Upload
trinhngoc
View
231
Download
5
Embed Size (px)
Citation preview
Pre-control and Some Simple Alternatives
Stefan H. SteinerDept. of Statistics and Actuarial Sciences
University of WaterlooWaterloo, N2L 3G1 Canada
Pre-control, also called Stoplight control, is a quality monitoring scheme
similar to a control chart. However, the goal of Pre-control is the
prevention of nonconforming units. The appeal of Pre-control is due
mainly to its ease of implementation, lack of assumptions, and reported
successes. However, it has come under much criticism since many of its
decision rules appear ad hoc. In addition, there is some confusion
regarding different versions of Pre-control suggested in the literature. This
article compares the various Pre-control schemes and contrasts them with
more traditional control charts such as acceptance control charts and X
charts. Based on this analysis the conditions under which Pre-control is
appropriate are laid out. Also, alternatives that retain much of the simplicity
of Pre-control but have improved operating characteristics are suggested.
Key Words: Acceptance Control Charts; Grouped Data; Modified Pre-
control; Operating Characteristic; Process Capability; Stoplight Control
2
1. Introduction
Pre-control (PC), sometimes called Stoplight control, was developed to monitor the
proportion of nonconforming units produced in a manufacturing process. Implementation
is typically very straight-forward. All test units are classified into one of three groups:
green, yellow, or red, where the colors loosely correspond to good, questionable, and poor
quality products (see Figure 1). Decisions to stop and adjust the process or to continue
operation are made based on the number of green, yellow, and red units observed in a
small sample. The goal of PC is to detect when the proportion of nonconforming units
produced becomes too large. Thus, PC monitors the process to ensure that process
capability ( Cpk ) remains large.
PC was initial proposed in 1954 (see Shainin and Shainin, 1989 and Ermer and
Roepke, 1991 for more details) as an easier alternative to Shewhart charts. However, since
that time, at least three different versions of PC have been suggested in the literature. In
this article the three versions are called:
• Classical PC
• Two-stage PC
• Modified PC
Classical PC refers to the original formulation of Pre-control as described by
Shainin and Shainin (1989) and Traver (1985). Two-stage PC is a modification discussed
by Salvia (1988) that improves the method’s operating characteristics by taking an
additional sample if the initial sample yields ambiguous results. Modified PC was
suggested by Gurska and Heaphy (1991), and represents a departure from the philosophy
of Classical and Two-stage PC in that it compromises between the design philosophy of
Shewhart type charts and the simplicity of application of PC.
3
PC schemes are defined by two attributes; their group classification method and
their decision criteria. A third criteria, the qualification procedure, is the same for all three
PC versions, and is this not discussed in this article. The most important difference
between the three versions of PC is the group classification method. Classical PC and
Two-stage PC base the classification of units on engineering tolerance or specification
limits. A unit is classified as green if its quality dimension of interest falls into the central
half of the tolerance range. A yellow unit has a quality dimension that falls into the
remaining tolerance range, and a red unit falls outside the tolerance range. Assuming,
without loss of generality, that the upper and lower specification limits (USL and LSL) are
1 and –1 respectively, group classification is based on the endpoints: t = [–1, –.5, .5, 1].
The classification scheme is illustrated in Figure 1. The colored circle can be used for the
ease of the operators as a dial indictor overlay (Shainin and Shainin, 1989).
Red
Green
YellowYellow
Green YellowYellow RedRed
Measurement-1 -.5 .5 1
USLLSL
Figure 1: PC Classification Criteria
Let the quality dimension of interest be Y. Then, given a probability density
function for observations f y( ), the group probabilities for Classical and Two-stage PC are
given by (1.1).
P green( ) = Pg = f y( )dyy=−.5
.5
∫P yellow( ) = Py = f y( )dy
y=−1
−.5
∫ + f y( )dyy=.5
1
∫ (1.1)
P red( ) = Pr = f y( )dyy=−∞
−1
∫ + f y( )dyy=1
∞
∫
4
Using Modified PC, on the other hand, the group classification scheme is the same
as that used with Classical PC and Two-stage PC with one very important difference.
