Professional Documents
Culture Documents
Keywords: CUSUM scheme; distance chart; joint monitoring; max chart; parameter estimation
1. Introduction
tatistical process control and monitoring schemes, such as control charts, are widely used in various sectors starting from
S manufacturing industry to several non-manufacturing environments. These schemes are primarily used for monitoring the
process distribution and detecting any changes in the process parameters. Most of the traditional control techniques monitor
a single process parameter, usually the mean or the variance. In recent years, host of researchers advocated in favour of a joint
monitoring scheme for a more efficient process control. The issue of the joint monitoring is particularly important in situations where
a process could be affected by special causes that may result in concurrent shifts in both the mean and variance. Gan et al.1 provided
an interesting example of an improperly fixed stencil in a circuit manufacturing. This may lead to a shift in the mean and variance of
the thickness of the solder paste printed on circuit boards. Similarly, the need for simultaneously comparing the location and scale
arises also out of fields like climate dynamics2, bioinformatics3, supply chain4, maintenance5, finance,6 and medicine7. Interested
readers may refer to Marozzi8 that addressed a wide scope of applications of simultaneous tests of the location and scale. Marozzi8
also comprehensively reviewed and compared several nonparametric statistics for simultaneous tests. Being nonparametric, the
control schemes based on the plotting statistics, as in Marozzi8, have broader applicability. Those statistics can be applied to various
sectors beyond manufacturing industries, as mentioned in the preceding texts, where the process distribution is, in general, non-
normal. Interested readers may also refer to Mukherjee and Marozzi9 for more details. In this paper, we, however, focus only on
the joint monitoring within the traditional Gaussian framework.
There is a plethora of literature on the joint monitoring schemes. We can see broadly two major groups of works. Some works
addressed monitoring the mean and variance jointly on a single chart, such as Chen et al.10, Costa and Rahim11, Wu et al.12, Li
et al.13, Zhang et al.14 and Ou et al.15, and among others. The rest of the works use a more classical approach of combining two
separate charts, one apiece for monitoring the mean and the variance, for example, Gan16, Costa17, He and Grigoryan18, and Lee19.
a
Department of Systems Engineering and Engineering Management, City University of Hong Kong, Hong Kong, China
b
XLRI-Xavier School of Management, Production Operations and Decision Sciences Area, Jamshedpur, India
c
School of Management, State Key Laboratory for Manufacturing Systems Engineering, The Key Lab of the Ministry of Education for Process Control Efficiency
Engineering, Xi’an Jiaotong University, Xi’an, China
*Correspondence to: Qin Su, Xi’an Jiaotong University, School of Management, Xi’an, China.
†
E-mail: qinsu@mail.xjtu.edu.cn
Copyright © 2016 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2016,
C. LI ET AL.
In reality, one-chart schemes are more attractive for both practical and statistical reasons. Joint monitoring based on a single chart
usually has performance advantage, especially when the process parameters are unknown. Further, such an approach typically
requires fewer process control resources and is easier to implement and follow up. Interested readers may see the review literature
by Cheng and Thaga20 that covers early contributions in this area. Recently, McCracken and Chakraborti21 reviewed the new
developments between 2006 and 2012. They also highlighted the current trends in joint monitoring research. For very recent papers
on joint monitoring, readers may refer to, among others, Haridy et al.22, Ou et al.,23 and Mukherjee et al.24.
The majority of the available one-chart schemes for jointly monitoring the process mean and variance are designed with the
presumption that the true parameters are known and the situation is described by the term ‘standards known’ or ‘Case-K’. Nevertheless,
more common and practical situation is that the one or more process parameters are unknown and are estimated from a reference
(Phase-I) sample. The term ‘standards unknown’ or ‘Case-U’ describes such a situation. It is well known that when the parameters are
unknown and are estimated from a reference sample (especially when the sample size is small), the estimation has a great impact on
the in-control (IC) performance metrics, such as the average run length (ARL) and the false alarm rate (FAR) and is more often, detrimental
to the out-of-control (OOC) performance of the charts. It may result in increasing the observed FAR or delaying the detection. In a nicely
written review paper by Jensen et al.25, they thoroughly discussed the effects of estimation of unknown parameters on the performance
of a number of control charts. Specially, for traditional control schemes to monitor the mean and variance simultaneously under Case-U,
there is an emerging concern about the use of certain improper plotting statistics. Therefore, we recommend modifying the Case-K
control charts so that the revised charts always yield a desired nominal IC-ARL (or called ARL0) when a reference sample is used to
estimate the unknown parameters. Further, we suggest amending the control limits based on the corrected plotting statistics.
Recently, McCracken et al.26 rendered an effective solution for simultaneous monitoring of unknown mean and variance of a
normally distributed process. They considered the Shewhart-type control schemes under Case-U by suitably modifying two well-
known schemes for Case-K, namely, the max chart introduced by Chen and Cheng27 and the distance chart proposed by Razmy28.
