You are on page 1of 13

Received: 4 March 2018 Revised: 7 September 2018 Accepted: 18 November 2018

DOI: 10.1002/qre.2438

RESEARCH ARTICLE

CCC‐r charts' performance with estimated parameter for


high‐quality process

Min Zhang | Xuejun Hou | Hui Chen | Shuguang He

College of Management and Economics,


Abstract
Tianjin University, Tianjin, China
CCC‐r charts are effective in detecting process shifts in the nonconforming rate
Correspondence especially for a high‐quality process. The implementation of the CCC‐r charts
Hui Chen, College of Management and
Economics, Tianjin University, Tianjin
is usually under the assumption that the in‐control nonconforming rate is
300072, China. known. However, the nonconforming rate is never known, and accurate esti-
Email: chenhui2017@tju.edu.cn mation is difficult. We investigate the effect of estimation error on the CCC‐r
Funding information charts' performances through the expected value of the average number of
National Key R&D Program of China, observations to signal (EANOS) as well as the standard deviation of the average
Grant/Award Number: 2016QY02D0301;
number of observations to signal (SDANOS). By comparing the in‐control per-
National Natural Science Foundation of
China, Grant/Award Numbers: 71872123, formance of the CCC‐r charts, the CCC‐r chart with a larger value of r is more
7161101136 and 71472132 susceptible to the effects of parameter estimation. Meanwhile, the performance
of the CCC‐r charts can converge when detecting upward shifts in p of out‐of‐
control processes. We recommend the use of the CCC‐4 chart when consider-
ing its effectiveness in detecting shifts as well as its easier construction in
practice. Furthermore, it is investigated that the CCC‐4 chart is less sensitive
to parameter estimation while being more effective in detecting different pro-
cess shifts when compared with Geometric CUSUM chart and synthetic chart.

KEYWORDS
average number of observations to signal, Bernoulli data, CCC‐r charts, estimated parameter, high‐
quality processes

1 | INTRODUCTION

A high‐quality process, ranging from production of microelectronics to surveillance systems used for health care, is
always associated with a low nonconforming rate (i.e., p0).1,2 For instance, the nonconforming rate can be as small as
116 parts per million (ppm) for the thermo‐sonic wire bonding.1,3 Control charts based on Bernoulli data are highly rec-
ommended for its effectiveness in detecting shifts in p0, while charts based on variable data are costly and sometimes
unavailable in monitoring a high‐quality process.2,4,5
The optimality properties of various charts with Bernoulli data for high‐quality processes have been reviewed exten-
sively.6 Among them, CCC‐r chart, which is defined as the control chart based on the number of cumulative count of
items until r nonconforming items are observed, is found to be effective in detecting process deterioration.7,8 When
using the CCC chart (r = 1), a decision made about whether the process is in control or out of control is based only
on the number of conformities between successive nonconformities.9 However, the CCC chart is not effective in detect-
ing moderate shifts in p0 compared with the CCC‐2 chart.10 CCC‐r chart for r > 1 is more sensitive in detecting small
changes in process fraction nonconforming.11,12

946 © 2018 John Wiley & Sons, Ltd. wileyonlinelibrary.com/journal/qre Qual Reliab Engng Int. 2019;35:946–958.
ZHANG ET AL. 947

There are generally two phases in the application of control charts. In phase I, a set of time‐ordered data is collected
to learn about the process, estimate the process parameters, and determine an appropriate phase II monitoring scheme.
In phase II, the control chart is used to monitor changes from the assumed in‐control model.13 Generally, most of the
statistical process control (SPC) researches focus on phase II control charts by assuming that the parameters are known.
Nevertheless, the parameters of the phase I process are rarely known, and a considerably large sample size, which is
seldom available, is needed to accurately estimate p0.14,15 Thus, it is helpful to provide phase I guidelines so as to make
it easier for the practitioners to understand the effects of parameter estimation error on the performance of the control
charts.16
Furthermore, appropriate metrics are essential for the comparison of the surveillance schemes when investigating
the effects of parameter estimation on the control charts. When considering the number of conformities between two
adjacent nonconformities in a high‐quality process, the matter of great importance is not the number of consecutive
nonconformities when there is an out‐of‐control signal but the total number of both the conforming and
nonconforming observations to signal.14 Moreover, as different phase I sample sizes the practitioners choose do have
an influence on the estimation of the control limits, the practitioner‐to‐practitioner variability in the in‐control average
number of observations to signal (ANOS) should also be considered, which is the standard deviation of the average
number of observations to signal (SDANOS).17-20 Among them, the in‐control ANOS represents the average number
of observations to signal when there is a false alarm in the in‐control process. A smaller value of out‐of‐control ANOS
means that the nonconforming rate p0 has shifted.21 The effect of different phase I sample sizes on CCC chart when p0 is
estimated has been well investigated.14,17 We will further investigate the effect of parameter estimation on the CCC‐r
charts' (r > 1) performances with the use of the ANOS and SDANOS metrics in our study.
Moreover, CUSUM chart based on geometric counts, which is also known as Geometric CUSUM chart, is highly rec-
ommended for monitoring process deterioration of a high‐quality process.5,22 It is argued that the Geometric CUSUM
chart excels the CCC‐r charts in detecting shifts when either the in‐control nonconforming rate or the out‐of‐control
nonconforming rate is large while the CCC‐r charts outperform the Geometric CUSUM chart if either p0 or p is small
when parameters are known.7 Besides, the synthetic chart can result in a great reduction in the out‐of‐control average
time to signal.23,24 Therefore, we will investigate differences of the effect of parameter estimation on the CCC‐r charts
(1 ≤ r ≤ 7), Geometric CUSUM chart, and synthetic chart by using the ANOS and SDANOS metrics.
This paper is organized as follows: First, the background on the CCC‐r control charts, the Geometric CUSUM chart,
and the synthetic chart as well as the performance metrics are introduced. Second, we consider the performance of the
CCC‐r charts with different Bayesian estimators. Performance comparisons are made for the CCC‐r charts with in‐
control and out‐of‐control high‐quality processes. Then we compare the performance of the CCC‐r chart, the Geometric
CUSUM chart, and the synthetic chart, all with estimated parameters. Finally, we add an example to explain the use of
the CCC‐4 chart when parameters are estimated.