Modified PC uses control limits, as defined in Shewhart charts, rather than tolerance limits
to define the boundaries between green, yellow and red units. As shown later, this change
is fundamental and makes Modified PC much more like a Shewhart chart than like Classical
PC. Group probabilities for the Modified PC procedure are given in (1.2). Note that to
avoid confusion, through this article the current process mean and standard deviation are
denoted µ and σ , whereas estimates of the in-control mean and standard deviation used to
set the control limits for Modified PC and Shewhart type charts are denoted µc and σc .
Pm green( ) = f y( )dyµc −1.5σc
µc +1.5σc
∫Pm yellow( ) = f y( )dy
µc −3σc
µc −1.5σc
∫ + f y( )dyµc +1.5σc
µc +3σc
∫ (1.2)
Pm red( ) = f y( )dy−∞
µc −3σc
∫ + f y( )dyµc +3σc
∞
∫
The group probabilities given by Equations (1.1) and (1.2) are the same when µc =
0 and σc = 1/3. This makes sense because in that case control limits and tolerance limits
are the same. It has been suggested that Classical PC and Two-stage PC are only
applicable if the current process spread (six process standard deviations) covers less than
88% of the tolerance range (Traver, 1985). With specification limits at ±1, as defined
previously, this condition corresponds to the constraint σ < 0.29333.
The second important distinction between the three PC versions is their decision
criteria. With Classical PC the decision to continue operation or to adjust the process is
based on only one or two sample units. The decision rules are given below:
Classical PC Decision Procedure
Sample two consecutive parts A and B
• If part A is green continue operation (no need to measure B).
5
• If part A is yellow measure part B. If part B is green continue operation,
otherwise stop and adjust process.
• If part A is red stop and adjust process (no need to measure B).
The Two-stage PC and Modified PC decision criteria are more complicated since
they are based on a two-stage procedure. If the initial two observations do not provide
clear evidence regarding the state of the process, additional observations (up to three more)
are taken. The decision criteria for Two-stage and Modified PC is given below:
Two-stage and Modified PC Decision Procedure
Sample two consecutive parts.
• If either part is red stop process and adjust.
• If both parts are green continue operation.
• If either or both of the parts are yellow continue to sample, up to an additional
three units. In the combined sample, if three greens are obtained continue
operation; and if three yellows or a single red are observed stop the process.
The advantage of this more complicated decision procedure is that more information
regarding the state of the process is obtained and thus decision errors are less likely. The
disadvantage is that, on average, larger sample sizes are needed, and thus implementation is
more demanding. Table 1 summarizes the comparison of the three versions of PC.
Table 1: Comparison of Pre-control Versions
Pre-controlVersion
GroupClassification
DecisionCriteria
Classic tolerance limits two observations
Two-stage tolerance limits five observations
Modified control limits five observations
Based on these comparisons it can be concluded that the purpose of the various
versions of PC are quite different. By design, Classical PC, and Two-stage PC tolerate
6
some deviation in the target mean of a process so long as the proportion nonconforming
does not become too large. Thus, process stability is not required, and Classical and Two-
stage PC can be quickly implemented since there is no need to estimate the current process
mean and standard deviation and show stability. Thus Classical and Two-stage PC are
very similar to Acceptance Control Charts. By contrast, the goal of Modified PC is to
monitor the process for deviations from stability. As a result, mean drifts are not tolerated,
and Modified PC is very similar to Shewhart type charts such as X charts.
