
The rangeor standard deviationchart
plots the within subgroup variation (common
cause variation).
How Common Are Common Causes?
Shewhart charts require the center of the process
to remain constant over time and the variation due
to only common causes for process behavior to be
predictable.
However, one type of special cause is not part of
the common cause variation but, within bounds, is
predictable. This type of cause is often called an
economically allowable special cause (EC), because
only minor cost benefit results from its elimination.
Another name for this type is environmental cause.
If the EC has a consistent and predictable behav-
ior, the behavior of the total process variation can
also be predicted within bounds. If the EC is incon-
sistent or exhibits chaotic behavior, the control
methods in the example that follows will not be
effective.
A classic example of a process with an EC is a
screw machine. In this process, the tool wear is
rapid. Figure 2 (p. 27) is an example of a rapidly
drifting process center, often evident within the
span of a single shift. Because of this EC, the process
can have additional variation caused by setup varia-
tion.
Organizations often use conventional control
limits, with the process center forced to be at the
midpoint of the specification limits. This can result
in overcontrol and decreased productivity.
A screw machine is used to manufacture small
shafts. The shafts are produced continuously on
the same machine by two shifts per day, six days
per week. The measurement under study is the flat
width of the shaft spindle.
The process center m increases rapidly as the
machine tools wear. The process may run for an
entire eight-hour shift before tool maintenance is
required. The process standard deviation, !X,
remains fairly constant during the course of a tool
wear cycle.
Once the effect of tool wear is removed, there is
strong evidence the product measurements are
being generated by a white noise process.
The objectives of the control plan are:
Keep tool change time to a minimum.
Minimize operator overcontrol.
Maximize the length of the tool wear cycle.
Ensure the process remains in control once the
tool wear variation has been taken into
account.
STATISTICS
68
64
60
56
52
0 10 20 30 40 50 60 70 80 90 100
A
v
e
r
a
g
e

f
o
r

s
a
m
p
l
e
Sample number
Sample Averages FIGURE 3
Allow the operator to stay as
close as possible to traditional
Shewhart procedures.
QUALITY PROGRESS
I
JUNE 2006
I
29
Minimize the probability of producing a non-
conforming part.
The first step in trying to meet these objectives is
to realize once again the fundamental equation of
process control must be revised in this situation to
read:
Total variation = common cause variation + tool
wear variation + other special cause variation.
Isolate and measure tool wear variation sepa-
rately from other sources of variation. Understand
the trade-off between long tool wear cycles and the
need to deal with fussy tool change problems.
Allow the operator to stay as close as possible to
traditional Shewhart procedures.
How To Do It Right
The first step in dealing with tool
wear trends is to collect the right kind
of data. For this process the approach
selected is:
Draw samples of size three once
every 675 pieces, which keeps the
sampling frequency close to its
original value of a sample once
every hour.
Use a pan holding approximately
675 pieces to determine when a
sample needs to be drawn.
The values from the first 101 samples are
shown in Figure 3. Each dot represents the average
for a sample of size three.
It is difficult to see tool wear cycles in these data.
Examination suggests tool wear trends in the first
part of the data. These trends can be highlighted by
removing the lines connecting the last point in one
cycle with the first point in the next cycle.
Extensively annotated control charts provided
by the operators helped identify individual cycles.
It was determined most of the later data were not
collected under the proper conditions and could
not be used for estimating tool wear.
Tool wear patterns lurking in the data emerge
once extraneous lines and dots are removed (see
Figure 4).
The first step is to determine whether the tool
(the special cause) is consistent over time (tool-to-
tool variability is predictable). To find a common
60
62
58
56
54
52
0 1 2 3 4 5 6 7
A
v
e
r
a
g
e

v
a
l
u
e

o
f

s
u
b
g
r
o
u
p

r
e
a
d
i
n
g
Sample number (675 pieces per pan)
Slope = 0.7962
Superimposed Patterns With
Regression Line and X-bar Chart
FIGURE 5
S
u
b
g
r
o
u
p