2 | BACKGROUND

2.1 | CCC‐r chart

CCC‐r chart is defined as a control chart based on the number of cumulative count of items until r nonconforming items
are observed.25 Suppose that the in‐control nonconforming rate of a high‐quality manufacturing process is p0 and the
shifted nonconforming rate is p1. Let x be the number of cumulative count of items inspected until r nonconforming
observations are observed.
When the nonconforming rate p is very small, there must be r nonconformities for the CCC‐r chart to signal. If p1 is
much less than p0, the waiting time until a nonconformity appears will be very long, not to mention r independent iden-
tical geometric counts.5 CCC‐r charts are very sensitive in detecting an upward shift in p0, which means process dete-
rioration.7 In practice, upper‐sided shifts are more meaningful and practical in a manufacturing process.21 Thus, we just
consider the CCC‐r charts for upward shifts in p0 in our study here.
For the CCC chart, x is a geometric random variable with the in‐control nonconforming rate p0. x for the CCC‐r
chart (r ≥ 2) follows a negative binomial distribution with parameters (r, p0). The probability mass function is

 
k−1
P ðx ¼ k Þ ¼ p0 r ð1−p0 Þk−r ; k ¼ r; r þ 1; (1)
r−1
948 ZHANG ET AL.

and the cumulative distribution function is


 
k k i−1
F ðk; r; p0 Þ ¼ ∑ Pðx ¼ iÞ ¼ ∑ p0 r ð1−p0 Þi−r : (2)
i¼r i¼r r−1

Based on Equation 2, LCL for the CCC‐r chart can be illustrated as


 
LCL i−1
∑ p0 r ð1−p0 Þi−r ¼ αL : (3)
i¼r r−1

αL represents the probability of the alarm rate when there is an upward shift in the nonconforming rate p0. αL for the
CCC‐r chart is given by Wu et al,7 which is

r
αL ¼ : (4)
ANOS0 × p0

ANOS0 is the desired average number of observations to signal when process is in‐control.
The out‐of‐control ANOS is given by

r
ANOS ¼     : (5)
LCL i−1
∑i¼r pr ð1−pÞi−r × p0
r−1

2.2 | Geometric CUSUM chart

Geometric CUSUM chart is also known to be an effective chart in monitoring a high‐quality process.2 Let xi be the ith
geometric counts between adjacent nonconformities. Then the upper‐sided Geometric CUSUM statistic {Gi} is

Gi ¼ maxð0; Gi−1 − x i þ kG Þ; for i ¼ 1; 2; …; (6)

where G0 is the initial Geometric CUSUM statistic. We consider the circumstance when there is no head‐start feature
(G0 = 0). Here, the acceptable reference value kG is based on the SPRT approach so that the Geometric CUSUM chart
can be designed to detect a shift10,26 from p0 to p1 (p1 = g × p0).

lnðp1 =p0 Þ
kG ¼ : (7)
ln½ð1 − p0 Þ=ð1 − p1 Þ

It should be noted that kG can be optimally determined using different optimization algorithms.27 The constant hG
represents the decision interval for the Geometric CUSUM chart, and a signal is given if Gi ≥ hG. hG can be calculated
based on the Markov chain approach by using a binary search approach with a given ANOS0.10

2.3 | Synthetic chart

Synthetic chart is a synthesis of np‐chart and CCC chart that was proposed by Wu et al,23 where n is the sample size of
inspected units in phase II and p is the ratio of the number of nonconforming units in a population to the total number
of units in that population. The way that a synthetic chart functions is basically similar to the CCC chart, except that
each item in the CCC chart is replaced with a sample of n items.28 In other words, conforming run length (CRL) with
a sample of n items is considered. If the number of nonconforming items in a sample exceeds an acceptance number c, it
is a nonconforming sample, and otherwise, it is a conforming sample. Since the number of nonconforming items d in a
sample follows a binomial distribution,23 the probability of a nonconforming sample is

c
Q ¼ 1 − ∑ Cni pi ð1−pÞn−i : (8)
i¼0
ZHANG ET AL. 949

The underlying distribution of the number of samples in a CRL is geometric,29 where the cumulative distribution
function is

F ðk; QÞ ¼ 1 − ð1−QÞk : (9)

The mean value of CRL is 1/Q. L is the CRL‐based lower control limit. In case the number of samples k including
the current nonconforming sample as well as the preceding conforming sample since the last nonconforming sample
falls below L, there is an out‐of‐control signal, which means the process has deteriorated. The out‐of‐control ANOS is

1 1
ANOS ¼ ×n× : (10)
Q F ðL; QÞ

The method for choosing the suitable parameters n, c, and L of the synthetic chart is given by Wu et al.23