2. Comparison with Traditional Control Charting Methods
In this section the various PC methods are compared with appropriate traditional
control charting techniques. In addition, the performance of PC under some special
situations is explored. Classical and Two-stage PC are compared with Acceptance Control
Charts (ACCs), and Modified PC is compared with an X chart. Previous comparisons of
Classical and Two-stage PC with more traditional control charts were made with X & R
charts (Mackertich, 1990, Ermer and Roepke, 1991), however this is the wrong
comparison. For both ACC and Classical and Two-stage PC the goal is not statistical
control, but rather producing parts within specification. An ACC is designed to monitor a
process when the process variability is much smaller than the specification (tolerance)
spread. Under that assumption moderate drifts in the mean (from the target value) are
tolerable since they do not yield a significant increase in the proportion of nonconforming
units. Like Classical and Two-stage PC, ACCs are based on specification limits.
However, ACC limits are derived based on a distributional assumption and require a
known and constant process standard deviation. ACC limits are usually set assuming a
normal process, although if justified other assumptions could be made. For more
information on the design of ACCs see Ryan (1989). Modified PC, on the other hand,
bases the grouping criteria on the control limits, and thus Modified PC has the same goal as
a Shewhart X chart, namely statistical control of a process.
7
PC and traditional control charting techniques differ in a number of ways. The first
obvious difference is that PC uses only information from the classified (or grouped)
observations, whereas traditional control charts ( X charts and ACCs) use variables data.
Grouping data results in decision rules that are easy to implement, but clearly, the grouping
also discards information. The loss of information can be quantified by calculating the
expected statistical (Fisher) information available in the PC grouped observations.
Calculating the Fisher information available in grouped data is discussed in Kulldorff
(1961) and Steiner, Geyer and Wesolowsky (1995). Figure 2 shows the expected
information about the mean and standard deviation for various parameter values. The plots
are scaled so that variables data (infinite number of groups) would provide an expected
information content of unity for all values of µ and σ . The expected information graph
about µ is symmetric about µ = 0, thus negative mean values are not shown. Figure 2
shows that the group classification scheme used in PC is not very efficient when µ is close
to zero and/or σ is small. The small amount of information available when µ equals zero
may not be a major concern if at that mean value it is very easy to conclude that the process
should continue operation. This would be the case, for example, if the process standard
deviation were very small compared with the specification range. This issue is explored
further in Section 3.
Exp
ecte
d In
form
atio
n
µ0 0.5 1 1.50.3
0.4
0.5
0.6
0.7
0.8
0.2 0.3 0.4 0.5 0.6 0.7 0.80.3
0.4
0.5
0.6
0.7
0.8
0.9
σ
Exp
ecte
d In
form
atio
n
Figure 2: Expected Information about µ (left, assume σ =0.2933)and σ (right, assume µ =0), t = [–1, –.5, .5, 1]
8
PC and traditional control charts also differ in their decision rules. All control
charts require some decision rules to determine whether the process should be allowed to
continue operation or if it should be stopped. It is difficult to evaluate the effectiveness of
the decision rules in isolation from the grouping criteria. However, the overall
effectiveness of various control charts can be compared through their Operating
Characteristic (OC) curve. An OC curve shows the probability a control chart signals (or
fails to signal) for different parameter values. OC curves for ACCs and X charts can be
derived from results given in Ryan (1989). The probability Classical and Two-stage PC
schemes continue operating, called the probability of acceptance of the process, can be
found by calculating the probability of each combination of green and yellow units that
leads to acceptance. For Classical and Two-stage PC the results are given by equations
(2.1) and (2.2), see Salvia (1988).
Paccept Classical( ) = Pg + PyPg (2.1)
Paccept Two − stage( ) = Pg + Py( )2 − 2PgPy(1 − Pg3 − 3Pg2Py) − Py2 1 − Pg3( ) (2.2)
where Pg and Py are given by Equations (1.1).
Figure 3 shows the OC curves for Classical PC, Two-stage PC, as derived from
(2.1) and (2.2), and an ACC with sample size n = 2. Figure 3 shows that an ACC with
sample size n = 2 is superior to Classical PC since it has significantly better power to detect
mean shifts. Similarly, the Two-stage PC scheme is superior to an ACC with n = 2. Note
that different σ values simply shift the OC curves horizontally without changing the
ranking. Also, a process whose quality characteristic is approximately normally distributed
is assumed. A different underlying distribution for the quality characteristic may
dramatically change the probabilities of acceptance, but generally all the considered
procedure are similarly effected.
