See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/232977086

Optimization design of control chart systems

Article in IIE Transactions · August 2010

DOI: 10.1080/07408170490426152

CITATIONS READS
18 914

4 authors, including:

Shiping Zhang Mohammad Shamsuzzaman

Dupont University of Sharjah
6 PUBLICATIONS 862 CITATIONS 80 PUBLICATIONS 1,135 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Mohammad Shamsuzzaman on 20 May 2014.

The user has requested enhancement of the downloaded file.

IIE Transactions (2004) 36, 447–455
Copyright 
C “IIE”

ISSN: 0740-817X print / 1545-8830 online

DOI: 10.1080/07408170490426152

Optimization design of control chart systems

ZHANG WU∗ , YEE CHEONG LAM, SHENG ZHANG and M. SHAMSUZZAMAN
School of Mechanical and Production Engineering, Nanyang Technological University, Singapore 639798
E-mail: [email protected]

Received November 2001 and accepted November 2003

This article proposes an algorithm to design an integrated control chart system consisting of several individual control charts, each
of which is used to monitor a critical process stage in the manufacture of a product. The design algorithm considers all the charts
within a system in an integrative and optimal manner. Consequently, the performance characteristics of the system as a whole can
be significantly improved and the product quality will be further enhanced. Such an improvement is achieved without requiring
additional cost and effort for on-line inspections. Furthermore, the floor operators can utilize and understand the optimal control
chart system as easily as they do the existing system. Some useful guidelines have also been highlighted to aid the users to adjust the
control limits of the control charts in a system.

1. Introduction drilled and milled, respectively, and dimensions x3 and x4

are determined. It is noteworthy that x2 depends directly
The fabrication of a product usually includes many pro- on x1 , and both x3 and x4 are directly affected by x2 . Four
cess stages. The integration of all these stages results in a control charts as a system are used to carry out the on-line
manufacturing system. Specifically, in the manufacture of a monitoring of the processes of the four machining stages x1
mechanical part, each stage corresponds to the machining to x4 . Furthermore, since two identical machines are used
of a dimension. As an example, several process stages are side by side to mill dimension x4 , control chart 4 has two
required to fabricate the dimensions of the part shown in duplicates, each controls one of the two parallel machines.
Fig. 1(a). Some of the dimensions are critical to the overall A block diagram of this control chart system is displayed
quality of the product and the corresponding process stages in Fig. 1(b). All non-functional process stages are ignored
have to be monitored using control charts. A control chart in the block diagram and are not included in the design of
system is the combination of all the charts that are used to the control chart system. It is assumed that the quality re-
monitor the critical stages of a manufacturing system. quirements on the non-functional dimensions can be easily
Due to differences in production rates and other factors, met and on-line monitoring is unnecessary.
some stages (e.g., stages with lower production rates) may In this article, a process means one single-stream stage
have more than one parallel stream (e.g., machine). In some or one stream in a multi-stream stage. The total number,
applications, a single Shewhart chart or group chart is used M, of the processes that are monitored by the control chart
to monitor the outputs from all the streams of a stage. How- system is equal to:
ever, in this article, a separate chart will be applied to the
output of each individual stream. This scenario helps to de- 
s
tect and diagnose the out-of-control stream (Montgomery, M= gi , (1)
2001). Usually, the parallel streams in a single stage have the i=1
same mean, standard deviation and target (Runger et al.,
1996), and therefore, all of them use an identical control where, s is the number of functional stages and gi is the
chart. In the machining of the part shown in Fig. 1(a), four number of streams in the ith stage (gi ≥ 1). However, since
critical dimensions xi (i = 1, 2, 3 and 4) are formed in four the gi streams in the ith stage are monitored by identical
stages. Dimension x1 is formed by facing surface 1 using control charts (with the same sample size, sampling inter-
surface 0 as the datum. Dimension x2 is associated with val and control limits), there are only s different charts in the
surface 2 which is machined with reference to surface 1. control chart system. In Fig. 1(b), the production flow en-
Then, using surface 2 as the datum, surfaces 3 and 4 are ters the system from the left-hand side. It first goes through
stages 1, 2 and 3 sequentially; then, branches into two par-
allel and identical streams in stage 4; and finally reaches the
∗
Corresponding author output point on the right-hand side. The production rate in