2.4 | Performance metrics

A phase I stable high‐quality process with geometric counts is considered here. When the in‐control nonconforming
rate p0 is unknown, we have the design model as

Objective: minjANOSðb
p0 Þ − ANOS0 j
(11)
Constrai nts: ANOS0 ¼ 100; 000 :

b
p0 is the estimator of the in‐control process nonconforming rate p0. Design variables differ from chart to chart. For
b is the design variable. For the Geometric CUSUM chart designed
the CCC‐r chart, the estimated lower control limit LCL
to quickly detect a shift from p0 to p1 (p1 = g × p0), the design variables are kG and b
hG . The design variables of synthetic
b
chart are n, c, and lower control limit L. The optimal design of the chart will search for the optimal values of the design
variables so as to minimize the difference between the ANOSðb
p0 Þ with estimated parameters and the desired in‐control
ANOS0.
We choose the expected values of average number of observations to signal ANOSavg as a performance metric. We
have
 
m m
ANOSavg ¼ E N ½ANOSðb
p0 Þ ¼ ∑ ANOSðb
p0 Þ pk0 ð1−p0 Þm−k ; (12)
k¼0 k

where m is the phase I sample size of the historical geometric count and k is the number of nonconformities of the
sample.17
Moreover, as practitioners may choose different phase I samples for the estimation of the parameters, practitioner‐
to‐practitioner variation in the in‐control ANOS should be considered.17 Thus, we have
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
h i
2
SDN ½ANOSðb
p 0 Þ ¼ E N ANOSðb p0 Þ − E N 2 ½ANOSðb p0 Þ; (13)

where
h i  
2 m
m
2
E N ANOSðb
p0 Þ ¼ ∑ ANOSðb
p0 Þ pk0 ð1−p0 Þm−k : (14)
k¼0 k

3 | BAYESIAN ESTIMATOR F OR p 0

The Bayesian approach that incorporates the knowledge of practitioners is used to estimate the nonconforming rate p0
here. It can avoid the problem when there is no observed nonconforming item compared with the maximum likelihood
estimator (MLE).17 Suppose the prior distribution is beta(a, b), then the mean of the prior distribution, that is, b
p0;p , is
expressed as
950 ZHANG ET AL.

a
b
p0; p ¼ : (15)
aþb

With the observed phase I data, the posterior distribution changes to beta(N + a, m − N + b), and we have the mean
b
p0;B , which is

N þa
b
p0;B ¼ ; (16)
mþaþb

where N is the number of nonconforming items in the sample of size m. There are three types of beta distribution, which
are strictly decreasing model (a ≤ 1 and b > 1), U‐shaped model (a < 1 and b < 1), and unimodal model (a > 1 and
b > 1). The effect of parameter estimation with combinations of different parameters (a, b) on the CCC‐r chart will
be investigated. For simplicity, we take the well estimated case ( b p0;p ¼ p0 ¼ 0:001) with parameters (1, 999), (2,
1998), (3, 2997), and (4, 3996); the overestimated case with parameters (1, 499); and the underestimated case with
parameters (1, 1999) into consideration.
The values of E N ½ANOSðb p0 Þ of the CCC‐r charts (1 ≤ r ≤ 7) with different Bayesian estimators when the desired
ANOS0 = 100 000 and p0 = 0.001 are shown in Figure 1. It is obvious that E N ½ANOSðb p0 Þ values become closer to the
desired ANOS0 with an increase in the sample size m. With an increase in the value of r, the E N ½ANOSðb p0 Þ values
for different CCC‐r charts deviate gradually from the desired ANOS0 for a fixed m, except the trend for the CCC‐2 chart
with a = 3 and b = 2997 when compared with CCC‐3 chart. But it is clear that the CCC‐r charts are not very sensitive to
the choice of Bayesian parameters for the similar trends under different circumstances. Thus, any Bayesian parameters
we discussed above can be used for the prior estimation of p0.
It is also of great importance to investigate whether the probability distributions of number of observations to signal
(NOS) remain skewed and widely dispersed for the CCC chart with Bayesian parameter.17 Taking CCC‐2 chart as an
example, Figure 2 shows the histogram of the in‐control NOS distribution with beta (1999) when m = 1 000 000 and
p0 = 0.001. It is obvious that the distributions of NOS remain skewed and widely dispersed, which is consistent with
the conclusion of CCC chart.17 Thus, we recommend taking SDANOS as well as ANOS into consideration.

N þa
FIGURE 1 Values of EN ½ANOSðbp0 Þ for the CCC‐r chart when ANOS0 = 100 000, p0 = 0.001, and b p0 ¼ : A, a = 1, b = 999; B,
mþaþb
a = 2, b = 1998; C, a = 3, b = 2997; D, a = 4, b = 3996; E, a = 1, b = 499; F, a = 1, b = 1999 [Colour figure can be viewed at
wileyonlinelibrary.com]
ZHANG ET AL. 951

FIGURE 2 The probability distribution of in‐control NOS for ANOS0 = 100 000, m = 1 000 000, and p0 = 0.001 for CCC‐2 chart [Colour
figure can be viewed at wileyonlinelibrary.com]

4 | PERFORMANC E O F THE CC C ‐r CHARTS WITH ESTIMATED


PARAMETERS

As we discussed above, the choice of Bayesian parameter does not have much influence on the performance of the
CCC‐r charts. We will investigate the in‐control performance and the out‐of‐control performance of the CCC‐r charts
with Bayesian parameter (1, 999) here.