a
v
e
r
a
g
e
Subgroup or time
Random Sample FIGURE 6
68
64
60
56
52
0 10 20 30 40 50 60 70 80 90 100
A
v
e
r
a
g
e

v
a
l
u
e

o
f

s
u
b
g
r
o
u
p

r
e
a
d
i
n
g
Sample number
Sample Data FIGURE 4
30
I
JUNE 2006
I
www.asq.org
trend in the various tool wear patterns, all cycles
were superimposed and a single simple linear
regression model was fit to the entire collection of
tool wear trends (see Figure 5, p. 29).
The regression model explained 77% of the
datas variation. The unexplained variation
exhibited the properties of white noise.
The tool wear slope of 0.7962 was estimated
using a simple linear regression model. This regres-
sion model also estimated a standard error of
0.0571 for the slope coefficient. Three standard
error tool wear growth limits can be calculated by
(3 x .0571) applied to the expected aver-
age line.
When someone is manually controlling
processes with a tool, he or she can lay a
transparency of Figure 6 (p. 29) on top of
the control chart.
The control limits in Figure 5 (p. 29) are
the same as those in a conventional X-bar
chart except they follow the average tool
wear line rather than a horizontal center-
line. This approach requires the special
causetool wearto exhibit a predictable
and consistent behavior, which requires
control by the supplier and purchasing.
The second step is to verify and quanti-
fy the tool wear behavior over time by
studying a random sample of tools over
their life (see Figure 6). This may require
100% sampling if the tool life is short.
This analysis must determine whether
the expected tool life pattern and common
cause variation are consistentand thus
predictable. Then this information should
be used to establish process control charts
to control the process variation and tool
wear variation (see Figure 7).
As confidence is gained in the process,
the diagonal control charts can be replaced
by I charts monitoring specific tool life fea-
tures (see Figure 8):
Setup control. Ensures the setupthe
starting point of the process trendof
each tool is consistent. This, with the
change control, also determines the
capability indexes of the process.
Wear control. Ensures there is no
change in the tool life model.
Change control. Ensures the useful
life of the tool has been reached and
the end life model of the tool has not
changed. This, with the setup control,
also determines the capability indexes
of the process.
STATISTICS
S
u
b
g
r
o
u
p