Both of the max and the distance charts for Case-K are based on the Z and the chi-squared statistics, respectively, used for testing
hypotheses about the mean and the variance. The plotting statistics of the max and distance charts are given by
X i μ 1 ðn 1ÞS2i
Mi ¼ max pffiffiffi ; Φ Fχ 2 ;
σ= n ðn1Þ σ2
and
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2 2
Xi μ 1 ðn 1ÞS2i
Di ¼ p ffiffiffi þ Φ Fχ 2 :
σ= n ðn1Þ σ2
Here the functions Φ(.) and F χ 2 ð:Þ denote, respectively, the cumulative distribution functions (cdf) of the standard normal
ðn1Þ
distribution and the chi-squared distribution with (n 1) degrees of freedom; the parameters μ and σ 2 denote the true process mean
and variance, and X i ands S2i are the mean and the variance of the ith test sample of size n. McCracken et al.26 noted that when both
parameters are unknown and are estimated from an IC reference sample, instead of the Z and the chi-squared statistics, it would be
ideal to combine the Student’s t statistic, suitable for testing a hypothesis concerning μ when σ 2 is unknown, and the F statistic,
recommended for testing a hypothesis concerning σ 2 when μ is unknown. Following this idea, McCracken et al.26 modified the
Shewhart-type max and distance control charts for Case-U. In this context, it is worth mentioning that the joint monitoring procedures
may also be viewed as the online version of combined and multi-aspect testing. Readers may refer to, for example, Pesarin and
Salmaso29 and Marozzi30 for more details on multi-aspect testing.
The Shewhart-type schemes are usually good for detecting large and abrupt shifts in a process. However, in practice, more often the
magnitude of shifts in either or both parameters are small and of persistent nature. It is well known that the CUSUM schemes are superior
to the Shewhart-type schemes in such situations. Therefore, it is an important open problem to study the design and implementation of the
CUSUM-based schemes for joint monitoring of normally distributed processes under Case-U. In this research, we embrace the method by
McCracken et al.26 and propose two modified CUSUM schemes, namely, the max-CUSUM chart and the distance-CUSUM chart (hereafter
referred to as M-CUSUM chart and D-CUSUM chart) for Case-U and study their implementation and performance properties.
The rest of the paper is organized as follows. Section 2 is devoted to the new design of the CUSUM schemes for joint monitoring
under Case-U. The implementation and performance investigation are studied in Section 3. In this section, we present a
comprehensive comparison of the two modified CUSUM charts with the two Shewhart-type charts proposed by McCracken et al.26,
and then further investigate the effect of the reference value of the CUSUM chart on the chart performance as well as the optimal
choice of the reference value with unknown shift sizes. Section 4 provides an illustrative example to offer more practical insights.
Section 5 concludes with some remarks.
Copyright © 2016 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2016,
C. LI ET AL.
of the statistic and corresponding control limits invariably induce additional variability that significantly influences the ARL-based
performance of a control chart. In such cases, use of traditional control schemes designed for Case-K for joint monitoring of the mean
and variance is troublesome as well as unethical. The problem is particularly severe when the size of the reference sample is small and
the test (Phase-II) sample size is relatively large.
McCracken et al.26 showed that when the reference sample size is small (usually 100 or less, as in most practical situations),
adjustments of the plotting statistic and control limits of Case-K charts are an absolute necessity. A large number of reference samples
are rarely available in practice. Therefore, it is better to consider appropriate schemes for Case-U such that the ARL0 will be close to a
nominal value. Otherwise, there might be many false alarms or unnecessary delay in detecting the shifts in the mean and/or variance.
Such phenomena will diminish the practical utility of these charts. This is certainly unwanted for an effective monitoring scheme. To
overcome the drawbacks, we propose two new CUSUM schemes based on the corrected plotting statistic and derive the control limits
for Case-U.
It is well known that the Ūi ’ s and S2Ui ’s are independent with the former following a normal distribution with mean μi and variance
σ 2i
ni, and the later assuming a chi-squared distribution with (ni 1) degrees of freedom. Writing N = n0 + n1, introduce, as in McCracken
et al.26, two statistics that correctly account for the influence of the reference sample on the plotting statistics, namely,
rffiffiffiffiffiffiffiffiffiffi
n0 n1 U1 U0 S2
W1 ¼ and W 2 ¼ U2 1 :
N SU 0 SU 0
McCracken et al.26 noted that given Ū0 and S2U0 , W1 and W2 are mutually independent. Further, W1 follows a normal distribution with
pnffiffiffiffiffiffiffi
0 n1 ð 1
μ U0 Þ σ2 ðn1 1ÞS2U
mean N SU and variance nN0 S21 , and σ2
0
W 2 follows a chi-squared distribution with (n1 1) degrees of freedom.
0 U0 1
Unconditionally, when the process is IC, we have W1 follows a t distribution with (n0 1) degrees of freedom, and W2 follows an F
distribution with (n1 1) and (n0 1) degrees of freedom, but W1 and W2 are not mutually independent.