4.1 | In‐control performance

Table 1 shows the values of E N ½ANOSðbp0 Þ and SDN ½ANOSðb p0 Þ when ANOS0 = 100 000 and p0 = 0.001. The first row
corresponds to the values of E N ½ANOSðb
p0 Þ, and the second row corresponds to the values of SDN ½ANOSðb p0 Þ. For a sin-
gle column, with the increase of phase I sample size m, E N ½ANOSðb p0 Þ and SDN ½ANOSðb
p0 Þ decrease and become closer
to ANOS0 and zero, respectively. Taking CCC‐4 chart, for example, the E N ½ANOSðb p0 Þ = 103 164 and
SDN ½ANOSðb p0 Þ = 29 171 when m = 100 000, which are relatively large. Then E N ½ANOSðb p0 Þ decreases to 100 028
accompanied by a small SDN ½ANOSðb p0 Þ = 4002 when m = 5 000 000. Furthermore, it should be noted that the

N þ1
TABLE 1 Values of EN ½ANOSðb
p0 Þ and SDN ½ANOSðb
p0 Þ when ANOS0 = 100 000, p0 = 0.001, and b
p0 ¼
m þ 1 þ 999
m CCC CCC‐2 CCC‐3 CCC‐4 CCC‐5 CCC‐6 CCC‐7

50 000 98 589 102 310 104 207 106 129 108 035 110 025 111 763
23 890 28 889 35 703 41 382 46 267 50 781 54 580
100 000 99 448 100 562 102 234 103 164 104 113 105 075 105 936
18 772 19 443 25 387 29 171 32 405 35 233 37 615
200 000 99 787 100 323 101 130 101 570 102 034 102 469 102 911
14 142 14 199 17 914 20 546 22 712 24 513 26 128
500 000 99 935 100 075 100 400 100 564 100 774 100 968 101 119
9 331 9 011 11 184 12 901 14 255 15 396 16 326
1 000 000 99 971 99 980 100 169 100 286 100 381 100 433 100 531
6697 6292 7912 9073 10 045 10 833 11 490
2 000 000 99 987 99 846 100 055 100 134 100 179 100 231 100 263
4772 4407 5601 6455 7084 7653 8151
5 000 000 99 995 99 955 100 325 100 028 100 449 100 148 100 103
3032 3546 4052 4002 5144 5004 5163
∞ 100 000 100 080 100 179 100 046 100 029 100 017 100 007
0 0 0 0 0 0 0
952 ZHANG ET AL.

E N ½ANOSðb p0 Þ values are not exactly the desired ANOS0 when referring to the CCC‐r charts (r ≥ 2) due to the rounded
control limits. It is also found from Table 1 that E N ½ANOSðb p0 Þ converges more quickly to the desired ANOS0 than
SDN ½ANOSðb p0 Þ converges to zero. When comparing the values of different CCC‐r charts in Table 1, a much larger sam-
ple size m is needed with the increase of r so as to get the similar in‐control performance. For example, if the practi-
tioner aims to achieve the SDN ½ANOSðb p0 Þ value to be 10% of the E N ½ANOSðbp0 Þ, the CCC‐7 chart needs a sample
size of about 2 000 000 while the CCC chart needs a sample size of about m = 500 000, which is relatively small.
In sum, the CCC‐r charts with a large value of r are susceptible to parameter estimation when the high‐quality pro-
cess is in control.

4.2 | Out‐of‐control performance

The effects of parameter estimation on the out‐of‐control performance of the CCC‐r charts will be further investigated
here. Figures will be mainly used to show the trends of all E N ½ANOSðb p0 Þ and SDN ½ANOSðb p0 Þ values in order to stream-
line the paper and show the difference of CCC‐r charts under different shifts of p. Figures 3 and 4 provide an overview
that shows the effects of parameter estimation on out‐of‐control performance of the CCC‐r charts when p = 0.002, 0.003,
0.004, 0.005, 0.006, and 0.007 with the in‐control ANOS0 = 100 000 and p0 = 0.001. There are six parts from Figure 3A to
Figure 3F where two subparts are included in each part. The left one corresponds to the trend of E N ½ANOSðb p0 Þ for the
charts with the CCC chart and the right one without the CCC chart. The values of E N ½ANOSðb p0 Þ and SDN ½ANOSðb p 0 Þ
for the CCC‐4 chart when the process is out of control are shown in Table 2.
The trends of E N ½ANOSðb p0 Þ and SDN ½ANOSðb p0 Þ in Figures 3 and 4 indicate that, when the sample size m or the
out‐of‐control nonconforming rate p is large, the out‐of‐control performance ( EN ½ANOSðb p0 Þ) for the CCC‐r charts
can almost achieve the parameter‐known results while the SDN ½ANOSðb p0 Þ values are small. However, even a large sam-
ple size still results in large values of SDN ½ANOSðb p0 Þ when there is a small upward shift in p0. It should be noted that
the SDANOS metric does not seem as powerful, particularly when there is a large shift in the process as the SDANOS for
all charts is very small. Considering the recommendation of Lee et al,18 we would like to suggest using SDN ½ANOSðb p 0 Þ
value to be 10% of E N ½ANOSðb p0 Þ value in practice in order to make full use of the SDANOS in the out‐of‐control cases.
From Figures 3 and 4, for different values of p, the E N ½ANOSðb p0 Þ and SDN ½ANOSðb p0 Þ values of the CCC chart are
larger than all those values of CCC‐r (r > 1). Thus, it is not a good choice to select the CCC chart. Furthermore, for small
shifts of p, the CCC‐r charts with larger r value (5 ≤ r ≤ 7) achieve quite good performance for the EN ½ANOSðb p0 Þ values

