Sie sind auf Seite 1von 15

This article was downloaded by: [Nanyang Technological University]

On: 18 July 2014, At: 19:36


Publisher: Taylor & Francis
Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer House,
37-41 Mortimer Street, London W1T 3JH, UK

International Journal of Production Research


Publication details, including instructions for authors and subscription information:
http://www.tandfonline.com/loi/tprs20

Binomial CUSUM chart with curtailment


ab a a c
Salah Haridy , Zhang Wu , Songlin Chen & Sven Knoth
a
School of Mechanical and Aerospace Engineering, Nanyang Technological University,
Singapore, Singapore
b
Department of Mechanical and Industrial Engineering, Northeastern University, Boston,
MA, USA
c
Department of Economics and Social Sciences, Helmut Schmidt University, Hamburg,
Germany
Published online: 06 Feb 2014.

To cite this article: Salah Haridy, Zhang Wu, Songlin Chen & Sven Knoth (2014) Binomial CUSUM chart with curtailment,
International Journal of Production Research, 52:15, 4646-4659, DOI: 10.1080/00207543.2014.882026

To link to this article: http://dx.doi.org/10.1080/00207543.2014.882026

PLEASE SCROLL DOWN FOR ARTICLE

Taylor & Francis makes every effort to ensure the accuracy of all the information (the “Content”) contained
in the publications on our platform. However, Taylor & Francis, our agents, and our licensors make no
representations or warranties whatsoever as to the accuracy, completeness, or suitability for any purpose of the
Content. Any opinions and views expressed in this publication are the opinions and views of the authors, and
are not the views of or endorsed by Taylor & Francis. The accuracy of the Content should not be relied upon and
should be independently verified with primary sources of information. Taylor and Francis shall not be liable for
any losses, actions, claims, proceedings, demands, costs, expenses, damages, and other liabilities whatsoever
or howsoever caused arising directly or indirectly in connection with, in relation to or arising out of the use of
the Content.
This article may be used for research, teaching, and private study purposes. Any substantial or systematic
reproduction, redistribution, reselling, loan, sub-licensing, systematic supply, or distribution in any
form to anyone is expressly forbidden. Terms & Conditions of access and use can be found at http://
www.tandfonline.com/page/terms-and-conditions
International Journal of Production Research, 2014
Vol. 52, No. 15, 4646–4659, http://dx.doi.org/10.1080/00207543.2014.882026

Binomial CUSUM chart with curtailment


Salah Haridya,b*, Zhang Wua, Songlin Chena and Sven Knothc
a
School of Mechanical and Aerospace Engineering, Nanyang Technological University, Singapore, Singapore; bDepartment of
Mechanical and Industrial Engineering, Northeastern University, Boston, MA, USA; cDepartment of Economics and Social Sciences,
Helmut Schmidt University, Hamburg, Germany
(Received 15 November 2012; accepted 26 December 2013)

The binomial cumulative sum (CUSUM) chart has been widely used to monitor the fraction nonconforming ( p) of a
process. It is a powerful procedure for detecting small and moderate p shifts. This article proposes a binomial CUSUM
Downloaded by [Nanyang Technological University] at 19:36 18 July 2014

control chart using curtailment technique (Curt_CUSUM chart in short). The new chart is able to improve the overall
detection effectiveness while holding the false alarm rate at a specified level. The results of the comparative studies show
that, on average, the Curt_CUSUM chart is more effective than the CUSUM chart without curtailment by 30%, in terms
of Average Number of Defectives, under different circumstances. The Curt_CUSUM chart can be applied to a 100%
inspection as well as a general random sampling inspection.
Keywords: quality control; statistical process control; control chart; CUSUM chart; curtailment