0740-817X 
C 2004 “IIE”
448 Wu et al.
being set at three. However, if all of the charts in a sys-
tem are designed in an integrated and optimal manner, the
overall effectiveness of the system may be improved. The
optimization design of the control chart system may result
in different k values or allocate a different power to dif-
ferent charts in a system based on the values of certain pa-
rameters (e.g., process capability, sampling interval, sample
size) that would affect the performance of the control chart
system.
The idea of applying multiple univariate control charts
to individual variables has been discussed in Montgomery’s
textbook on SPC (Montgomery, 2001). However, hitherto,
little literature can be found in the optimization design
of an integrated control chart system for a manufactur-
ing system. Woodall and Ncube (1985) investigated the
multivariate CUSUM control procedures and defined the
out-of-control condition for a system comprising several
charts. Their procedures for independent quality character-
istics can be applied to multi-stage manufacturing systems.
Nonetheless, in their study, the charting parameters (e.g.,
control limits) are not designed in an integrative and opti-
mal manner. Instead, identical parameter values are used
by individual charts. Hawkins (1991) discussed multivari-
ate quality control based on regression-adjusted variables.
The purpose of this technique is to make the controls more
effective than those based on individual variables. Nelson
(1986), and Mortell and Runger (1995) proposed and de-
veloped the group control chart for monitoring the out-
put from multiple streams of a single stage. Runger et al.
(1996) used principle components analysis to develop con-
trol charts that are able to detect both common and spe-
cial assignable causes in multiple process streams. Zhang
Fig. 1. (a) The dimensions of the part; (b) the process stages; (1989), Fong and Lawless (1998), and Lawless et al. (1999)
(c) the quality linkages. discussed the variation transmission through multiple pro-
cess stages. Wade and Woodall (1993), Ding, Ceglarek and
Shi (2002a, 2002b) and Ding, Shi and Ceglarek (2002) have
each stage may not be exactly the same and buffers may be published several papers studying multi-stage processes and
used between the stages when necessary. the diagnosis problems. Zantek et al. (2002) used quality
If any process is out of control, the control chart system linkages to handle the observed interdependence between
should produce a signal as quickly as possible, otherwise the quality characteristics in multi-stage processes. How-
the overall quality of the product is likely to deteriorate. ever, these papers did not study the multi-stage processes
However, if all the charts in the system report an in-control from the viewpoint of on-line SPC monitoring.
status, the whole manufacturing system is assumed to be in Similar to a single control chart, the power or effective-
control and the product quality should be satisfactory. ness of a control chart system is measured by the average
Conventionally, each individual chart in a system is de- time to signal, ATS. In this article, ATS is defined as the av-
signed in isolation. For example, the X̄ chart is a charting erage time that one of the control charts in the chart system
tool for monitoring the process mean shifts and is used most gives an out-of-control signal subsequent to any process
widely in SPC (Statistical Process Control). Its lower con- in a manufacturing system going out of control (Woodall
trol limit LCL and upper control limit UCL are calculated and Ncube, 1985). When any process is out of control, the
as follows: users want the control chart system to signal quickly, i.e.,
√ √ to have a small out-of-control ATS value. Conversely, when
LCL = µ0 − kσ/ n, UCL = µ0 + kσ/ n, (2)
the whole manufacturing system is in control, the users want
where, n is the sample size, µ0 and σ are the in-control the chart system to produce minimum false alarms, i.e., to
process mean and standard deviation respectively and k is have a large in-control ATS0 value.
the control limit coefficient. In conventional chart system A few assumptions and conventions will be adopted in
designs, k is identical for every individual chart, typically this article.
Optimization design of control chart systems 449