9
0 1 2 3 4 50
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Mean Shift in Sigma Units
Prob
abili
ty o
f A
ccep
tanc
e Classical Precontrol
2-stage Precontrol
ACC n = 2
Figure 3: OC curve Classical and Two-stage PC versus ACC with n=2N( µ , σ =.2933) process
Note however, that on average the Two-stage PC scheme requires a larger sample
size. Two-stage PC uses a partially sequential procedure and requires sample sizes of
between one and five units. In the appendix, the average sampling number of the Two-
stage PC procedure, denoted E n( ) is derived. E n( ) depends on the group probabilities,
given by (1.1), that are in turn dependent on the process parameters µ and σ . Figure 4
shows plots of E n( ) versus the process mean assuming normality. For example,
E n ;µ = 0,σ =.29333( ) = 2.4, and E n ;µ = 2σ,σ =.29333( ) = 3.5,
0 0.2 0.4 0.6 0.8 11.5
2
2.5
3
3.5
4
E n( )
µ 0 0.2 0.4 0.6 0.8 11.5
2
2.5
3
3.5
4
E n( )
µ
Figure 4: Average sample size E n( ) for Two-stage PC as function of µσ = 0.29333 on left, σ = 0.2 on right, N( µ , σ ) process
10
These results suggest a comparison of Two-stage PC with ACCs having larger
sample sizes. OC curves of Two-stage PC and ACCs with sample sizes of 3, 4 and 5 are
shown in Figure 5. Clearly, of the compared curves the PC procedure has least power to
detect mean shifts. In addition ACCs are likely easier to implement since they require fixed
sample sizes. On the other hand, the power of Two-stage PC is fairly competitive with an
ACC with n = 3, and ACCs have the disadvantage of requiring an estimate of σ , an
assumption that σ does not change, and precise or variables measurements rather than
grouped data. In addition, Two-stage PC requires smaller sample size on average when the
process is centered at µ = 0. In conclusion, Two-stage PC appears a reasonable
alternative to ACCs when little process knowledge is available or if estimating the process
mean and standard deviation are expensive.
0 0.5 1 1.5 2 2.5 3 3.5 40
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Prob
abili
ty A
ccep
t
Mean Shift in Sigma Units
2-stage Precontrol
ACC n=3
ACC n=5
ACC n=4
Figure 5: Comparison of Two-stage PC and ACC with n=3, n=4, n=5N( µ , σ =.29333) process
Classical and Two-stage PC schemes are designed to prevent defectives, but they
do not adapt to changes in process variability. Figures 6 and 7 explore the relation between
the probability Classical and Two-stage PC signal and the probability of a defect. Figure 6
11
shows a contour plot of the probability of a defect for various combinations of µ and σ ,
assuming USL = 1 and LSL = –1. The probability of a defect, or a nonconforming unit, is
given by Pr in Equations (1.1). Figure 7 shows contour plots of probability that Classical
and Two-stage PC signal (one minus the probability of acceptance as given by (2.1) and
(2.2). Ideally Figures 6 and 7 would look similar, however, this is not the case. The PC
methods are too conservative when σ is small. When σ is small compared with the
specification limits a significant drift in the process mean can be tolerated before the
proportion defective becomes large. However, using PC, large mean values together with
small σ values are likely to lead to a signal since then many yellow units would be
observed. As a result, PC is not applicable when σ is very small compared with the
specification limits.