1. Introduction
Quality control (QC) is nowadays one of the most important activities in modern manufacturing industries and service
sectors. Customers are usually very concerned with the quality of products. Producing good quality to satisfy customers’
requirements is critical to the survival of a company both now and in the future. Essentially, QC covers the monitoring,
examination and improvement of the quality and productivity of a product, service or process.
Statistical process control (SPC) is one of the most popular and important topic in QC. The prominence of SPC in
manufacturing and service industries can be demonstrated by the quality and costs benefits that have been ascribed in
literature to the intervention of SPC (Bounds 1988; Dondero 1991). SPC effectively monitors the process and reports
trouble spots before defective items are produced. It is essential to help avoid damages and serious economic losses.
Early detection of problems through SPC can reduce wasted time and resources. Additionally, production processes can
be streamlined through the identification of bottlenecks, wait times and other sources of delay by use of SPC. SPC has
been adopted and widely employed in different areas, such as manufacturing industries (Castillo 1995; Tannock 2003;
Yu and Liu 2011; Shu, Hsu, and Hu 2012), health care (Albers 2011; Cannon et al. 2012) and service management
(Tao et al. 2012; Wang et al. 2012).
The control chart is one of the most powerful techniques in SPC. No tool can capture the voice of a process better
than the control chart. It is an effective tool to monitor the process, reduce variation, improve productivity and ensure
quality. Recently, many new charts and SPC techniques have been proposed (Chen et al. 2012; Haridy, Wu, and Flaig
2012; Yang et al. 2012; Knoth and Steinmetz 2013). The binomial cumulative sum (CUSUM) chart is one of the com-
monly used charts for attributes. It is used to monitor the fraction nonconforming p of a process. The fraction noncon-
forming p is defined as the ratio of the number of non-conforming units in a population to the total number of units in
that population (Montgomery 2009). The widespread applications of the binomial CUSUM chart and other attribute
charts are attributable to several factors, such as the simplicity of handling attribute quality characteristics, the capability
of checking multiple quality requirements, the ease to communicate between people at different levels and the
prevalence of count data in many industries and non-manufacturing sectors.
The CUSUM chart is an effective tool for detecting small and moderate process shifts. The first CUSUM chart was
introduced by Page (1954) for monitoring the mean of a quality characteristic in a production process. This chart
incorporates the cumulative information from all samples up to the last one (Lucas 1985a). A statistic Ct is updated and
plotted for the tth sample in a binomial CUSUM chart for detecting upward p shifts.

*Corresponding author. Email: salah@pmail.ntu.edu.sg

© 2014 Taylor & Francis


International Journal of Production Research 4647

C0 ¼ 0
(1)
Ct ¼ maxð0; Ct1 þ dt  kÞ
where dt is the number of non-conforming units found in the tth sample and k is a reference parameter. When an
increasing p shift occurs, Ct tends to become larger and larger. Sooner or later, a sample point will make Ct exceed the
control limit H of the CUSUM chart, and thereby, an out-of-control signal is produced.
CUSUM charts have been increasingly recognised across industries for variable and attribute SPC applications (Lim
et al. 2002; Wu and Tian 2005; Perry and Pignatiello 2011; Lee et al. 2012). Their popularity is mainly attributed to the
fact that they are sensitive for detecting small and moderate shifts in many different SPC applications. Kemp (1962)
described the use of the CUSUM chart for controlling the percentage of defective items being produced. Lucas (1985b)
detailed the design and implementation of the CUSUM chart for attributes. Lucas (1989) also studied the performance of
the binomial CUSUM chart when the defect level is very low. Gan (1993) further developed an optimal algorithm to min-
imise the ATS of the binomial CUSUM chart at a particular p shift value. Radaelli (1994) examined the Poisson approxi-
mation to the binomial CUSUM chart. Woodall (1997) provided a comprehensive review of the control charts for
attributes including the binomial CUSUM chart. Some researchers compared the performance of the Poisson CUSUM
Downloaded by [Nanyang Technological University] at 19:36 18 July 2014