1. Since the processes will often operate in the in-control cording to the available resources (e.g., the instrument and
condition for most or relatively long periods of time manpower) and the expected detection power. The distribu-
(Montgomery, 2001), it is assumed that only one pro- tion parameters µ0,i and σi are usually estimated from the
cess is out of control at any moment in a manufacturing data observed in pilot runs or the process capability study.
system. It is a conservative assumption, because, if two The variable τ is specified based on the trade-off between
or more processes are out of control simultaneously (e.g., the false alarm rate and detection power. If the cost of han-
due to some common mode failures), the out-of-control dling the false alarms is high, a larger τ should be used in
ATS will be even smaller compared with the ATS under order to reduce the false alarm frequency. However, a large
this assumption. τ may at the same time impair the effectiveness of the con-
2. Only the X̄ chart will be studied in detail in this study. trol chart. The resultant (or actual) in-control ATS0 must
3. The underlying probability distribution of the quality be equal to or larger than τ . The maximum allowable mean
characteristic x (e.g., the dimension of a mechanical shift di indicates the rejectable quality level (the barely toler-
part) in each process is assumed normal with a constant ated mean shift). It may be determined by the QA (Quality
standard deviation. Assurance) engineer directly based on factors such as qual-
4. The gi parallel streams in the ith stage have the same ity, manufacturing cost and customers’ requirements. Or
mean, standard deviation and target and use identical more preferably, di is calculated from the specified toler-
control charts (with same sample size, sampling interval ance and the allowable process capability ratio Cpk . Each
and control limits). process must have a minimum allowable value cpk,min for the
process capability ratio. The product quality will degrade
significantly if Cpk is lower than cpk,min . Usually, cpk,min is
2. Optimization design of the control chart system decided by considering, for example, whether the process
is an existing one or a new one, or whether the product is
In this article, the control chart system is intended to moni- critical to safety (Montgomery, 2001). A default value of
tor all of the functional processes (or simply, the processes) one may be used for cpk,min . It results in a nonconforming
in a manufacturing system. For the design of this chart sys- parts per million; PPM, value of 2700. For the ith process
tem, the following parameters are specified. stage:
s = number of stages in the control chart system; cpk,min,i = USLi − (µ0,i + di ) /3σi . (3)
gi = number of streams in the ith stage;
Here, it is assumed that the in-control process mean µ0,i
hi = sampling interval of the control chart in the ith
coincides with the center between the lower specification
stage;
limit LSLi and the upper specification limit USLi (LSLi and
ni = sample size of the control chart in the ith stage;
USLi can be found in the engineering drawings). Therefore,
µ0,i = in-control process mean in the ith stage;
the maximum allowable mean shift di for the ith process
σi = constant standard deviation of the process in the
stage can be calculated as:
ith stage;
di = maximum allowable value of the mean shift δi in di = USLi − µ0,i − 3σi cpk, min,i (4)
the ith stage;
pi = probability of out-of-control occurrence in the ith Next, the probability pi that an out-of-control case takes
stage; place in the ith stage is estimated from historical data on
τ = minimum allowable in-control ATS0 ; out-of-control cases. For example, if in a manufacturing
Vi = a vector [v1 , v2 , . . . ]T indicating the cause stage (up- system, 10 out of 25 out-of-control cases occurred in the
stream influencing stage) numbers. The dimension first stage, then p1 ≈ 10/25 = 0.4. If such historical data
xi produced in the ith stage is directly dependent are not available, it may be reasonable to estimate pi by the
on the dimensions produced in all of these cause following formula:
stages; 
s
pi = gi / (gj ). (5)
i = induced mean shift. It is the mean shift undergone
j=1
by the ith stage due to (or incurred by) the mean
shifts of the cause stages. i is expressed as a func- This means that the estimated probability that the out-of-
tion fi (δv1 , δv2 , . . .) in terms of the output mean shifts control case happens in the ith stage is proportional to the
of all cause stages. number of parallel streams in this stage. Finally, the vectors
Vi and the formulae for i can be determined with the help
Most of these specifications can be determined easily. The
of the quality linkages (Zantek et al., 2002). For example,
numbers s and gi can be found from the block diagram (e.g.,
the cascade quality linkages in Fig. 1(c) indicate that x2
Fig. 1(b)). The parameters hi , ni , µ0,i , σi and τ are those usu-
is directly affected by x1 , and both x3 and x4 are directly
ally required by conventional chart designs. Generally, the
affected by x2 . Correspondingly:
sampling interval hi depends on the rational subgrouping
or the production shift. The sample size ni is decided ac- V2 = [1]T , V3 = [2]T , V4 = [2]T . (6)
450 Wu et al.
The formulae for i can take many forms and may be de- The major step in the optimization procedure is to