-1 -0.5 0 0.5 10.1
0.15
0.2
0.25
0.3
0.35
0.4
1e-05
0.001
0.005
0.02
0.02
0.05
0.05
0.1
0.1
Mean
Stan
dard
Dev
iatio
n
Figure 6: Contour plot of the probability of a defect
12
-1 -0.5 0 0.5 10.1
0.15
0.2
0.25
0.3
0.35
0.4
1e-05
0.001
0.005
0.02
0.05
0.1
0.1
0.5
0.5
0.9
0.9
Mean
Stan
dard
Dev
iatio
n
-1 -0.5 0 0.5 10.1
0.15
0.2
0.25
0.3
0.35
0.4
1e-05
0.001
0.005
0.02
0.05 0.1
0.1
0.5
0.5
0.9 0.9
Mean
Stan
dard
Dev
iatio
n
Figure 7: Contour plot of P(signal) for Classical & Two-stage PC
ACCs do not suffer from this shortcoming since they adjust their control limits for
different σ values. Table 2 gives specific values shown on the contour plots and
equivalent values from an ACC with n = 5 for a direct comparison. Most enlightening are
the two cases: µ = 0, σ = .2, and µ = .6, σ = .1. In both cases the probability of a
defect is fairly small and approximately the same. However, when µ = 0, σ = .2 the
probability the PC scheme signals is very small, whereas when µ = .6, σ = .1 the
probability of a signal using PC is very large. Clearly, when σ is small, PC rejects some
processes that yield very few defects. This is appropriate if the goal is also to keep the
mean centered on the target value. However, the goal of PC is defect detection and
prevention. Thus, PC is too conservative when σ is small, say when 6σ covers less than
40% of the tolerance range. As shown in Table 2 an ACC with n = 5 does not suffer from
this problem to the same degree.
Table 2: Probability of defect and signals
µ σ P(defect)P(signal)
Classical PCP(signal)
Two-stage PCP(signal)
ACC n = 5
0 .1 ~ 0 3.2 e–13 1.7 e–18 ~ 0
0 .2 5.7 e–5 1.5 e–4 1.8 e–5 2.6 e–7
0 .3 8.6 e–4 .0099 .0088 not applicable
.5 .1 2.9 e–7 .25 .4688 1.3 e–7
.6 .1 3.2 e–5 .7079 .9540 .0018
.7 .1 .0013 .9550 .9994 .2489
13
Modified PC, unlike Classical and Two-stage PC, classifies units based on control
limits. As a result, Modified PC is the same in philosophy as an X control chart, since its
goal is to monitor the stability of a process. With Modified PC and X charts ideally any
drift in the process mean would be detected. The operating characteristics of Modified PC
can be determined using Equation (2.2) and (1.2) with µ = 0, σ = 1/3. Modified PC is
compared with traditional X control charts in Table 3. The average sample size of
Modified PC can be determined through the result in the appendix setting µ = 0 and σ =
1/3. A plot of the result would be similar to those shown in Figure 4.
Table 3: Modified PC compared with X control charts
µ
Psignal
ModifiedPC
Psignal
X chartn = 3
Psignal
X chartn = 4
Psignal
X chartn = 5
0 .0238 .0027 .0027 .0027
±1σ .2097 .1024 .1587 .2225
±2σ .8370 .6787 .8413 .9295
Table 3 shows that Modified PC has a very high false alarm rate that is an order of
magnitude larger than that for X charts. When the process is stable ( µ = 0) a Modified
PC chart signals a problem on average almost 2.4% of the time. This false alarm rate is too
high, since searching for assignable causes is usually time consuming and expensive, and
too many false alarms greatly reduces the credibility of the control chart and may lead to it
being ignored.
In summary, Classical and Two-stage PC are fairly competitive with ACCs, except
under certain circumstances. When the process variation is very small compared with the
tolerance range ( 6σ < .4T , where T is the tolerance range), Classical and Two-stage PC
lead to undesirable results such as rejecting a process that is producing virtually all parts
within specification. On the other hand, when the process variation is relatively large
compared to the tolerance range ( 6σ > .88T ) Classical and Two-stage PC yield
14
excessively large false alarm rates. However, if 6σ lies between 40% and 88% of the
tolerance range (with specification limits of –1 and 1, .29333 > σ > .13333) Classical or
Two-stage PC yield good results. Modified PC, on the other hand, is designed to have the
same goal as a Shewhart chart. Thus, Modified PC is similar to Two-stage PC with 6σ
equal to 100% of the tolerance range. As a result, Modified PC is not a good method since
it yields too many false alarms.