Write W 1 ¼ Φ1 ζ n0 1 ðW 1 Þ and W 2 ¼ Φ1 F n1 1;n0 1 ðW 2 Þ , where ζυ(.) and F υ1 ;υ2 ð:Þ denote, respectively, the cdf of a t
distribution with υ degrees of freedom and that of an F distribution with υ1 and υ2 degrees of freedom. Following the results as in
pffiffiffiffi
n0 ðU0 μ0 Þ ðn0 1ÞS2
McCracken et al.26, the joint cdf of W 1 and W 2 conditioning on Z U0 ¼ σ0 and V U0 ¼ σ2 U0 is given by
0
ϕ ðzU0 Þf χ 2 ðv U0 Þdz U0 dv U0 :
ðn0 1Þ
pffiffiffiffiffi
where b ¼ ζ 1 1
n0 1 fΦðυ1 Þg, d ¼ F n1 1;n0 1 fΦðυ2 Þg, τ ¼ σ 0 =σ 1 , and δ ¼
2 2
n1 ðμ0 μ1 Þ=σ 0 ; the functions ϕ(.) and f χ 2 ð:Þ denote the
ðn0 1Þ
probability distribution function (pdf) of the standard normal distribution and the chi-squared distribution with (n0 1) degrees of
freedom, respectively.
Notice that when δ = 0 and τ = 1, the joint (conditional and unconditional) cdf will not depend on any nuisance parameters. As a
consequence, the IC run-length distributions of all control charts based on W 1 and W 2 are free from any unknown parameters.
Therefore, W1* and W2* can be easily used for jointly monitoring the unknown process mean and variance. Now, introducing the
max-type and the distance-type operators, we combine W1* and W2*, and obtain two statistics for joint monitoring purposes, namely,
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Ψ ¼ max W ; W and Δ ¼ W *2 þ W *2. Based on these statistics, we develop two (Phase-II) CUSUM control charts for Case-U,
1 2 1 2
Copyright © 2016 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2016,
C. LI ET AL.
say, the modified M-CUSUM chart and the modified D-CUSUM chart. In this context, we use the term ‘modified’ to distinguish these
charts from their classical versions (Case-K). The charting procedures are provided in the following subsection.
X m ¼ ðX 1 ; X 2 ; …; X m Þ:
Step 2. Let Yj,n = (Yj1, Yj2, …, Yjn) be the jth test sample of size n, j = 1, 2, ….
Step 3. Identify the U0’s with the X’s and the U1’s with the Y’s, respectively, and compute the statistics W 1 and W 2 for the jth test
sample, for j = 1, 2, ….
Step 4A. Obtain Ψ μm;n ¼ E Ψ jICÞ ¼ E max W 1 j; jW 2 jICÞ via numerical method.
n o
Step 5A. Calculate the max-type statistic: Ψj ¼ max jW 1j j; jW 2j j for j = 1, 2, ….
h i
Step 6A. Compute the plotting statistic: C Mj ¼ max 0; C Mj1 þ Ψj Ψ μm;n k , with the starting value CM0 = 0. Here, k(≥0) is
called reference value.
Step 7A. Note that Ψ j ≥0 by definition and whatever be the direction of shift, Ψ j will be larger under an OOC set-up. Therefore, only
upper one-sided CUSUM chart is recommended. So, plot CMj against the upper control limit (UCL) HM. Determination of
HM is discussed later in subsequent section.
Step 8A. If CMj exceeds HM, the process is declared OOC at the jth test sample. If not, the process is thought to be IC, and testing
continues to the next test sample.
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ð Þ
Steps 1–3. Same as before.
Step 4B. Obtain Δμm;n ¼ E ðΔjICÞ ¼ E W *1 2 þ W 2*2 jIC via numerical method.
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Step 5B. Calculate the distance-type statistic: Δj ¼ W *1j2 þ W 2j *2 for j = 1, 2, ….
h i
Step 6B. Compute the plotting statistic: C Dj ¼ max 0; C Dj1 þ Δj Δμm;n k , with the starting value CD0 = 0.
Step 7B. Here also, Δj ≥ 0 by definition and whatever be the direction of shift, Δj will be larger under an OOC set-up. Therefore,
plot CDj against the UCL HD. Determination of HD is discussed later in subsequent section.
Step 8B. If CDj exceeds HD, the process is declared OOC at the jth test sample. If not, the process is thought to be IC, and testing
continues to the next test sample.
Note that here we can also use Δ2j ¼ W 1j *2 þ W *2 for j = 1, 2, …, as the distance-type statistic. In particular,
2 2j
* *
E Δ jIC ¼ E W 1 þ W 2 jIC ¼ 2, whatever be m and n. This property can be useful in practice allowing for facilitating charting
2 2
procedures.
2.2.3. Follow-up procedures
If the process is declared OOC at a certain step, it is important to find out whether there is a shift in mean or in variance or in both
parameters. To this end, we may suggest a unified p-value-based approach in the line of McCracken et al.26 and Chowdhury et al.31,32
for both of the CUSUM charts. Note that the CUSUM chart is basically intended for small and/or persistent shift. Therefore, we
recommend to perform a two sample t-test for mean and a F-test for variance based on m Phase-I observations and all the Phase-
II observations drawn till the point of signal and compute respective p-values of the two tests, say, p1, p2. Unlike the testing of
hypothesis where p-value below 0.05 is often considered as an indication of shift, traditionally, control charts operate at a very small
p-value (such as 0.002 for a Shewhart X chart with ARL0 = 500) to protect against FARs.