termined by many different methods. For part machining, express the objective function ATS (Equation (8)) and
these formulae may be determined based on the informa- the constraint function ATS0 (Equation (9)) in terms of
tion of design dimensioning and process planning. For the α1 , α2 , . . . , αs . The detailed formulae are derived in the
example in Fig. 1(a): Appendices (Equation (A2) for calculating ATS0 , and
Equation (A12) for calculating ATS).
2 = f2 (δ1 ) = δ1 , 3 = f3 (δ2 ) = δ2 ,
= f4 (δ2 ) = δ2 .
4
The optimization design can be carried out by using any
(7)
software for nonlinear constrained programming. In our
Here, for example, ( 2 = δ1 ) indicates that stage 2 will have
implementation, a dynamic search algorithm is employed.
an induced mean shift of 2 if stage 1 has an output mean
In this algorithm, the optimal values of αi of the first (s−1)
shift of δ1 , and 2 is equal to δ1 .
control charts are searched step by step in (s−1) levels, using
Based on the above specifications, the optimization de-
the same step size dα. The last α s is finally determined so
sign of the control chart system can be conducted by using
that the resultant ATS0 is exactly equal to the specified τ
the following nonlinear optimization model:
(Equation (9)).
Objective function: ATS = minimum, (8) Combining Equations (A1) and (A2) in the Appendix,
Constraint function: ATS0 ≥ τ, (9) and replacing ATS0 by τ , the value of any individual αi can
Design variables: k1 , k2 , . . . , ks , be expressed in terms of other αj (j = 1, 2, . . . , s, j = i) and
τ:
where, ATS is the average time to signal of the control chart   1/gi 
system when any process in the manufacturing system is out 1 − 1/τ
of control (or more specifically, when any process mean µi αi = hi 1 − s . (12)
j=1,j =i (1 − αj / hj )
gj
has shifted by an amount of di ). ATS0 is the average time
to signal when the whole system is in control (or all process This equation can be used to determine the range of possible
means remain in µ0,i ). α i value in the ith level search. As a simple example, let us
In the actual optimization search, the design variables consider a manufacturing system with three stages (s = 3),
k1 , k2 , . . . , ks (control limit coefficients) are replaced by the where g1 = g2 = g3 = 1, h1 = 100, h2 = 120, h3 = 200.
corresponding type-I error probabilities α1 , α2 , . . . , αs as- There are two levels of searches. The first level is to search
sociated with the s control charts. The variable αi is the the optimal α 1 . The maximum possible value of α 1 can be
probability that the ith control chart produces an out-of- found from Equation (12) by setting both α 2 and α 3 equal
control signal when the ith process is in fact in control: to zero (i.e., all the power is allocated to the first chart):
−1
ki = (1 − 0.5αi ), (10) α1,max = 100/τ. (13)
where −1 () is the inverse function of the cumulative prob- Therefore, α1 will be searched in the range of [0, 100/τ ] in
ability function of the standard normal distribution. Opti- the first level. The second level is to search the optimal α2,
mizing αi is equivalent to optimizing ki , but the former is while α1 is given. The maximum possible value α 2,max of
easier to handle during the optimization search. Because α 2 should be determined dynamically, because it is depen-
αi has a finite lower bound (zero) and upper bound (one) dent on the current value of α1 given in the first level. The
for the search (conversely, the upper bound of ki is infinite) maximum value α2,max is also found from Equation (12) by
and a simple equation among α 1 , α 2 , . . . , α s (Equation setting α3 equal to zero:
(12)) can be established easily. The objective of the opti-
mization model is to identify the optimal values for (α 1 , α2,max = 120(1 − (1 − 1/τ )/(1 − α1 /100)). (14)
α2 , . . . , α s ) that will jointly minimize ATS (maximize ef-
Therefore, α2 will be searched between zero and α 2,max in
fectiveness) and ensure ATS0 (the false alarm rate) so that
the second level. Finally, α 3 is directly determined using
it is equal to the specification. The underlying idea of the
Equation (12) without any search. That is, α 3 (or α s in a
optimization design is to enhance the power of those charts
general case) will take the maximum possible value in order
which take a very long period to produce a signal when the
to make full use of the power of the control chart system
corresponding processes become out of control. This pro-
(determined by the specified τ ). In this example:
cedure must sacrifice the power of some other charts in the
system. However, if the α i values of all charts are adjusted α3 = 200(1 − (1 − 1/τ )/[(1 − α1 /100)(1 − α2 /120)]). (15)
in an integrated and optimal manner, the overall (average)
effectiveness of the control chart system will be improved. It is clear that the resultant design point (α 1 , α 2 , α 3 ) will
The final outputs of the optimization design are the op- ensure that ATS0 is equal to τ .
timal control limits for each control chart. They can be de- The overall optimization search is outlined as follows:
rived from the optimal value of αi through ki (see Equation Step 1. Specify s, gi , hi , ni , µ0,i , σ i , di , pi ,τ , Vi and i .
(10)). Step 2. Initialize ATSmin as a large number, say 1000 000.
√ √
LCLi = µ0,i − ki σi / ni , UCLi = µ0,i + ki σi / ni . (11) ATSmin is used to record the minimum ATS.
Optimization design of control chart systems 451