N þ1
FIGURE 3 Values of EN ½ANOSðbp0 Þ with the in‐control ANOS0 = 100 000, p0 = 0.001, and b
p0 ¼ : A, p = 0.002; B, p = 0.003; C,
m þ 1 þ 999
p = 0.004; D, p = 0.005; E, p = 0.006; F, p = 0.007 [Colour figure can be viewed at wileyonlinelibrary.com]
ZHANG ET AL. 953

N þ1
FIGURE 4 Values of SDN ½ANOSðb p0 Þ with the in‐control ANOS0 = 100 000, p0 = 0.001, and b
p0 ¼ : A, p = 0.002; B, p = 0.003;
m þ 1 þ 999
C, p = 0.004; D, p = 0.005; E, p = 0.006; F, p = 0.007 [Colour figure can be viewed at wileyonlinelibrary.com]

TABLE 2 Out‐of‐control values of EN ½ANOSðb


p0 Þ and SDN ½ANOSðb
p0 Þ for the CCC‐4 chart with the in‐control ANOS0 = 100 000,
N þ1
p0 = 0.001, and b
p0 ¼
m þ 1 þ 999
m p = 0.002 p = 0.003 p = 0.004 p = 0.005 p = 0.006 p = 0.007

50 000 8267 2580 1361 922 711 588


2352 499 166 66 28 12
100 000 8130 2553 1351 917 708 586
1665 354 118 47 20 8
200 000 8055 2538 1345 914 706 585
1176 251 84 33 14 6
500 000 8008 2529 1342 912 705 585
740 158 53 21 9 4
1 000 000 7995 2526 1341 912 705 585
521 111 37 15 6 3
2 000 000 7988 2525 1341 912 705 585
370 79 26 10 4 2
5 000 000 7983 2524 1340 911 705 585
230 49 16 6 3 1
∞ 7983 2524 1340 911 705 585
0 0 0 0 0 0

are smaller (see Figure 3A,B). On the other hand, for large shifts of p, the E N ½ANOSðb p0 Þ values of CCC‐2 chart and
CCC‐3 chart decrease and become more effective (see Figure 3D‐F). The CCC‐4 chart, in general, performs relatively
well for all cases of shifts. In practice, it is usually difficult to predict the process shift sizes.27,30 Meanwhile, the
CCC‐4 chart is relatively robust to parameter estimation when the high‐quality process is in control. Thus, we recom-
mend choosing CCC‐4 chart in practice for its good performance in detecting process shifts as well as its lower cost of
construction.7
954 ZHANG ET AL.

5 | PERFORMANCE C OMPARISON BETWEEN GEOMETRIC CUSUM


C H A R T , C C C‐ 4 CHART, A ND SYNTHETIC C HART

CCC‐r chart outperforms the Geometric CUSUM chart (g = 5) on condition that p0 or p shifts are not that large.7 As the
estimated parameters will affect the performance of the charts and the performance of Geometric CUSUM chart can be
varied when Geometric CUSUM chart is designed to quickly detect different shifts, Geometric CUSUM chart with dif-
ferent g values as well as p0 or p is not that large but still needs discussion especially when parameters are unknown.14
And the performances of Geometric CUSUM chart and CCC‐4 chart with estimated parameters are compared here.
Besides, we also take synthetic chart with optimal sample size n = 95 and acceptance number c = 1 into comparison
for its effectiveness in detecting p shifts.
The in‐control values of E N ½ANOSðb p0 Þ and SDN ½ANOSðb p0 Þ for the Geometric CUSUM chart with different g values
(1.5 ≤ g ≤ 7), the CCC‐4 chart, and the synthetic chart when the desired ANOS0 = 100 000 and p0 = 0.001 are shown in
Table 3. For the Geometric CUSUM chart with a fixed sample size m, both E N ½ANOSðb p0 Þ and SDN ½ANOSðb p0 Þ values
decrease with the increase of g value (1.5 ≤ g ≤ 4). Then both E N ½ANOSðb p0 Þ and SDN ½ANOSðb p0 Þ values increase when
g grows (4 ≤ g ≤ 7). For instance, a sample size m that is a little smaller than 2 000 000 would be appropriate for the
Geometric CUSUM charts with g = 4 to achieve the SDN ½ANOSðb p0 Þ value to be 10% of the E N ½ANOSðb p0 Þ value. How-
ever, the Geometric CUSUM charts with other values of g (g ≠ 4) in Table 3 will need a sample size much larger than
2 000 000 if the practitioner wants to achieve a similar level of parameter estimation effects. For the CCC‐4 chart, the
E N ½ANOSðb p0 Þ values are the closest to the desired ANOS0 and SDN ½ANOSðb p0 Þ values are always the smallest when
compared with other charts shown in Table 3. It is obvious that CCC‐4 chart is the least sensitive to parameter estima-
tion. The synthetic chart is the second least sensitive to parameter estimation. For example, a sample size m = 1 000 000
achieves SDN ½ANOSðb p0 Þ = 9073 to be 9.1% of E N ½ANOSðb p0 Þ = 100 286 for the CCC‐4 chart while m = 1 000 000
achieves SDN ½ANOSðb p0 Þ = 11 918 to be 11.8% of E N ½ANOSðb p0 Þ = 100 436 for the synthetic chart. Meanwhile,
m = 2 000 000 achieves SDN ½ANOSðb p0 Þ = 9439 to be 9.4% of E N ½ANOSðb p0 Þ = 100 396 for the Geometric CUSUM chart
with g = 4 compared with SDN ½ANOSðb p0 Þ = 8549 to be 8.54% of E N ½ANOSðb p0 Þ = 100 132 for the synthetic chart.
Figures 5 and 6 show E N ½ANOSðb p0 Þ and SDN ½ANOSðb p0 Þ values for the Geometric CUSUM chart with different g
values (1.5 ≤ g ≤ 7), the CCC‐4 chart, and the synthetic chart when p = 0.002, 0.003, 0.004, 0.005, 0.006, and 0.007 with
the in‐control ANOS0 = 100 000 and p0 = 0.001. For the Geometric CUSUM chart, it can be designed to quickly detect a
certain upward shift in p. For instance, the Geometric CUSUM chart with g = 3 has the best performance when detect-
ing p = 0.003 (see Figure 5B). But the Geometric CUSUM chart with g = 3 will not be the best when detecting other
TABLE 3 Values of E N ½ANOSðb p0 Þ and SDN ½ANOSðb
p0 Þ for the geometric CUSUM chart (g = 1.5 to 7), the CCC‐4 chart, and the synthetic
N þ1
chart (n = 95, c = 1) with the in‐control ANOS0 = 100 000, p0 = 0.001, and b p0 ¼
m þ 1 þ 999