(i) If p1, p2 < 0.002, a shift in both mean and variance is indicated.
(ii) If both p1, p2 > 0.05, a false alarm is indicated.
(iii) If p1 < 0.002 and p2 > 0.05, a shift in mean is indicated.
(iv) If p2 < 0.002 and p1 > 0.05, a shift in variance is indicated.
(v) If p1 < 0.002 and 0.002 ≤ p2 ≤ 0.05, a major shift in mean with possible associated shift in variance is indicated.
(vi) If p2 < 0.002 and 0.002 ≤ p1 ≤ 0.05, a major shift in variance with possible associated shift in mean is indicated.
(vii) If p1 > 0.05 and 0.002 ≤ p2 ≤ 0.05, that could be an early sign of shift in variance.
Copyright © 2016 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2016,
C. LI ET AL.
(viii) If p2 > 0.05 and 0.002 ≤ p1 ≤ 0.05, that could be an early sign of shift in mean.
(ix) If 0.002 ≤ p1, p2 ≤ 0.05, that could be an early sign of shift in both mean and variance.
In this context, it is worth mentioning that the ordinary t-test is slightly affected when there is a shift in the variance. Similarly, the
power of the F-test is also reduced when there is a shift in the mean. That leads to a multiplicity issue in follow-up. In the current work,
we omit multiplicity issue for simplicity. Further researches to avoid multiplicity issues in follow-up and designing improved joint
monitoring procedures are highly warranted. Some direction to this end is included in concluding remarks.
Table I. The UCL values of the modified M-CUSUM and D-CUSUM charts of various k for various combinations of m and n and a
nominal ARL0 of 500
Proposed M-CUSUM chart Proposed D-CUSUM chart
Copyright © 2016 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2016,
C. LI ET AL.
Table II. Performance comparisons for various OOC cases when μ0 = 0, σ 0 = 1, m = 100, and n = 5
Copyright © 2016 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2016,
C. LI ET AL.
Table III. Performance comparisons for various OOC cases when μ0 = 0, σ 0 = 1, m = 50, and n = 5.
Copyright © 2016 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2016,
C. LI ET AL.
the same basic statistics W1* and W2*. Comparative performance of the schemes for m = 75, 30 and n = 5 are very similar (and
consistent) with the displayed tables, so that we omit the details for brevity. In Tables II and III, each cell shows the ARL and (SDRL)
values. The cells correspond to the best performing chart (with the lowest OOC-ARL) are shaded in tables.
The simulation results reveal that for all these charts in a given situation, the OOC-ARL values and the SDRL values decrease steeply
with the increasing shift in the mean and also with the increasing shift in the variance, which is expected. Specifically, the charts
detect an upward shift in the variance faster than that in the mean, as we see from Tables II and III that the reduction in the ARL
for an increase in the variance when the mean is IC is greater than that for an increase of the same percentage in the mean when
the variance is IC. When there is a large shift in the mean (e.g. μ1 = 3) accompanied by a shift in the variance, we notice an interesting
phenomenon. We find out that in this scenario, the OOC-ARL values increase slightly instead of being reduced, with the increasing
shift in the variance. For example, the OOC-ARL under μ1 = 3 and σ 1 = 2 is found to be marginally higher than the OOC-ARL values
under μ1 = 3 and σ 1 = 1.75. This phenomenon is also seen in McCracken et al.26 and Chowdhury et al.32, although they did not report
it. Besides, it is also observed that the OOC performance is improved by enlarging the reference sample size. This is intuitively
justifiable. When we have more knowledge about the unknown process parameters, it improves the effectiveness of the chart.
Beyond the general overview, from the tables, we observe that the two proposed CUSUM charts perform rather similarly,
though the D-CUSUM chart slightly outperforms the M-CUSUM chart in most cases. The M-CUSUM chart is marginally better in
detecting moderate-to-large shifts in the mean (1.0 to 3.0) accompanied by smaller or no shifts in the variance. For example, when
(m, n, k) = (100, 5, 0.5), the mean shift from 0 to 1.5 and there is no shift in the variance, the ARL of the modified M-CUSUM chart is
1.96, while that of the modified D-CUSUM chart is around 6% larger. However, the modified D-CUSUM chart has a shorter ARL in
the majority of the cases, including small shifts in the mean (less than 0.75) with no shift in the variance and moderate-to-large shifts
in the variance accompanied by any shift in the mean. Thus, in general, the modified D-CUSUM chart may be preferable.
We also see from Tables II and III that for various choices of k (0.25, 0.5, and 1) the two proposed CUSUM charts always display
superior performance as expected, in detecting small-to-moderate shifts in the mean and/or variance, compared with the modified
Shewhart-type charts. Later we further find that by choosing an appropriate k, the CUSUM charts are still able to marginally
outperform the Shewhart-type charts even for the larger shifts. Hence, from a performance standpoint, we highly recommend the
two proposed CUSUM schemes for jointly monitoring unknown mean and variance, instead of the Shewhart-type schemes proposed
Figure 1. OOC-ARL values of the four types of charts for various values of k and various combinations of μ1 and σ 1 when σ 1 = 1.0
Copyright © 2016 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2016,
C. LI ET AL.
by McCracken et al.26. The CUSUM schemes are adaptive and offer very fast detection for targeted shifts. In the next section, we
systematically discuss the empirical design issues regarding the reference value k for optimal performance of the proposed CUSUM
charts.