3. Alternatives to Traditional Pre-control
As discussed in Section 2, PC has a number of important advantages over
traditional control charts, mainly in terms of simplicity of implementation. However, its
design choices, in particular the grouping criteria and the decision rules, appear quite ad
hoc. We may ask if the performance of PC could be improved while retaining its
simplicity? In this section, three simple variations of PC called Ten Unit PC, Mean Shift
PC, and Simplified PC are considered. Each variation is very similar to Two-stage PC,
and requires only minor modifications in the design. The goal is not to develop the optimal
approach under particular assumptions, but rather to consider simple changes that retain the
flavor of PC. If optimal charts are desired, and more process information is available, the
grouped data ACCs and Shewhart type charts developed by Steiner, Geyer, and
Wesolowsky (1994, 1995) should be considered.
One reason why Two-stage PC is fairly competitive with ACCs in terms of power
is due to its partially sequential decision procedure. As a result, one possible improvement
approach is to make the decision process more sequential. A totally sequential procedure is
theoretically feasible, but would be difficult to implement since it would occasionally
require large sample sizes and defining the acceptance/rejection criteria would be difficult.
A version of PC that allows a maximum of ten units at each decision point is a
compromise. The proposed decision rules for Ten Unit PC are given below. This
sampling procedure should be repeated each time the process is to be monitored.
15
Ten Unit PC Decision Process
Take sample units one at a time defining #G and #Y as the number of green and
yellow units observed so far respectively.
• Stop the process
– if there are any red units, or
– if at any time #Y – #G 2 together with #Y 3, or
– if at any time #Y 5.
• Continue operation of the process and stop sampling if at anytime #G – #Y 2.
• Otherwise, continue to sample.
By enumerating all the possible paths to acceptance and rejection of Ten Unit PC it
is possible to derive expressions for the probability of a signal and the average sample size.
Table 4 shows a comparison of the operating characteristics of the proposed Ten Unit PC
scheme and Two-stage PC for various combinations of process mean and standard
deviation. Table 4 shows that Ten Unit PC has lower false alarm rates, more power to
detect two-sigma shifts in the mean, and requires approximately the same average number
of units as Two-stage PC.
Table 4: Comparison of Ten Unit and Two-stage PC
µ σP(signal)
Two-stage PCE n( )
Two-stage PCP(signal)
Ten Unit PCE n( )
Ten Unit PC
0 .29333 .0069 2.37 .0045 2.41
±1σ .29333 .1058 2.96 .0895 3.27
±2σ .29333 .7155 3.32 .7427 3.74
0 1/3 .0238 2.55 .0174 2.65
±1σ 1/3 .2097 3.11 .1959 3.52
±2σ 1/3 .8370 2.95 .8513 3.17
A second improvement idea stems from the realization that using PC, units are
classified into one of five groups (see Figure 1) but only three groups (green, yellow and
red) are used during the decision process. As a result, important information pertaining to
16
the process mean is ignored. The proposed Mean Shift PC utilizes this information. Mean
Shift PC is a very simple adaptation of Two-stage PC since it does not requiring any
changes in the data collection process. The Mean Shift PC method classifies observations
into one of four groups: Green, Yellow+, Yellow–, and Red. Red units above or below
the specification limits are not distinguished because they are very rare and automatically
lead to a signal in any case. The Mean Shift PC decision rules, given below, are similar to
those used in Two-stage PC.
Mean Shift PC Decision Process
Sample two consecutive parts.
• If either unit is red stop the process.
• If both units are green continue operation.
• If either or both units are yellow, sample an additional three units. If in the
combined sample three greens are obtained continue the process. If three
Yellow+ or three Yellow– or a single Red are observed then stop the process.