chart with other charts such as the c chart (White, Keats, and Stanley 1997). Bourke (2001) investigated the operating
characteristics of the CUSUM chart under 100% inspection. Wu, Jiao, and Liu (2008) studied a unique feature of the
binomial CUSUM chart which is able to increase the detection effectiveness depending on the size of process shift.
The speed of signalling an out-of-control condition is usually measured by the average time to signal (ATS), that is,
the average time required to signal an out-of-control condition after its occurrence. The smaller the out-of-control ATS
is, the earlier the problem is signalled and the more effective the chart is. Furthermore, the minimisation of ATS also
improves the capability of the control chart to diagnose the assignable cause. Because, the shorter the ATS is, the easier
it is to determine the time when the p shift takes place (Pignatiello and Samuel 2001). In general, the ATS of a control
chart can be calculated by
ATS ¼ h  ARL (2)
where ARL is the average run length, i.e. the average number of samples required to signal an out-of-control condition
and h is the sampling interval.
It is relatively common to base an attribute control chart on 100% inspection of all process output (Montgomery
2009). It is mainly because of the simplicity of the inspection in attributes. In view of this, the discussions in this article
mainly refer to a 100% inspection. Under 100% inspection, there may be no natural divisions of the products into
groups of units that would form the samples for the control charts. Meanwhile, the sampling intervals may be much
smaller than the working shifts. Under such circumstances, the implementation of rational subgrouping may not be
enforced. The samples could be formed by designating n consecutive units as a sample for administrative convenience
or just artificially (Duncan 1986; Reynolds and Stoumbos 1999; Montgomery 2009). In fact, Hawkins and Olwell
(1998) argued that the idea of rational grouping plays no useful part in CUSUM design and use.
Under a 100% inspection, the sampling interval h is equal to the product of the sample size n and the time t
required to inspect a unit. Thus, the ATS of a control chart can be calculated as follows:
ATS ¼ h  ARL ¼ n  t  ARL (3)
In this study, the interval t is always used as the time unit (or in other words, ATS is measured in terms of t). There-
fore, for the 100% inspection,
ATS ¼ n  ARL ¼ n  t  ARL (4)
In this article, the out-of-control ATS is calculated by using the steady-state mode. Under 100% inspection, the
steady-state mode implies that the process starts and stays in an in-control condition for a long time and then a process
shift occurs at some random time within a sample.
The sample size n is a critical factor to ATS. Traditionally, n is rarely determined analytically. Many researchers
(British Standard Institution 1985; Saniga, Davis, and Mcwilliams 1995) have proposed approaches for the determina-
tion of the sample size of the attribute charts. However, a sample size determined by these methods is unlikely to result
in an ATS that is close to the minimum.
It is well known that a small sample size will make the chart less sensitive to small process shifts and result in a
large ARL, and vice versa. Hence, a change of n in either direction will cause one of the two factors (n and ARL in
Equation (4)) of ATS to increase and the other to decrease. Logically, there must be an optimal value of n that minimis-
es ATS or maximises the detection effectiveness of the control chart.
4648 S. Haridy et al.

It is well known in SPC that the performance of the control chart can be enhanced by the incorporation of different
features or by using various techniques. The basic binomial CUSUM chart for attributes was firstly developed by Lucas
(1985a). In the last century, many researchers put a lot of effort into improving this CUSUM chart by different
approaches and from diverse perspectives. For instance, Gan (1993) proposed a design algorithm for optimising the
binomial CUSUM chart. Wu, Jiao, and Liu (2008) introduced a unique feature of the binomial CUSUM chart in which
the difference (dt − d0) is replaced by (dt − d0)w in the formulation of the CUSUM. Mousavi and Reynolds (2009) devel-
oped a Markov binary CUSUM chart based on a log-likelihood-ratio statistic. Recently, Haridy, Wu, Yu, et al. (2013)
presented an optimisation algorithm for the design of the combined CUSUM chart and np chart. All these approaches
resulted in a considerable enhancement in the performance of the binomial CUSUM chart and each of these papers has
its distinctive contribution to the improvement of the basic binomial CUSUM chart in SPC literature.
This article proposes a binomial CUSUM chart with curtailment (Curt_CUSUM chart) to monitor the fraction non-
conforming p. The curtailment technique is widely used in acceptance sampling plans in which the inspection of a sam-
ple is terminated and the associated lot is rejected as soon as the number of observed non-conforming units exceeds an
acceptance number (Montgomery 2009). The curtailment can result in a significant reduction in average sample number
in an acceptance sampling plan.
Downloaded by [Nanyang Technological University] at 19:36 18 July 2014