Figure 2. OOC-ARL values of the four types of charts for various values of k and various combinations of μ1 and σ 1 when σ 1 = 1.5
Copyright © 2016 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2016,
C. LI ET AL.
Figure 3. OOC-ARL values of the four types of charts for various values of k and various combinations of μ1 and σ 1 when σ 1 = 2.0
Table IV hierarchically for a better display of the relationship between the reference value k and the magnitude of shifts in the mean
and variance. Thus, we can see clearly that, for example, k ≤ 0.25 is especially suitable for small shifts, such as μ1 ≤ 0.75 and σ 1 ≤ 1.25.
Indeed, the optimal values of k may not be unique for those very large shifts (especially in the mean), which can easily be seen from
the fourth subfigure (d) in Figures 1–3. We consider the smallest k in such cases. Besides, we notice that the optimal values of k for the
modified D-CUSUM charts are generally larger than that of the modified M-CUSUM charts.
Traditionally, when the specific magnitude of the shifts in the mean and/or variance is assumed, the CUSUM charts can be
optimally designed in terms of (minimum) OOC-ARL. In practice, however, the nature and magnitude of process shifts in the mean
and/or variance are often unknown. The CUSUM charts may perform poorly when the actual shift size is significantly different from
the assumed size. To deal with such a problem, we often combine several CUSUM charts implemented with a different k. However,
this approach adds much complexity to the design and implementation of CUSUM charts.
Ryu et al.33 recommended a method based on the expected weighted run length (EWRL), an ARL-based performance measure, to
optimally design a CUSUM chart to detect a process mean shift of unknown size. The core idea behind the method is to optimize the
expected value of weighted ARLs based on the probability distribution of the shift. A similar consideration can be found in Wu et al.34
and Yang et al.35. Here, we apply the similar method to our problem. We consider two adaptive evaluation criteria, respectively, the
expected relative ARL (ERARL), given as
ARLðμ1 ; σ 1 Þ
ERARL ¼ ∬ f μ1 ðμ1 Þf σ1 ðσ 1 Þdμ1 dσ 1 :
D ARL opt ðμ1 ; σ 1 Þ
‡
and the expected quality loss (EQL),
EQL ¼ ∬ μ21 þ σ 21 1 ARLðμ1 ; σ 1 Þf μ1 ðμ1 Þf σ1 ðσ 1 Þdμ1 dσ 1 :
D
where ARL(μ1, σ 1) is produced by the operational CUSUM chart with a specific k at (μ1, σ 1). In the previously mentioned expressions, D
represents the process shift domain (D ∈ [μ1, σ 1|(μ1,min ≤ μ1 ≤ μ1,max) ∩ (σ 1,min ≤ σ 1 ≤ σ 1,max)]) Correction added on 30 May 2016, after first
‡
Correction added on 30 MAY 2016, after first online publication: “1/AD” in these two equations was deleted
Copyright © 2016 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2016,
C. LI ET AL.
Table IV. The optimal values of k and (the corresponding OOC-ARL values) for the modified M-CUSUM and
D-CUSUM charts for various OOC cases and a nominal ARL0 of 500
online publication: “and AD denotes its area” was deleted. f μ1 ðμ1 Þ and f σ1 ðσ 1 Þ indicate the probability distribution of μ1 and σ 1, respectively.
Moreover, in the expression of ERARL, ARLopt(μ1, σ 1) means the (globally) optimal ARL with the corresponding optimal k at (μ1, σ 1).
More often, it is difficult in practice to estimate the exact distribution of process shifts. Instead, on the basis of the knowledge about
the process, practitioners usually assume certain common form of the distribution of shifts. To this end, the uniform and triangular
distributions are often considered. In this paper, we consider only the uniform distribution of μ1 ∈ U[0.5, 2] and σ 1 ∈ U[1, 2] for
illustration. We present our findings in Table V for m = 100 and n = 5 related to the optimal choice of k. Both performance measures,
ERARL and EQL are displayed.
Table V. The optimal values of k for the modified M-CUSUM and D-CUSUM charts for μ1 ∈ U [0.5, 2] and σ 1 ∈ U [1, 2] when m = 100
and n = 5
The modified M-CUSUM chart The modified D-CUSUM chart
Performance
measure k EWRL value k EWRL value
Copyright © 2016 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2016,
C. LI ET AL.
The two charts look very similar and indicate that the two methods are almost equally efficient. The first signal occurs at the 15th
sample for the modified M-CUSUM chart and at the 16th sample for the modified D-CUSUM chart. As both charts signal, we use the
proposed post-signal diagnostic procedures to follow up. We see that the p-value of two-sample t-test considering Phase-I and the
14th test sample is 0.0155, which gives us an early indication of location shift at sample No. 14. The p-value for two-sample t-test
for Phase-I sample and the 15th test sample is 0.0269, which also indicates a possible location shift could happen. Then, a tectonic
shift in location takes place in sample No. 16 (p-value is 0.0076) and charts detect it. In this context, we observe only a location shift.