Step 3. Carry out the searches in (s – 1) levels. by Equation (16). Therefore, the value of τ (the minimum
3.1 Each design point (α 1 , α 2 , . . . , α s ) will auto- allowable in-control average time to signal) is fixed at 64 750
matically make ATS0 equal to τ . throughout the sensitivity study.
3.2 Calculate ATS by Equation (A12) in the The performance of a two-stage manufacturing system
Appendix. will be studied for six different cases. The two stages are
3.3 If this ATS value is smaller than the current assumed to be independent of each other in the sensitivity
ATSmin , replace ATSmin by ATS and record study. In each case, all the parameters take their nominal
the current design point (α 1 , α 2 , . . . , α s ) as the values in both stages, except that one selected parameter
temporary optimal solution. (called the active parameter) takes its low value in stage
1 and high value in stage 2. For example, in case 1, the
Step 4. When the whole search is completed, the optimal first parameter gi is designated as the active parameter.
design point (α 1 , α 2 , . . . , α s ) is finalized. Therefore,
Step 5. Calculate the control limits of the s control charts
using Equations (10) and (11). g1 = 1, g2 = 4;
This search algorithm is simple and reliable. A computer h1 = h2 = 175;
program system.c written in the C language has been devel- n1 = n2 = 6;
oped to carry out the optimization design. The value of the σ1 = σ2 = 0.025;
step size dα is decided so that the total number of search d1 = d2 = 0.03;
points is always limited to 100 000 regardless of the number p1 = p2 = 0.5. (17)
of αi . Usually, an optimal solution is obtained in a few CPU
seconds using a personal computer. The program system.c Each case will be investigated for two different control
is available by contacting the authors. chart systems. The first one is the conventional system,
where the α i values (or equivalently, the control limit co-
efficients) for the two control charts are made equal. The
3. Sensitivity study second one is the optimal system, in which the α i values of
the two control charts are determined by the optimization
This section will study the effects of six parameters (gi , hi , design proposed in this article. Both control chart systems
ni , σ i , di and pi ) on the performance of the control chart make ATS0 equal to τ (= 64 750). The ratio R between the
system. Each of these parameters has a low value, a nominal out-of-control ATS values resulting from each control chart
value and a high value, as shown in Table 1. system would be calculated as:
The nominal value denotes the value under normal cir- R = ATSoptimal /ATSconvention . (18)
cumstances. The nominal value of gi is equal to its low value,
because it is assumed that there is usually only one stream Obviously, if R is less than one, the optimal control chart
in a stage, unless the production rate in this stage is much system will outperform the conventional system when the
lower than that in other stages. corresponding active parameter has different values in dif-
ATS0 can be calculated from the in-control average ferent stages in a manufacturing system. The R values for
run length, ARL0 , that is the average number of samples all six cases are listed in Table 2 under column (7). It can be
counted from the beginning of the process to the first false observed that in all of the six cases, the optimal control chart
alarm: system achieves an obvious improvement in the effective-
ness compared to the conventional system. It is especially
ATS0 = h × ARL0 . (16) the case if any of the four parameters hi , ni , σ i or di has
In the sensitivity study, if the sampling interval hi takes different values in the two stages, the optimal control chart
its nominal value of 175 time units (the time unit is de- system will have a substantially smaller ATS than the con-
termined by the applications) and ARL0 assumes a typical ventional system. Among all the parameters, di (the maxi-
value of 370 as in a 3σ X̄ chart, ATS0 is found to be 64 750 mum allowable mean shift) promises the greatest potential
for using the optimal control chart system. A reduction of
ATS by 33.1% can be achieved, when di has different values
Table 1. The parameter values (d1 = 0.01, d2 = 0.05 in case 5) and all other parameters are
Parameter Low Nominal High fixed. The performance of the optimal control chart system
is relatively less sensitive to the parameters gi and pi . The
gi 1 1 4 actual R value would differ for different systems and cir-
hi 50 175 300 cumstances. However, it is believed that the optimal control
ni 3 6 10 chart system would be effective in general and beneficial to
σi 0.01 0.025 0.04 many real systems. The reason for this is that, in any man-
di 0.01 0.03 0.05
ufacturing system, the influential parameters would be in
pi 0.1 0.5 0.9
general different among several stages.
452 Wu et al.
Table 2. Sensitivity study