Table 5: Comparison of Two-stage PC and Mean Shift PC
µ σPsignal Two-stage PC
Psignal MeanShift PC
0 1/3 .0238 .0120
±1σ 1/3 .2097 .2031
±2σ 1/3 .8370 .8369
0 .29333 .0069 .0031
±1σ .29333 .1058 .1029
±2σ .29333 .7155 .7154
Table 5 compares the power of the Mean Shift PC and Two-stage PC. The table
shows that Mean Shift PC has a much smaller false alarm rate than Two-stage PC, and
virtually identical power to detect two sigma mean shifts. This advantage in terms of
operating characteristics will typically outweigh the slightly more complicated decision
17
process. Thus, Mean Shift PC is a good alternative to Two-stage PC, though the process
standard deviation should also be monitored in some way.
The third variation, called Simplified PC uses only green and yellow units. The
group probabilities and decision process for Simplified PC are given below:
P green( ) = Pg = f y( )dyy=−.5
.5
∫P yellow( ) = Py = f y( )dy
y=−∞
−.5
∫ + f y( )dyy=.5
∞
∫ (3.1)
Simplified PC Decision Process
Sample five consecutive parts
• If three or more units are yellow, then stop the process
• Otherwise continue operation.
The loss of efficiency using Simplified PC rather than Two-stage PC is very slight
when the process variation is in the range where PC is appropriate. This only small loss of
efficiency arises because the probability of a red unit is usually very small, and thus the red
classification provides very little information. In fact, as designed, Simplified PC has a
slightly smaller false alarm rate, though not quite as much power to detect mean shifts.
Table 6 shows a comparison. Compared with Two-stage PC, Simplified PC has the
advantage of requiring less effort in the group classification, and the decision rules are
more straight forward. In addition, although Simplified PC requires a larger sample size
on average often the required sample sizes are the same and implementation is usually
easier if the sample size is fixed.
18
Table 6: Comparison of Two-stage PC and Simplified PC
µ σPsignal Two-stage PC
Psignal
Simplified PC
0 1/3 .0238 .0193
±1σ 1/3 .2097 .1831
±2σ 1/3 .8370 .8258
0 .29333 .0069 .006
±1σ .29333 .1058 .0972
±2σ .29333 .7155 .7103
Simplified PC is also appealing since it can fairly easily be adapted when the
process variation is much smaller than the tolerance range. As shown in Section 2, PC is
not appropriate if the process standard deviation is less than around 1 15th of the tolerance
range. When σ is small, observing yellow units with the traditional grouping criterion as
defined by (1.1), is not necessarily evidence of a problem. See Figures 6 and 7. In this
circumstance the Simplified PC group limits [–.5, .5] can be replaced with the group limits
[–c, c], where c is closer to the tolerance limits (.5 < c < 1). By adjusting the group limits
it is possible to only detecting process mean shifts that yield an excessive proportion of
non-conforming units. The decision process remains unchanged.
The best value for c depends on the process standard deviation and the acceptable
and rejectable process levels (APL and RPL). A good choice for c is the midpoint between
the APL mean and the RPL mean. APL and RPL mean values can be estimated from
acceptable and unacceptable process capability values. Process capability is defined as Cpk
= min USL − µ( ) 3σ, µ − LSL( ) 3σ( ), where LSL and USL is the lower and upper
specification limits respectively. For example, assume the process standard deviation is σ
= .1, and that an acceptable process capability value is Cpk = 4/3 (an average of 6 defects
out of 100,000 units) whereas Cpk = 2/3 (an average of 94 defects out of 10,000 units) is
unacceptable. Then looking only at the upper specification limit (the problem is symmetric
on the lower limit) the corresponding APL and RPL mean values are µ = .6 and µ = .8
respectively. In this case choosing c = .7 is reasonable. Figure 8 illustrates the operating
19
characteristics obtained in this example for various parameter values. For example, the
probability of a signal when µ = .6 is .031 and the probability of a signal when µ = .8 is
.969. If this false alarm rate is too large adjusting the decision barrier c slightly higher may
be appropriate. An OC curve for Two-stage PC would show a much great chance of
rejecting the process at the acceptable mean value µ = .6.