a
v
e
r
a
g
e
Subgroup or time
Control Chart FIGURE 7
Change control
Wear control
Setup control
Features
Controlling Features FIGURE 8
Change control
Expected number of parts
Setup control
Controlling Features
Continual Improvement
FIGURE 9
QUALITY PROGRESS
I
JUNE 2006
I
31
As confidence in the process and
tooling increases, monitoring can use
the tool life or number of parts before a
required change (see Figure 9).
In this stage of the continual improve-
ment process, there is confidence the
tooling wear trend will be consistent and
acceptable due to the actions of the sup-
pliers of the tooling and materials. The
need for redundant inspection can be
eliminated.
As confidence in the process and
tooling further increases, monitoring
can use the tool life and periodic sys-
tems audits (see Figure 10).
Based on process knowledge gained,
there is confidence the tooling wear
trend and life will be consistent and
acceptable due to the actions of the
suppliers of the tooling and materials.
The need for redundant inspection can
be eliminated.
In the final stages, setup control
moves offline, and control uses the
tool life or block tool change (see
Figure 11).
In this stage of the continual improve-
ment process, there is confidence:
The tooling wear trend and life
will be consistent and acceptable
due to the actions of the suppliers
of the tooling and materials.
The variation caused by the machine will be
consistent and acceptable due to preventive
maintenance.
Setup consistency will be controlled by offline
setup activities.
The need for any online inspection can be elimi-
nated. This does not eliminate the need for period-
ic system and product audits to verify the process
controls are still valid or ensure an unknown spe-
cial cause hasnt crept into the process.
At any stage, the controlling cycle must restart if:
The tooling exhibits erratic behavior.
A different vendor begins to supply the tool-
ing or material.
Other Types of Control Charts
The new edition of SPC Manual discusses the
need to understand the underlying model of vari-
ation and physics of the process. To enable the
reader to select the appropriate tool for a specific
process model, the manual includes sections on
the following charts:
6
Probability based charts.
Short run control charts.
Charts for detecting small changes.
Non-normal charts.
Multivariate charts.
Regression control charts.
Residual charts.
Autoregressive charts.
The manual describes and identifies the use of
each and provides references for further study but
provides how-to instructions only for the basic
charts.
Expected number of parts
Setup control
Controlling Features
Further Continual Improvement
FIGURE 10
Expected number of parts
Offline
setup control
Controlling FeaturesContinual
Improvement Final Stage
FIGURE 11
32
I
JUNE 2006
I
www.asq.org
STATISTICS
Warnings
It is important to understand the real meaning
of random. Selecting a random sample requires
specific techniques.
In practice, many people think blind selection
is random selection. In reality, this may be hap-
hazard or convenience sampling. Using haphaz-
ard or convenience sampling when random
sampling is required can lead to biased and
therefore erroneous conclusions.
SPC is useful and necessary for continual
improvement, but many applications do not
realize its full benefit because they lack knowl-
edge of the tools and the processes to be ana-
lyzed, maintained and improved.
Although basic control charts, covered in all
introductory SPC courses, have a wide applica-
tion to random processes, there also are many
stationary processes or ones with predictable
ECs that need advanced charts or the application
of a basic chart in a manner reflecting the actual
process nature.
If the goal is to eliminate the need for charting
by building knowledge of and confidence in the
process, SPC charts need to be used so organiza-
tions can increase their understanding of the
common causes and special causes affecting their
processes. Then they can replace SPC charts with
robust policies and techniques governing process
control.
REFERENCES AND NOTES
1. SPC Manual, second edition, Automotive Industry
Action Group, 2005.
2. The SPC Manual, second edition, includes discussion
and references to many of these advanced charts.
3. Improving Heart Failure Care Through
Education, www.ihi.org/ihi/topics/improvement/
improvementmethods/improvementstories/
improvingheartfailurecarethrougheducation.htm.
4. W.E. Deming, The New Economics: For Industry,
Government, Education, second edition, MIT Press, 2000.
5. There are several classifications of stochastic
stationary processes. This example is just one type.
6. The manual does not maintain these are all the
possible charts that can be used. The ones discussed do
cover the majority of situations.
GREG GRUSKA, a Fellow of ASQ, is the vice president of
product development for Omnex Systems and a principal
consultant in performance excellence and a Six Sigma Master
Black Belt for Omnex, Ann Arbor, MI. He directed the devel-
opment and initial implementations of Comprehensive
Process Control Planning, a book published by Omniface
Corp. Gruska is a writing member of the measurement sys-
tems analysis, SPC and failure mode and effects analysis
manual subcommittees of the Automotive Industry Action
Group supplier quality requirements task force, which is part
of the international task force governing ISO/TS 16949. With
masters degrees in mathematics and engineering from Michi-
gan State University and Wayne State University, Gruska is
also an ASQ certified quality engineer. He has been a mem-
ber of the board of examiners and a judge for the Michigan
Quality Leadership Award since 1994.
CHAD KYMAL, an international trainer and consultant, is
chief technical officer and founder of Omnex Inc. He wrote
the ISO/TS 16949:2002 Implementation Guide and the
Auditor Handbook to ISO/TS 16949:2002A Guide
to the Automotive Process Approach to Audits, both
published by Paton Press. He has served on the Malcolm
Baldrige National Quality Award board of examiners and
as an RABQSA certified lead auditor. Kymal has a masters
degree in industrial and operations engineering and an
MBA, both from the University of Michigan.
comment
Please
If you would like to comment on this article,
please post your remarks on the Quality Progress
Discussion Board at www.asq.org, or e-mail
them to [email protected].