Optimal αi
Case Active parameter α1 α2 ATSoptimal ATSconvention Ratio R
(1) (2) (3) (4) (5) (6) (7)

1 g1 = 1 g2 = 4 0.001 0835 0.000 4048 463.723 490.265 0.945 86

2 h1 = 50 h2 = 300 0.000 2495 0.003 1364 321.917 454.122 0.708 88
3 n1 = 3 n2 = 10 0.002 4103 0.000 2924 581.421 700.603 0.829 89
4 σ 1 = 0.01 σ 2 = 0.04 0.000 0019 0.002 7008 714.924 1006.509 0.710 30
5 d1 = 0.01 d2 = 0.05 0.002 6886 0.000 0141 4074.377 6085.955 0.669 47
6 p1 = 0.1 p2 = 0.9 0.000 4683 0.002 2344 320.297 354.257 0.904 14

The optimal α i values of the optimal control chart system hi , take their nominal values in both stages. In cases 1 and
are also enumerated in Table 2 (columns (3) and (4)). They 2, one, and only one, of the two parameters n and h takes
provide the users with the general guidelines for adjusting different values in the two stages (in case 1, n1 = n2 , but
the α i values (or equivalently, the control limit coefficients) h1 = h2 ; in case 2, h1 = h2 , but n1 = n2 ); the values of the
for different charts in a system. For example, in case 2, hi ratio R are significantly smaller than one for both cases. In
(sampling interval) is the active parameter. In this case, a case 3, n1 > n2 and h1 < h2 . Both inequalities lead to looser
larger α value should be allocated to the chart with the control limits for the optimal control chart in stage 1 and
longer sampling interval. This means that, the control lim- tighter control limits for the optimal control chart in stage 2.
its of the chart with the longer sampling interval should be The effects of n and h are synergistic, because they enhance
made relatively tighter. This scenario is quite heuristic, be- each other in each of the two stages. It makes the value of
cause, if a chart with a longer sampling interval is capable of R extremely small (only 0.434 52), or in other words, makes
giving a signal earlier, the ATS value of the system could be the optimization design very effective. Conversely, in case
reduced to a greater degree. As a guideline, the control lim- 4, n1 < n2 and h1 < h2 . While the first inequality leads to
its of the following control charts should be made relatively tighter control limits in stage 1 and looser control limits in
tighter (or its α i is set larger): stage 2, the second inequality results in opposite changes in
the control limits in the two stages. In this case, the effects
1. The control chart that is used in a stage with a smaller of n and h are anti-synergistic, because they balance out
number of streams. each other in each of the two stages. As a result, the value
2. The control chart that has a longer sampling interval. of R is almost equal to one. In general, the effectiveness of
3. The control chart that has a smaller sample size. the optimization design is more notable when a synergistic
4. The control chart that is used to monitor a process with effect exists between the parameters. That is, the differences
a larger variance. of all the parameters drive the control limits of some charts
5. The control chart that is used to monitor a process with in a system to be tighter and the control limits of other
a smaller allowable mean shift. charts to be looser.
6. The control chart that is used to monitor a process stage
where an out-of-control case is more likely to occur.
4. Example
With these guidelines in mind, the users can adjust the
control limits of each chart in a system rationally and effec- A part is machined in a manufacturing system as shown in
tively, even if more complicated computation has not been Fig. 1(a). Stage 1 turns the upper surface (datum: surface 0);
carried out. stage 2 turns surface 2 (datum: surface 1); stage 3 drills the
Finally, it is interesting to look into the synergistic or anti- hole (datum: surface 2); and stage 4 mills the slot (datum:
synergistic effects between the parameters (e.g., between the surface 2). The dimensions and tolerances on the drawing
sample size n and sampling interval h). Table 3 displays the are determined by the design engineers. Since the process
results of four cases where all the parameters, except ni and in stage 4 is much more time consuming, two streams are
used in parallel. Among all dimensions, only four (x1 , x2 ,
x3 and x4 ) obtained at the above four stages are functional
Table 3. Synergistic or anti-synergistic effects between n and h
(or critical to the overall quality of the part) and have to
Case n1 h1 n2 h2 Ratio R be closely monitored by the X̄ charts. The block diagram
of the control chart system is displayed in Fig. 1(b) and
1 3 50 10 50 0.802 85 the quality linkages in Fig. 1(c). The specifications of the
2 3 50 3 300 0.645 87 system are listed in Table 4.
3 10 50 3 300 0.434 52
Originally, the system adopts the 3σ X̄ charts for all
4 3 50 10 300 0.994 93
the processes, i.e., all charts have the same type-I error
Optimization design of control chart systems 453
Table 4. The system specifications And stages 3 and 4 will be influenced indirectly (Equation
(A7)):
Number of stages in s=4
the system δ3 = 3 = δ2 = 0.030,
Number of streams g1 = g2 = g3 = 1, g4 = 2
δ4 = 4 = δ2 = 0.030.
Sampling intervals h1 = h2 = 100, h3 = h4 = 200
(minutes)
Sample sizes n1 = n2 = 5, n3 = n4 = 6 On the other hand, if stage 2 is out of control, only stages
In-control process µ0,1 = 16.00, µ0,2 = 13.00,
3 and 4 will be affected:
means (mm) µ0,3 = 8.00, µ0,4 = 11.00
Process standard σ1 = 0.020, σ2 = 0.019, σ3 = 0.039, δ3 = 3 = d2 /g2 = 0.013/1 = 0.013,
deviations (mm) σ4 = 0.018
δ4 = 4 = d2 /g2 = 0.013/1 = 0.013.
The computer program system.c works out the optimal
probability of 0.0027. As a result, the in-control ATS0 of control chart system for this example in 1.219 CPU seconds
the system is equal to 10584 minutes. In order to improve on a personal computer (Pentium IV, CPU 1.6 GHz). The
the effectiveness of the control chart system, optimization results are listed below:
design is carried out using the algorithms proposed in
this article. Thus, some more specifications for di , pi , Stage 1: LCL = 15.965 mm, UCL = 16.035 mm,
τ , Vi and i are required. The values of the maximum (α = 0.000102);
allowable mean shift di will be calculated from the upper Stage 2: LCL = 12.967 mm, UCL = 13.033 mm,
specification limits USL (shown in Fig. 1(a)) and the (α = 0.000102);
minimum allowable process capability ratio cpk, min . In this Stage 3: LCL = 7.962 mm, UCL = 8.038 mm,
example, cpk,min is specified as one for all the processes in
(α = 0.016590);
the manufacturing system. Thus, when PPM hits 2700 in
any process, the control chart system is required to provide Stage 4: LCL = 10.976 mm, UCL = 11.024 mm,
a signal as soon as possible. The maximum allowable mean (α = 0.000950);
shifts are calculated by Equation (4): ATS0 = 10 584 minutes ATS = 1452 minutes.
d1 = 16.09 − 16.00 − 3 × 0.020 × 1 = 0.030 mm; It can be observed that, while both the optimal con-
d2 = 13.07 − 13.00 − 3 × 0.019 × 1 = 0.013 mm; trol chart system and the conventional 3σ system have
the same ATS0 value, the out-of-control ATS (=1452)
d3 = 8.13 − 8.00 − 3 × 0.039 × 1 = 0.013 mm;
of the optimal system is only 55.8% of that of the conven-
d4 = 11.08 − 11.00 − 3 × 0.018 × 1 = 0.026 mm. tional system (=2601). This is a significant improvement in
For this set of mean shifts, the out-of-control ATS of effectiveness.
the conventional 3σ system is equal to 2601 minutes. The In the optimal solution, the majority of the type-I error
probabilities pi of the out-of-control occurrences are esti- probability (or the power) is allocated to stage 3. The rea-
mated from historical records, which show that the number son for this is that this stage has a relatively long sampling
of out-of-control cases occurring in each of the four stages interval, a large variance, a small allowable mean shift and
are four, three, eight and six, respectively. Therefore: a high probability of out-of-control occurrences.