We see from Montgomery36, employing the traditional X and R charts, only X chart first signals at the 18th sample. According to the
later investigation of Montgomery36, the process mean probably suffered a shift since the 13th sample. Therefore, it is shown that
both the CUSUM schemes proposed in this paper are superior to the traditional ones. They are detecting a possible shift more quickly
than the traditional schemes. Additionally, here no prior knowledge about the true process parameters is required.
Copyright © 2016 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2016,
C. LI ET AL.
perform similarly, with the modified D-CUSUM chart having a slight advantage in most cases. Furthermore, the proposed CUSUM
schemes are superior to the Shewhart-type schemes proposed in McCracken et al.26 especially in detecting small-to-moderate shifts
in the mean and/or variance. We systematically investigate the empirical design of optimal reference value of the CUSUM schemes
and provide useful recommendations to the practitioners. When the true process parameters are unknown, two proposed CUSUM
schemes can be very useful in practice for jointly monitoring the mean and variance of a normally distributed process.
It will be an interesting future research problem to use a combination of a Welch t-type test, a robust test to detect the shift in
mean under natural heteroscedasticity and a Levene or Pan type test, a robust test to check the shift in variance under possible shift
in mean. This approach may help to address the multiplicity issue. Another potential direction left for further studies is to investigate
adaptive/dynamic joint monitoring schemes when no prior knowledge about the nature and magnitude of process shifts is available.
The attention here was focused on the problem of simultaneously monitoring the two parameters of a normal process. Nevertheless,
in various applications, such as monitoring consumer satisfaction, service quality, or duration of the contact labour strikes, the data are
often non-normal. Therefore, more investigations on the effective designing of joint monitoring schemes under Case-U when the
underlying process distribution is non-normal are highly warranted.
Acknowledgements
The authors are grateful to the anonymous reviewer for providing helpful suggestions and comments for revising this paper. The
work described in this paper was supported by grants from National Natural Science Foundation of China (No. 71371163 and No.
70872091). This work was also funded by MOE (Ministry of Education in China) Project of Humanities and Social Sciences (No.
13YJA630078) and City University of Hong Kong (No. 9380058).
Note that conditionally on X and S2X , the proposed scheme is a standard one-sided CUSUM chart for continuous observations whose
run-length distribution can be traced by Markov chain theory by artificial discretization of the support of C ’j s, that is the interval [0, H].
Suppose, the interval [0, H] is portioned into c subintervals of equal length, say s, where c is a very large positive integer. Further,
h i
assume that the probability density within the lth interval ðl1c ÞH ; lHc for l = 1, 2, … c, is concentrated as a probability mass at ðl0:5c
ÞH
.
n o
ð2c1ÞH
Then, it is legitimate to consider that when the process is IC, the possible values of Cj are given by 0; 2c H
; 3H
2c ; … ; 2c . Let us
adopt the simplified notation scheme of Mukherjee et al.37 and label these transient states as {1, 2, … , c + 1}. Define pij as the
one-step transition probability from state i to state j, and thus given Ū0 and S2U0 , it is easy to see that
(i) p11 ¼ P C k ¼ 0C k1 ¼ 0; U0 ; S2U0 ,
(ii) pi1 ¼ P C k ¼ 0ði2c ÞH < C k1 < ði1c ÞH ; U0 ; S2U0 for i = 2, 3, …, c + 1,
(iii) p1j ¼ P ðj2c ÞH < C k < ðj1c ÞH C k1 ¼ 0; U0 ; S2U0 for j = 2, 3, …, c + 1,
(iv) pij ¼ P ðj2c ÞH < C k < ðj1c ÞH ði2c ÞH < C k1 < ði1c ÞH ; U0 ; S2U0 for i = 2, 3, …, c + 1, and for j = 2, 3, …, c + 1.
Using an almost accurate approximation provided by equation (6.51) of Hawkins and Olwell38 we have the following:
(i) p11 = F(k),
(ii) pi1 = [F(is + 2s + k) + 4F(is + 1.5s + k) + F(is + s + k)]/6 for i = 2, 3, …, c + 1,
(iii) p1j = F(js s + k) F(js 2s + k) for j = 2, 3, …, c + 1,
(iv) pij = [F(js is + s + k) + 4F(js is + 0.5s + k) 4F(js is 0.5s + k) F(js is s + k)]/6 for i = 2, 3, …, c + 1, and for j = 2, 3, …, c + 1.
where F stands for the conditional cumulative probability density of the centred modified max (Ψ Ψ μm;n Þ or distance (Δ Δμm;n )
statistic given Ū0 and S2U0 . The conditional cdf of Ψ and Δ can be derived with the joint distribution of W 1 and W 2 given in
Section 2.1. The conditional cdf of Ψ is as follows:
Copyright © 2016 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2016,
C. LI ET AL.
where b ¼ ζ 1 1 1 1
m1 ðΦðgÞÞ, h ¼ ζ m1 ðΦðgÞÞ, d ¼ F n1;m1 ðΦðgÞÞ, e ¼ F n1;m1 ðΦðgÞÞ; δ and τ are as defined before, and the conditional
cdf of Δ is expressed as,
pffiffiffiffiffiffiffiffiffi
2 2
∞ g v 2
FðgjZ U0 ; V U0 Þ ¼ ∫
∞
∫ f W 1 ;W 2 ðv1 ; v2 jZ U0 ; V U0 Þdv1 dv2 :
pffiffiffiffiffiffiffiffiffi
g v 2
2 2
where f W 1 ;W 2 ðv 1 ; v 2 jZ U0 ; V U0 Þ is the conditional joint pdf of W 1 ; W 2 , which can be obtained by differentiating the conditional joint cdf
F W 1 ;W 2 ðv 1 ; v 2 jZ U0 ; V U0 Þ.