Geometric CUSUM Chart


CCC‐4 Synthetic
m g = 1.5 g=2 g=3 g=4 g=5 g=6 g=7 Chart Chart

50 000 180 433 153 193 122 476 114 893 126 820 121 812 122 316 106 129 110 863
512 808 264 320 89 358 68 129 128 283 106 840 104 354 41 382 55 837
100 000 132 921 123 470 110 913 107 788 111 945 109 961 110 646 103 164 105 686
139 353 102 216 54 532 45 314 63 552 56 814 57 644 29 171 38 340
200 000 114 642 110 982 105 354 103 952 106 233 105 782 105 565 101 570 102 848
66 337 55 264 35 723 30 935 39 190 37 528 36 899 20 546 26 307
500 000 105 480 104 316 102 121 101 614 102 297 102 354 102 036 100 564 101 020
34 005 30 386 21 550 19 109 22 230 22 024 21 543 12 901 16 746
1 000 000 102 686 102 212 101 054 100 802 101 111 101 128 101 251 100 286 100 436
22 741 20 434 15 001 13 401 15 298 15 009 14 818 9 073 11 918
2 000 000 101 257 101 072 100 527 100 396 100 398 100 442 100 536 100 134 100 132
15 572 14 076 10 526 9 439 10 487 10 432 10 671 6455 8549
5 000 000 100 686 100 351 100 211 100 162 100 457 100 265 100 880 100 028 100 042
9756 8672 6629 5953 6864 6710 7108 4002 5892
∞ 99 995 100 022 99 897 100 056 100 278 100 368 100 032 100 046 100 006
0 0 0 0 0 0 0 0 0
ZHANG ET AL. 955

FIGURE 5 Values of EN ½ANOSðbp0 Þ for the Geometric CUSUM chart (g = 1.5 to 7), the CCC‐4 chart, and the synthetic chart (n = 95,
N þ1
c = 1) with the in‐control ANOS0 = 100 000, p0 = 0.001, and b
p0 ¼ [Colour figure can be viewed at wileyonlinelibrary.com]
m þ 1 þ 999

FIGURE 6 Values of SDN ½ANOSðbp0 Þ for the Geometric CUSUM chart (g = 1.5 to 7), the CCC‐4 chart, and the synthetic chart (n = 95,
N þ1
c = 1) with the in‐control ANOS0 = 100 000, p0 = 0.001, and b
p0 ¼ [Colour figure can be viewed at wileyonlinelibrary.com]
m þ 1 þ 999

shifts in p (see Figure 5). With the increase of upward shifts in p, the synthetic chart becomes more effective in detecting
shifts. As for the CCC‐4 chart, when considering detecting different shifts in p, it always has very good performance and
sometimes even the best among charts in Figure 5 when 0.004 ≤ p ≤ 0.005. Meanwhile, the trend of SDN ½ANOSðb p0 Þ for
the CCC‐4 chart becomes more stable with the increase of p value (see Figure 6).
Consequently, the CCC‐4 chart can be always very effective in detecting different shifts in p (0.002 ≤ p ≤ 0.007)
while to be less sensitive to parameter estimation when compared with Geometric CUSUM chart and synthetic chart.
956 ZHANG ET AL.

By considering the difficulty of determining the kG and hG for Geometric CUSUM chart as well as n, c, and L for syn-
thetic chart,7 we recommend the CCC‐4 chart to be used in practice as a better choice here.