0 0.2 0.4 0.6 0.8 10
0.2
0.4
0.6
0.8
1
µ
Psignal
Figure 8: OC curve of the adapted Simplified PC procedure
4. Conclusions
This article compared and contrasts previously recommended PC approaches and
suggests simple variations that improve performance. Classical and Two-stage PC are
compared with acceptance control charts rather than X charts since this is a more
appropriate comparison. It is concluded that Classical and Two-stage PC are good
methods when the process standard deviation σ lies in the range T 15 ≤ σ ≤ 11T 75 ,
where T is the tolerance range. Two-stage PC is preferred over Classical PC unless the
additional sampling required is very onerous. Modified PC, on the other hand, has the
same goal as an X chart, but is shown to have an excessively large false alarm rate, and is
thus not recommended.
Three alternatives to PC are suggested that retain the great advantage of very simple
implementation. The Ten Unit approach is recommended when improved operating
20
characteristics are desired and taking some additional observations is not difficult. The
Mean Shift approach yields better results than Two-stage PC when the goal is the detection
of process mean shifts, and Simplified PC utilizes a simplified classification and decision
procedure while providing similar operating characteristics. Simplified PC can also easily
be adapted to handle situations when the process standard deviation is smaller than T 15.
Appendix
In this appendix an expression for the average sampling number E n( ) of the Two-
stage PC scheme, discussed in Section 2, is derived. E n( ) is determined by calculating
the probability of each possible path to either acceptance or rejection of the process. Below
in (A.1), pi is the probability that the Two-stage PC reaches a decision in i units. For
example, Two-stage PC would stop after two units if we observe either two green units
(process acceptable), or if the first observation is yellow or green and the second
observation is red (process rejectable).
p1 = pr , p2 = pg2 + pr py + pg( ), p3 = pr py
2 + 2 pg py pr + py3 (A.1)
p4 = 3py2 pg pr + 2 pg
2 py pr + 2 pg3 py + 3py
3 pg , p5 = 5py2 pg
2 pr + 5py3 pg
2 + 5pg3 py
2
E n( ) = ipii=1
5∑
References
Ermer, D.S., Roepke, J.R. (1991), “An Analytical Analysis of Pre-control,” ASQC
Quality Congress Transactions - Milwaukee, 522-527.
Gurska, G.F., Heaphy M.S. (1991), “STOP LIGHT CONTROL - Revisited,” ASQC
Statistics Division Newsletter, Fall.
Kulldorff, G. (1961), Contributions to the Theory of Estimation from Grouped and
Partially Grouped Samples, John Wiley and Sons, New York.
Mackertich, N. A. (1990), “Pre-control vs. Control Charting: A Critical Comparison,”
Quality Engineering, 2, 253-260.
21
Ryan, T.P. (1989), Statistical Methods for Quality Improvement, John Wiley & Sons,
New York.
Shainin, D., Shainin, P. (1989), “Pre-control Versus X & R Charting: Continuous or
Immediate Quality Improvement?”, Quality Engineering, 1, 419-429.
Salvia, A.A. (1988), “Stoplight Control,” Quality Progress, September, 39-42.
Steiner, S.H., Geyer, P.L., Wesolowsky, G.O. (1994), “Control Charts based on
Grouped Data,” International Journal of Production Research., 32, 75-91.
Steiner, S.H., Geyer, P.L., Wesolowsky G.O. (1995), "Shewhart Control Charts to
Detect Mean Shifts Based on Grouped Data," submitted to Quality and Reliability
Engineering International.
Traver, R.W. (1985), “Pre-control: A Good Alternative to X -R Charts,” Quality
Progress, September, 11-14.