Wu, Luo, and Zhang (2006) incorporated the curtailment technique into the simple Shewhart np chart. The curtail-
ment feature led to a substantial improvement in the detection effectiveness of the np chart. Specifically, the np chart
with curtailment was able to reduce the out-of-control ATS by nearly half, on average, compared to the conventional np
chart without using curtailment. Usually, the successful employment of a new feature in a simple chart (such as the np
chart) will stimulate researchers to further study the incorporation and the influence of this feature on more advanced
charts (such as the CUSUM chart). This has a practical significance, especially with the increasing recognition of the
sophisticated CUSUM and Exponentially Weighted Moving Average (EWMA) charts across industries in today’s SPC
applications (Shu, Jiang, and Tsui 2008).
In literature, there are many studies in which a particular feature was applied to different control charts. For example,
Epprecht, Costa, and Mendes (2003) used the Variable Sampling Intervals (VSI) feature to enhance the effectiveness of the
np chart. Epprecht, Simões, and Mendes (2010) also applied the same feature (VSI) to improve the performance of the
EWMA chart for attributes. Duncan (1956) developed an algorithm for the economic design of the Shewhart X-bar chart.
Serel and Moskowitz (2008) used the same algorithm to optimise the EWMA chart. However, there is no guarantee that
such research effort will always achieve satisfactory results. Sometimes, a certain feature may achieve a significant improve-
ment for only a specific type of control charts. For instance, while the runs rules significantly enhance the performance of
the X-bar and np charts, they only slightly improve the effectiveness of the EWMA and CUSUM charts. Meanwhile, while
the synthetic algorithm is very effective for the np and X-bar charts, it is not so for the EWMA and CUSUM charts.
The performance of the Curt_CUSUM chart will be compared with that of the conventional CUSUM chart (i.e. the
binomial CUSUM chart without using curtailment) in a systematic and analytical manner. Both charts are designed by
an optimal algorithm in which the objective function to be minimised is an overall performance index, the Average
Number of Defectives or AND in short (Haridy, Wu, Lee, et al. 2013). Minimising AND leads to an immediate reduc-
tion of non-conforming units and, therefore, directly benefits the industry. The results of the comparative studies show
that the Curt_CUSUM chart always has a better overall performance than the conventional CUSUM chart. The high
effectiveness of the Curt_CUSUM chart is mainly attributable to the curtailment technique.
In this article, it is assumed that the random number d (the number of non-conforming units found in a sample)
follows a binomial distribution with known in-control fraction nonconforming p0. Since the control charts for attributes
are most often used to detect an increase in fraction non-conforming or deterioration in quality (Lucas 1985a; Reynolds
and Stoumbos 1999), the focus of this research is to detect increasing p shifts. A decreasing p shift may be caused by a
real improvement in product quality, or a mistake made by the operator, or a malfunction of the instrument. It does not
do harm to the quality of the product directly.
The remainder of the article proceeds as follows. Firstly, the implementation of the Curt_CUSUM chart is presented.
Secondly, the objective function AND is introduced and the optimal design of the Curt_CUSUM chart is detailed. Then,
a comparative study is conducted for a general case, and then further through a factorial experiment. Subsequently, an
illustrative example is provided. The conclusions and discussions are drawn in the last section.

2. Implementation of the Curt_CUSUM chart


In a similar way to the conventional CUSUM charts, a Curt_CUSUM chart for detecting increasing p shifts is
characterised by the sample size n, reference parameter k and upper control limit H. The determination of n, k and H
will be discussed in the next section.
International Journal of Production Research 4649