6 | EXAMPLES

In this section, we will provide an example of application to explain the implementation of the CCC‐4 chart when
parameters are estimated with the in‐control ANOS0 = 100 000 and p0 = 0.001 for different sample sizes. In Table 4,
for a fixed sample size m, several numbers of nonconforming items N are chosen to estimate the CCC‐4 chart's control
limit LCL. The probability of the occurrence of such N values in practice is also provided in Table 4. When N is 0.001
times of m, the process parameters will be well estimated (bp0;B ¼ p0 ¼ 0:001). For example, when m = 1 000 000 and
N = 1000, the estimated lower control limit is equal to the parameter known case (LCL b ¼ LCL ¼ 1; 270). To illustrate
b and the true value of LCL is also
the estimation error in practice, the difference between the estimated control limit LCL
shown in Table 4. With an increase in the m value, LCL b becomes more converging to the LCL with known parameter.

TABLE 4 An example of application for the CCC‐4 chart when parameters are estimated with the in‐control ANOS0 = 100 000, p0 = 0.001,
N þ1
and b
p0 ¼
m þ 1 þ 999
m N Probability b
LCL LCL b
LCL‐LCL

50 000 30 0.000675 2470 1270 1200


40 0.021488 1699 1270 429
50 0.056353 1270 1270 0
60 0.020097 1002 1270 −268
70 0.001359 819 1270 −451
100 000 80 0.005190 1703 1270 433
90 0.025038 1459 1270 189
100 0.039881 1270 1270 0
110 0.023424 1121 1270 −149
120 0.005553 999 1270 −271
200 000 180 0.010551 1460 1270 190
190 0.022439 1359 1270 89
200 0.028212 1270 1270 0
210 0.021525 1191 1270 −79
220 0.010206 1119 1270 −151
500 000 480 0.012139 1340 1270 70
490 0.016300 1305 1270 35
500 0.017847 1270 1270 0
510 0.015999 1237 1270 −33
520 0.011788 1205 1270 −65
1 000 000 980 0.010422 1304 1270 34
990 0.012063 1287 1270 17
1000 0.012621 1270 1270 0
1010 0.011947 1254 1270 −16
1020 0.010243 1236 1270 −34
2 000 000 1980 0.008113 1286 1270 16
1990 0.008726 1279 1270 9
2000 0.008925 1270 1270 0
2010 0.008683 1262 1270 −8
2020 0.008037 1253 1270 −17
5 000 000 4980 0.005434 1276 1270 6
4990 0.005594 1274 1270 4
5000 0.005645 1270 1270 0
5010 0.005583 1267 1270 −3
5020 0.005413 1262 1270 −8
ZHANG ET AL. 957

b for the CCC‐4 chart when m = 5 000 000 is much closer to 1270 when compared with the circum-
For instance, LCL
stance when m = 100 000. Thus, we recommend the use of a large enough sample size to achieve the desired in‐control
performance. In practice, the ANOS and SDANOS values in Tables 1, 2 can be used to determine the minimum sample
size in phase I required for estimating the control limit to meet the performance requirements.

7 | CONCLUSION

CCC‐r charts have been recommended for being effective in monitoring a high‐quality process. We investigate the
effects of parameter estimation on the CCC‐r charts in this study. The CCC‐r chart with a larger r is more sensitive
to the effects of parameter estimation when the process is in control. When comparing the out‐of‐control performance,
the performance of the CCC‐r charts will converge when detecting increase in p. We propose the CCC‐4 chart for its
effectiveness in monitoring different process shifts and its easier construction in practice when compared with Geomet-
ric CUSUM chart and synthetic chart. Furthermore, CCC‐4 chart is less sensitive to parameter estimation while being
tuned more quickly to different shifts in p when compared with Geometric CUSUM chart.
Moreover, there may be some future research points on this topic. Firstly, other objective functions such as average
number of defectives (AND)27 may be used to measure the overall performance of attribute charts such as CCC‐r charts
for high‐quality processes. Secondly, we only obtained the ANOS values for different out‐of‐control p values to illustrate
the effect of estimated parameters. It is also very beneficial to investigate the charts performance with a specific distri-
bution of the out‐of‐control nonconforming rate, for example, Rayleigh distribution,31 when process parameters are esti-
mated. Finally, upward shifts under a steady‐state random shift model may be more realistic than the zero‐state model.
Future research is warranted to extend influence of estimated parameters on steady‐state random shift models.

A C K N O WL E D G M E N T S
This research work was supported by National Key R&D Program of China (no. 2016QY02D0301 and National Natural
Science Foundation of China (nos 71872123, 7161101136, and 71472132).

R EF E RE N C E S
1. Toubia‐Stucky G, Liao H, Twomey J. A sequential Bayesian cumulative conformance count approach to deterioration detection in high
yield processes. Qual Reliab Eng Int. 2012;28(2):203‐214.
2. Xie M, Goh TN, Kuralmani V. Statistical models and control charts for high‐quality processes. New York: Springer Science & Business
Media, LLC; 2012.
3. Chang TC, Gan FF. Modified Shewhart charts for high yield processes. J Appl Stat. 2007;34(7):857‐877.
4. Steiner SH, MacKay RJ. Effective monitoring of processes with parts per million defective. A hard problem! In: Lenz HJ, Wilrich PT, eds.
Frontiers in Statistical Quality Control 7. Heidelberg: Physica; 2004:140‐149.
5. Szarka JL III, Woodall WH. On the equivalence of the Bernoulli and geometric CUSUM charts. J Qual Technol. 2012;44(1):54‐62.
6. Szarka JL, Woodall WH. A review and perspective on surveillance of Bernoulli processes. Qual Reliab Eng Int. 2011;27(6):735‐752.
7. Wu Z, Yeo SH, Fan H. A comparative study of the CRL‐type control charts. Qual Reliab Eng Int. 2000;16(4):269‐279.
8. Chen YK. Cumulative conformance count charts with variable sampling intervals for correlated samples. Comput Ind Eng.
2013;64(1):302‐308.
9. Noorossana R, Saghaei A, Paynabar K, Samimi Y. On the conditional decision procedure for high yield processes. Comput Ind Eng.
2007;53(3):469‐477.
10. Bourke PD. Detecting a shift in fraction nonconforming using run‐length control charts with 100% inspection. J Qual Technol.
1991;23(3):225‐238.
11. Ohta H, Kusukawa E, Rahim A. A CCC‐r chart for high‐yield processes. Qual Reliab Eng Int. 2001;17(6):439‐446.
12. Xie M, Goh TN, Lu XS. Computer‐aided statistical monitoring of automated manufacturing process. J Comput Ind Eng. 1998;35(1–
2):189‐192.
13. Jones‐Farmer LA, Woodall WH, Steiner SH, Champ CW. An overview of phase I analysis for process improvement and monitoring. J
Qual Technol. 2014;46(3):265‐280.
14. Yang Z, Xie M, Kuralmani V, Tsui LK. On the performance of geometric charts with estimated control limits. Qual Control Appl Stat.
2003;48:481‐484.
958 ZHANG ET AL.