p1 = 4/(4 + 3 + 8 + 6) = 0.190;
p2 = 3/(4 + 3 + 8 + 6) = 0.143; 5. Conclusions and discussions
p3 = 8/(4 + 3 + 8 + 6) = 0.381;
p4 = 6/(4 + 3 + 8 + 6) = 0.286. This article discusses the optimization design of a control
chart system for use in multi-stage processes. It is found
The QA manager requires that the false alarm rate of that by properly allocating the power among the individual
the optimal control chart system be maintained at the same charts, based on the values of the influential parameters, the
level as for the conventional 3σ system. Thus, τ is specified effectiveness of the system as a whole can be significantly
as 10 584 minutes. The dependent relationship vectors Vi improved and therefore the product quality is further guar-
and the formulae for i were shown in Equations (6) and anteed. The improvement in effectiveness is achieved with-
(7): out requiring extra cost and effort for on-line inspection.
This is because each chart in the optimal system uses the
V2 = [1]T , V3 = [2]T , V4 = [2]T ,
same sample size and sampling interval as in the conven-
2 = δ1 , 3 = δ2 , 4 = δ2 . tional system. Most of the specifications of the optimization
So, if stage 1 goes out of control, stage 2 will be directly design are similar to that required by the conventional sys-
affected (Equation (A8)): tem design or available from manufacturing information
δ2 = 2 = d1 /g1 = 0.030/1 = 0.030. and records. Furthermore, the use of the optimal control
454 Wu et al.
chart system will not in any sense increase the difficulties Wu, Z., Zhang, X.L. and Yeo, S.H. (2001) Design of the sum-of-
of the shop floor operators to run and understand the con- conforming-run-length control charts. European Journal of Oper-
ational Research, 132(1), 187–196.
trol charts. This is because the control charts in the optimal
Zantek, P.F., Wright, G.P. and Plante, R.D. (2002) Process and prod-
system are almost the same as that in the conventional sys- uct improvement in manufacturing systems with correlated stages.
tem, except that the positions of the control limits have been Management Scinece, 48(5), 591–606.
adjusted. Zhang, G.X. (1989) New diagnosis theory with two kinds of quality, in
The design algorithm of the optimal control chart system ASQC, 43rd Annual Quality Congress Transactions, pp. 594–599,
Toronto, Ontario, Canada.
can be easily computerized. The design of a system has to
be conducted only once, and the resultant optimal system
can be used continuously until the process parameters have
changed. Appendices
It is believed that the general algorithms discussed in this
article can be modified and then applied to the designs of Appendix 1: Calculation of the in-control ATS0
other types of control charts in addition to the X̄ chart. In a single time unit, the probability that the ith control
It is also noted that the actual model, layout, scenario and chart (or a duplicate of the ith chart in a stream) in the
interactions between stages of a particular system may be ith stage produces a false alarm is approximately equal to
somewhat different from the model discussed in this article, α i /hi . The probability that it does not produce a false alarm
but the proposed algorithm will be generally applicable. The is nearly equal to (1−α i /hi ). It is recalled that each of the gi
procedures and formulae can be easily modified to cope streams in the ith stage runs a duplicate of the ith control
with different models. chart. Therefore, the probability P0 that the whole control
Finally, the general guidelines for adjusting the control chart system generates a false alarm in a time unit is:
limits of each chart in a system are particularly useful and g 
s  i  s
universally applicable. P0 = 1 − (1 − αi / hi ) = 1 − (1 − αi / hi )gi .
i=1 j=1 i=1
(A1)
Finally,
References
ATS0 = 1/P0 . (A2)
Ding, Y., Ceglarek, D. and Shi, J. (2002a) Fault diagnosis of multistage
manufacturing processes by using state space approach. ASME Appendix 2: Calculation of the out-of-control ATS
Transactions, Journal of Manufacturing Science and Engineering,
124, 313–322. It has been assumed that only one process (i.e., one stream
Ding, Y., Ceglarek, D. and Shi, J. (2002b) Design evaluation of multi- in a stage) in a manufacturing system may go out of control
station assembly processes by using state space approach. ASME
Transactions, Journal of Mechanical Design, 124, 408–418.
at any moment. The probability wi that the control charts
Ding, Y., Shi, J. and Ceglarek, D. (2002) Diagnosability analysis of multi- in the ith stage do not give a signal is calculated differently
station manufacturing processes. ASME Transactions, Journal of for the following three cases:
Dynamic Systems, Measurement, and Control, 124, 1–13.
Fong, D.D.T. and Lawless, J.F. (1998) The analysis of process variation Case 1. Suppose that one of the gk streams in the kth stage
transmission with multivariate measurements. Statistica Sinica, 8, (1 ≤ k ≤ s) is out of control (i.e., its mean shift is
151–164. equal to dk ). The probability β k of the type-II error
Hawkins, D.M. (1991) Multivariate quality control based on regression- that the corresponding control chart fails to give a
adjusted variables. Technometrics, 33, 61–75.
Lawless, J.F., MacKay, R.J. and Robinson, J.A. (1999) Analysis of vari-
signal in a single sample is:
ation transmission in manufacturing processes–part I. Journal of UCLk − (µ0,k + δk )
Quality Technology, 31, 131–142. βk = √
Montgomery, D.C. (2001) Introduction to Statistical Quality Control, σk / nk
Wiley, New York, NY. LCLk − (µ0,k + δk )
Mortell, R.R. and Runger, G.C. (1995) Statistical process control of mul- − √ , (A3)
tiple stream processes. Journal of Quality Technology, 27, 1–12.
σk / nk
Nelson, L.S. (1986) Control chart for multiple stream processes. Journal where, the control limits LCLk and UCLk are de-
of Quality Technology, 18, 255–256.
pendent on α k (Equations (10) and (11)). The prob-
Reynolds, M.R. and Stoumbos, Z.G. (2000) A general approach to mod-
eling CUSUM charts for a proportion. IIE Transaction, 32, 515–535. ability that none of the gk duplicates of the kth
Runger, G.C., Alt, F.B. and Montgomery, D.C. (1996) Controlling multi- chart gives a signal in a single sample is:
ple stream processes with principal components. International Jour-
nal of Production Research, 34(11), 2991–2999. wk = (1 − αk )gk −1 βk . (A4)
Wade, M.R. and Woodall, W.H. (1993) A review and analysis of cause-
selecting control charts. Journal of Quality Technology, 25, 161–169. Case 2. If the ith stage stays in control and is not influenced
Woodall, W. H. and Ncube, M. M. (1985) Multivariate CUSUM quality by the mean shift in the out-of-control stage k (e.g.,
control procedures. Technometrics, 27(3), 285–292. in Fig. 1(b), stage 1 is not influenced if mean shift
 Optimization design of control chart systems 455