Using pij as the elements, the transition matrix T with a size of (c + 1) × (c + 1) can be established. Now defining Ni as the run-length
n o
ð2c1ÞH
variable with a starting value C0 = s, where sϵ 0; 2c H
; 3H
2c ; … ; 2c and μ̃i = E(Ni|ZU0, VU0) as the conditional ARL for i = 1, 2, …, c + 1,
we have, from the properties of Markov chains,
→
~¼ μ
μ ~1 ; μ ~ cþ1 ’ ¼ ðI T Þ1 1 :
~ 2 ; …; μ
→
where I is an identity matrix and 1 is a vector with all elements equal to one. Thus, the unconditional ARL is given by integrating this
over the joint probability distribution of Ū0 and S2U0 , as follows
∞∞ ∞∞
~ 1
μ ¼ ∫ ∫ EðN1 jzU ; vU ÞdF Z
0 0 U0
ðzU0 ÞdF V U0 ðv U0 Þ ¼ ∫ ∫ μ~1 ϕðzU Þf χ
0 2 ðv U0 Þdz U0 dv U0 :
0 ∞ 0 ∞ ðm1Þ
The expressions for other run-length distribution characteristics, such as the SDRL and the percentiles, can be derived similarly
using properties and results of Markov chains.
References
1. Gan FF, Ting KW, Chang TC. Interval charting schemes for joint monitoring of process mean and variance. Quality and Reliability Engineering
International 2004; 20(4):291–303.
2. Zhang Q, Xu CY, Yang T. Variability of water resource in the Yellow River basin of past 50 years, China. Water Resources Management 2009;
23(6):1157–1170.
3. Neuhauser M, Senske R. The Baumgartner–Weiss–Schindler test for the detection of differentially expressed genes in replicated microarray
experiments. Bionformatics 2004; 20:3553–3564.
4. Shu J, Barton R. Managing supply chain execution: monitoring timeliness and correctness via individualized trace data. Production and Operations
Management 2012; 21(4):715–729.
5. Caballero Morales SO. Economic statistical design of integrated X-bar-S control chart with preventive maintenance and general failure distribution.
PLoS One 2013; 8(3):1–25.
6. Lunde A, Timmermann A. Duration dependence in stock prices: an analysis of bull and bear markets. Journal of Business & Economic Statistics 2004;
22(3):253–273.
7. Rice KL, Rubins JB, Lebahn F, Parenti CM, Duane PG, Kuskowski M, Joseph AM, Niewoehner DE. Withdrawal of chronic systemic corticosteroids in
patients with COPD. American Journal of Respiratory and Critical Care Medicine 2000; 162(1):174–178.
8. Marozzi M. Nonparametric simultaneous tests for location and scale testing: a comparison of several methods. Communication in Statistics -
Simulation and Computation 2013; 42(6):1298–1317.
9. Mukherjee A, Marozzi M. A distribution-free phase-II CUSUM procedure for monitoring service quality. Total Quality Management & Business
Excellence, 2016, DOI:10.1080/14783363.2015.1134266.
10. Chen GM, Cheng SW, Xie HS. Monitoring process mean and variability with one EWMA chart. Journal of Quality Technology 2001; 33(2):223–233.
11. Costa ABF, Rahim MA. Monitoring process mean and variability with one non-central chi-square chart. Journal of Applied Statistics 2004;
31(10):1171–1183.
12. Wu Z, Yang M, Khoo MBC, Yu FJ. Optimization designs and performance comparison of two CUSUM Schemes for monitoring process shifts in mean
and variance. European Journal of Operational Research 2010; 205(1):136–150.
13. Li Z, Zhang J, Wang Z. Self-starting control chart for simultaneously monitoring process mean and variance. International Journal of Production
Research 2010; 48(15):4537–4553.
14. Zhang J, Zou C, Wang Z. A new chart for detecting the process mean and variability. Communications in Statistics - Simulation and Computation
2011; 40(5):728–743.
15. Ou Y, Wu Z, Lee KM, Wu K. An adaptive CUSUM chart with single sample size for monitoring process mean and variance. Quality and Reliability
Engineering International 2013; 29(7):1027–1039.
16. Gan FF. Joint monitoring of process mean and variance using exponentially weighted moving average control charts. Technometrics 1995;
37(4):446–453.