15. Fallahnezhad MS, Golbafian V. Determining the sample size for estimation of the CCC‐R control chart parameters based on estimation
costs. Int J Eng. 2017;30:253‐260.
16. Woodall WH, Montgomery DC. Research issues and ideas in statistical process control. J Qual Technol. 1999;31(4):376‐387.
17. Zhang M, Peng Y, Schuh A, Megahed FM, Woodall WH. Geometric charts with estimated control limits. Qual Reliab Eng Int.
2013;29(2):209‐223.
18. Lee J, Wang N, Xu L, Schuh A, Woodall WH. The effect of parameter estimation on upper‐sided Bernoulli cumulative sum charts. Qual
Reliab Eng Int. 2013;29(5):639‐651.
19. Zhang M, Nie G, He Z, Hou X. The Poisson INAR (1) one‐sided EWMA chart with estimated parameters. Int J Prod Res.
2014;52(18):5415‐5431.
20. Woodall WH, Driscoll AR. Some recent results on monitoring the rate of a rare event. In: Knoth S, Schmid W, eds. Frontiers in Statistical
Quality Control 11. Cham: Springer; 2015:15‐27.
21. Reynolds JR, Marion R, Stoumbos ZG. A CUSUM chart for monitoring a proportion when inspecting continuously. J Qual Technol.
1999;31(1):87‐108.
22. Abbasi SA, Riaz M, Miller A. Enhancing the performance of CUSUM scale chart. Comput Ind Eng. 2012;63(2):400‐409.
23. Wu Z, Yeo SH, Spedding TA. A synthetic control chart for detecting fraction nonconforming increases. J Qual Technol.
2001;33(1):104‐111.
24. Haridy S, Wu Z, Khoo MBC, Yu FJ. A combined synthetic and np scheme for detecting increases in fraction nonconforming. Comput Ind
Eng. 2012;62(4):979‐988.
25. Xie M, Lu XS, Goh TN, Chan LY. A quality monitoring and decision‐making scheme for automated production processes. Int J Qual
Reliab Manage. 1999;16(2):148‐157.
26. Bourke PD. The geometric CUSUM chart with sampling inspection for monitoring fraction defective. J Appl Stat. 2001;28(8):951‐972.
27. Haridy S, Wu Z, Lee KM, Bhuiyan N. Optimal average sample number of the SPRT chart for monitoring fraction nonconforming. Eur J
Oper Res. 2013;229(2):411‐421.
28. Bourke PD. Performance comparisons for the synthetic control chart for detecting increases in fraction nonconforming. J Qual Technol.
2008;40(4):461‐475.
29. Dudewicz EJ, Mishra S. Modern mathematical statistics. New York: John Wiley & Sons, Inc; 1988.
30. Shu L, Jiang W, Wu S. A one‐sided EWMA control chart for monitoring process means. Commun Stat Simul Comput. 2007;36(4):901‐920.
31. Haridy S, Wu Z, Chen S, Knoth S. Binomial CUSUM chart with curtailment. Int J Prod Res. 2014;52(15):4646‐4659.

A UT H O R B I O G R A P H I E S
Min Zhang is a professor in College of Management and Economics, Tianjin University, China. She received her
PhD in management science and engineering from Tianjin University, China, in 2006. Her research interests focus
on statistical quality control and six sigma management. She has published more than 30 papers in research
journals, such as International Journal of Production Economics, International Journal of Production Research, and
Quality and Reliability Engineering International.

Xuejun Hou is a graduate student in College of Management and Economics at Tianjin University in China.

Hui Chen is a PhD student in College of Management and Economics at Tianjin University in China. His research
interest focuses on statistical quality control.

Shuguang He is a professor in College of Management and Economics, Tianjin University, China. He received his
PhD in management science and engineering from Tianjin University, China, in 2002. His research interests focus
on warranty data analysis and statistical quality control. He has published more than 30 papers in research journals,
such as Journal of Quality Technology, Annals of Operations Research, and International Journal of Production
Research.

How to cite this article: Zhang M, Hou X, Chen H, He S. CCC‐r charts' performance with estimated parameter
for high‐quality process. Qual Reliab Engng Int. 2019;35:946–958. https://doi.org/10.1002/qre.2438

You might also like