occurs in stage 2, 3 or 4; stage 2 is not impacted if length, ARL. Since an out-of-control case may occur in any
mean shift takes place in stage 3 or 4; and stages 3 of the s stages, the final ATS for the control chart system
and 4 are independent of each other), then, for the is:
duration of one sampling interval hk of stage k, the

s
probability that a duplicate of the ith chart gives a ATS = (atsk × pk ), (A12)
signal due to type-I error can be approximated by: k=1

θ = αi hk / hi . (A5) where pk is the probability that the out-of-control case takes

The ratio hk /hi is used to match the sampling inter- place in the kth stage.
val hi in the ith stage to hk in the kth stage. Thus, All of the formulae developed in the Appendices have
the probability that none of the charts in the ith been verified by Monte Carlo simulation. Admittedly, a
stage gives a signal during a time period of hk is: chart monitoring a process that is currently in control may
produce an out-of-control signal (a misleading signal). This
wi = (1 − θ)gi . (A6) is mainly because of the induced mean shifts discussed in
case 3. Therefore, whenever a control chart produces an out-
Case 3. If the ith stage stays in control by itself, but its mean of-control signal, all of the upstream processes, which have
value is affected directly or indirectly by the out- direct or indirect impacts on this process, must be investi-
of-control stage k, then each of its gi streams will gated. More advanced diagnostic methods can be found in
undergo an induced mean shift: many references (Ding, Ceglarek and Shi, 2002a; Ding, Shi
δi = i = fi (δv1 , δv2 , . . .), (A7) and Ceglarek, 2002; Zantek et al., 2002). The type-I error
may also generate a misleading signal, but it is negligible.
where, δv1 , δv2 , . . . are the output mean shifts of
the cause stages (or the immediate upstream stages
having impact on the ith stage). The cause stages
can be found from the vector Vi (or the quality link- Biographies
ages). If a cause stage is the out-of-control stage k,
the output mean shift is: Zhang Wu is an Associate Professor in the School of Mechanical and
Production Engineering, Nanyang Technological University, Singapore.
δk = dk /gk . (A8) His current research interests include the developments of algorithms in
quality control, reliability, nonlinear optimization and geometrical toler-
This is because one, and only one, of the gk streams ance. He is a member of the American Society for Quality.
in the kth stage is out of control (i.e., having a mean Ph.D. 1988, McMaster University, Canada.
shift of dk ). For the duration of one sampling inter- M.Eng 1984, McMaster University, Canada.
val hk of the out-of-control stage k, the probability B.Eng 1982, Huazhong University of Science and Technology, China.
that a duplicate of the ith chart gives a signal is Yee Cheong Lam is a Professor in the School of Mechanical and Pro-
calculated by: duction Engineering, Nanyang Technological University, Singapore. His
current research interests include the development of algorithms for the
θ = (1 − βi )hk / hi . (A9) simulation and optimization of various manufacturing processes. He is a
member of the Institution of Engineers, Australia.
This is similar to Equation (A5) except that α i is replaced by
(1−β i ), and β i is calculated by Equation (A3) as β k . Then, Ph.D. 1983, Melbourne University Australia.
B.Eng 1979, Melbourne University Australia.
the probability wi for case 3 can be evaluated by Equation
(A6). Sheng Zhang is a Postdoctoral Fellow in the School of Mechanical and
Combining all of the three cases, the probability that an Production Engineering, Nanyang Technological University, Singapore.
out-of-control signal is produced during a time period of His current research interests include quality control, condition monitor-
hk can be calculated as: ing and fault diagnosis.
Ph.D. 1998, Northeastern University, China.
 s
M.Eng 1996, Northeastern University, China.
qk = 1 − (wi | stage k is out of control). (A10) B.Eng 1994, Shenyang Institute of Golden Technology, China.
i=1

Therefore, given that one process in the kth stage goes out M. Shamsuzaman is a Ph.D. research student in the School of Mechanical
and Production Engineering, Nanyang Technological University, Singa-
of control, the value of the out-of-control ATS is: pore. His current research area is statistical quality control.
ATSk = (1/qk − 1)hk + 0.5hk (A11) M.Eng 2000, Asian Institute of Technology, Thailand
B.Eng 1994, Bangladesh Institute of Technology Chittagong,
This is the steady-state formula, which provides a more Bangladesh.
realistic evaluation of the out-of-control ATS than the zero-
state counterpart (Reynolds and Stoumbos, 2000 Wu et al.,
2001). In Equation (A11), 1/qk is actually the average run Contributed by the On-line Quality Engineering Department

View publication stats