17. Costa ABF. Joint X-bar and R charts with variable parameters. IIE Transactions 1998; 30(6):505–514.
18. He D, Grigoryan A. Joint statistical design of double sampling X-bar and s charts. European Journal of Operational Research 2006; 168(1):122–142.
19. Lee PH. Joint statistical design of X-bar and S charts with combined double sampling and variable sampling interval. European Journal of
Operational Research 2013; 225(2):285–297.
20. Cheng SW, Thaga K. Single variables control charts: an overview. Quality and Reliability Engineering International 2006; 22(7):811–820.
21. McCracken AK, Chakraborti S. Control charts for joint monitoring of mean and variance: an overview. Quality Technology and Quantitative
Management 2013; 10(1):17–36.
22. Haridy S, Wu Z, Lee KM, Rahim MA. An attribute chart for monitoring the process mean and variance. International Journal of Production Research
2014; 52(11):3366–3380.
Copyright © 2016 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2016,
C. LI ET AL.
23. Ou Y, Wu Z, Khoo MBC, Chen N. A rational sequential probability ratio test control chart for monitoring process shifts in mean and variance.
Journal of Statistical Computation and Simulation 2015; 85(9):1765–1781.
24. Mukherjee A, McCracken AK, Chakraborti S. Control charts for simultaneous monitoring of parameters of a shifted exponential distribution.
Journal of Quality Technology 2015; 47(2):176–192.
25. Jensen WA, Jones-Farmer LA, Champ CW, Woodall WH. Effects of parameter estimation on control chart properties: a literature review.
Journal of Quality Technology 2006; 38(4):349–364.
26. McCracken AK, Chakraborti S, Mukherjee A. Control charts for simultaneous monitoring of unknown mean and variance of normally distributed
processes. Journal of Quality Technology 2013; 45(4):360–376.
27. Chen GM, Cheng SW. Max chart: combining X-bar chart and S chart. Statistica Sinica 1998; 8(1):263–271.
28. Razmy AM. Schemes for joint monitoring of process mean and process variance. Master’s Thesis, Department of Statistics and Applied Probability,
NUS. Singapore, 2005.
29. Pesarin F, Salmaso L. Permutation Tests for Complex Data. Wiley: Chichester, 2010.
30. Marozzi M. A combined test for differences in scale based on the interquantile range. Statistical Papers 2012; 53(1):61–72.
31. Chowdhury S, Mukherjee A, Chakraborti S. A new distribution-free control chart for joint monitoring of unknown location and scale parameters of
continuous distributions. Quality and Reliability Engineering International 2014; 30(2):191–204.
32. Chowdhury S, Mukherjee A, Chakraborti S. Distribution-free Phase II CUSUM control chart for joint monitoring of location and scale. Quality and
Reliability Engineering International 2015; 31(1):135–151.
33. Ryu JH, Wan H, Kim S. Optimal design of a CUSUM chart for a mean shift of unknown size. Journal of Quality Technology 2010; 42(3):311–326.
34. Wu Z, Jiao J, Yang M, Liu Y, Wang Z. An enhanced adaptive CUSUM control chart. IIE Transactions 2009; 41(7):642–653.
35. Yang M, Wu Z, Lee KM, Khoo MBC. The X control chart for monitoring process shifts in mean and variance. International Journal of Production
Research 2012; 50(3):893–907.
36. Montgomery DC. Introduction to Statistical Quality Control (6th edn). John Wiley: New York, 2009.
37. Mukherjee A, Graham MA, Chakraborti S. Distribution-free exceedance CUSUM control charts for location. Communications in Statistics - Simulation
and Computation 2013; 42(5):1153–1187.
38. Hawkins DM, Olwell DH. Cumulative Sum Charts and Charting for Quality Improvement. Springer-Verlag: New York, 1998.
Authors' biographies
Chenglong Li is currently a joint PhD candidate in the School of Management, Xi’an Jiaotong University, China, and in the
Department of Systems Engineering and Engineering Management, City University of Hong Kong, Hong Kong. His research interests
are mainly on statistical quality control.
Dr. Amitava Mukherjee is an associate professor in the production, operations and decision sciences area of XLRI Jamshedpur,
School of Management, the oldest business school in India. He is a life member of International Indian Statistical Association, Indian
Science Congress Association, and Calcutta Statistical Association and a member of American Mathematical Society. Dr Mukherjee has
published a number of research areas including sequential analysis, nonparametric inference, statistical process control, and
geostatistics. His current research interests include applied sequential methodology, nonparametric hypothesis testing, and statistical
process control.
Qin Su is a professor and the Vice Dean of the School of Management, Xi’an Jiaotong University, China. She received her PhD from
Xi’an Jiaotong University in 1993. Professor Su is a member of China Association for Quality and a senior member of Chinese
Mechanical Engineering Society. Her research interests include quality management, supply chain management, and industrial
engineering.
Min Xie is currently the Chair Professor of Industrial Engineering at City University of Hong Kong, Hong Kong. He received his PhD
from Linkoping University in Sweden in 1987, specialising in quality and reliability. Since then, he has been active in quality and
reliability research and published numerous papers in this area. He has also authored several books. Professor Xie is a Fellow of IEEE
and an elected member of International Statistical Institute.
Copyright © 2016 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2016,