For a conventional CUSUM chart, a decision regarding the process status cannot be made until all the n units in a
sample have been inspected. On the contrary, a Curt_CUSUM chart is able to detect an out-of-control status before
having inspected all the n units in a sample. The statistic Ct−1 for the (t − 1)th sample of a Curt_CUSUM chart can be
found by Equation (1), after this sample is processed. Then, let gti be the count of non-conforming units i found in the
tth sample. The value of gti will increase one by one from zero during the inspection of the tth sample. Referring to
Equation (1), the process is thought to be out of control when:
Ct1 þ gti  k [ H (5)

or gti [ Gt (6)

where Gt ¼ H þ k  Ct1 (7)


Gt is defined as the threshold of curtailment for the tth sample. This means that, as soon as gti > Gt, the process is
deemed to be out-of-control and no further inspection of the remaining units in the tth sample is needed. Consequently,
an out-of-control signal may be produced earlier. This is the basic role of the curtailment.
Downloaded by [Nanyang Technological University] at 19:36 18 July 2014

Suppose, for a CUSUM chart, n = 100, k = 2 and H = 4. If it is a conventional chart, an out-of-control signal cannot
be produced until all the 100 units have been inspected. On the other hand, for a Curt_CUSUM chart (Ct−1 is assumed
to be equal to 3),
Gt ¼ H þ k  Ct1 ¼ 4 þ 2  3 ¼ 3 (8)
Now, if the 20th unit is identified as the 4th non-conforming unit (i.e. gti = 4 after 20 units have been inspected), an
out-of-control signal can be produced immediately since gti (= 4) becomes larger than the threshold Gt (= 3). In other
words, the out-of-control status is detected after only 20 units have been inspected compared with the sample size (n =
100).
A Curt_CUSUM chart can be implemented as follows:
(1) Initialise the statistic C0 in Equation (1) as zero and set t = 1.
(2) Determine the threshold Gt by Equation (7).
(3) At the beginning of the tth sample, set the count gti of the non-conforming units at zero.
(4) Increase gti by one whenever a non-conforming unit is found. If gti > Gt at any moment, the inspection is
terminated immediately and go to step (8).
(5) Otherwise (i.e. gti ≤ Gt up to the end of the sample), the process is considered to be in control currently and
make dt = gti.
(6) Update Ct by Equation (1)
Ct ¼ maxð0; Ct1 þ dt  kÞ (9)

(7) Increase t by one. Then, go back to step (2) and take the next sample.
(8) An out-of-control signal is produced and the process is stopped for further investigation.
It is clear that curtailment mechanism may detect the out-of-control condition in step (4) before completing the inspec-
tion of the entire sample and therefore, the signalling speed could be expedited. In fact, the curtailment is the distinctive
feature of the Curt_CUSUM chart compared to the conventional CUSUM chart.

3. Design of the Curt_CUSUM chart


3.1 Objective function
When a process shift occurs, the fraction nonconforming p will change to:
p ¼ d  p0 (10)
where δ (1 ≤ δ ≤ δmax) measures the increasing p shift in terms of p0. The process is in control when δ = 1 (i.e. p = p0)
and out of control when (1 < δ ≤ δmax) with a maximum p value at δ = δmax (i.e. pmax = δmax × p0).
Since the goal of the design algorithm is to achieve the best overall performance, the objective function should
measure the detection effectiveness against all possible p shifts within the range of (1 < δ ≤ δmax). In the past, the out-of-
control ATS values at one or a few specified process shifts (δ) are often used as the objective function to be minimised
in the optimal design of a control chart. However, this approach does not guarantee that the resultant control chart will
perform well over a wide range of process shifts.
4650 S. Haridy et al.

A sound measure of the overall effectiveness of an attribute control chart is the AND (Haridy, Wu, Lee, et al.
2013). It is the average number of non-conforming units produced in different out-of-control cases across a range of p
shifts (1 < δ ≤ δmax).
Z dmax
AND ¼ N d  p0  ATSðdÞ  fd ðdÞ dd (11)
1

where N is the number of units produced per time unit and ATS(δ) is the ATS value for a particular shift δ, and fδ(δ) is
the probability density function of δ. In Equation (11), the term (N) may be removed because it is a constant and has no
effect on the performance comparison and the optimal solution.
Z dmax
AND ¼ d  p0  ATSðdÞ  fd ðdÞ dd (12)
1

The integration in Equation (12) can be computed quickly and accurately by a numerical method such as
Legendre–Gauss Quadrature. It is noted that the simplified Equation (12) can be used to design and compare the charts.
But, in order to obtain the actual value of AND, the result of Equation (12) should be multiplied by N.
Downloaded by [Nanyang Technological University] at 19:36 18 July 2014

The index AND (average number of non-conforming units) directly relates the chart performance to the average
number of non-conforming units or the economic outcome. That is to say, a chart that produces smaller AND is thought
to be more economical over different shifts δ. Meanwhile, AND can be considered as a weighted average of ATS that
uses δ as the weight. If AND is used as the objective function to be minimised, then the larger the δ, the smaller the
corresponding ATS(δ) will result from the optimal design. This is justifiable, as a larger δ will lead to a greater loss in
quality and should be detected at a higher speed. Other researchers (Reynolds and Stoumbos 2004) also used a weighted
average of ATS to measure the overall performance in which the quadratic loss is used as the weight.
In any process, the shift δ in fraction nonconforming p is assumed to follow a probability distribution. However, it
is usually unknown and therefore one may have to design the control charts based on an assumed distribution over the
shift domain. In this article, it is assumed that the random shift δ follows a Rayleigh distribution (Haridy, Wu, Lee,
et al. 2013). This distribution is often used to characterise the positional deviation from a target in geometrical tolerance.
It was also adopted to model the mean shift of a normally distributed random variable (Wu, Xie, and Tian 2002). The
Rayleigh distribution is skewed to the right (i.e. δ tends to cluster to the lower end) and looks like a reasonable delegate
of the distributions of many process shifts. If the Rayleigh distribution is used, the probability density function in
Equation (12) is as follows:
!
pðd  1Þ pðd  1Þ2
fd ðdÞ ¼ exp  (13)
2ðld  1Þ2 4ðld  1Þ2
As shown, fδ(δ) is characterised by a single parameter μδ (the mean of δ). Figure 1 displays the probability density
functions of three Rayleigh distributions with different μδ values. The cumulative distribution function of the Rayleigh
distribution is
!
pðd  1Þ2
Fd ðdÞ ¼ 1  exp  (14)
4ðld  1Þ2

The value of μδ can be estimated from the historical data of the out-of-control cases. Suppose an out-of-control
condition is signalled and a follow-up investigation discovers that the fraction nonconforming p during the estimated
fallout duration is ^ pi =p0 is the estimate of the sample shift ^di for this out-of-control case. If m records of ^
pi . Then ^ di are
available, then μδ can be estimated by Pm ^
di
ld ¼ i¼1 (15)
m
The values of δmax in Equation (12) can be determined from Equation (14) so that the probability of (δ > δmax) is
negligible (say < 0.0001).

3.2 Optimal design


To carry out the optimal design of a Curt_CUSUM chart, four specifications have to be given: (1) the minimum
allowable value (τ) of the in-control ATS0, (2) the in-control fraction nonconforming p0, (3) the mean value (μδ) of the
random shift δ and (4) the minimum allowable value (nmin) of the sample size n.
International Journal of Production Research 4651
f

0.4
µ =3

0.3
µ =5

0.2
µ =7

0.1

0
1 3 5 7 9 11 13 15 17 19 21 23

Figure 1. Three Rayleigh probability density functions of δ.


Downloaded by [Nanyang Technological University] at 19:36 18 July 2014

The value of τ is decided by a quality engineer with regard to the false alarm rate. If the cost of handling false alarms is
high, a larger τ should be used in order to reduce the false alarm frequency. However, a large τ may impair the effectiveness
of the control chart in the meantime. The resultant (or actual) in-control ATS0 must be equal to or larger than τ. The value
of p0 is usually estimated from the historical data observed when the process is in control. The value of μδ is estimated from
the historical data of the out-of-control cases, as mentioned in Equation (15). Finally, the value of nmin is determined in
order to accommodate some managerial considerations. It is noteworthy that, in a 100% inspection, varying the sample size
n simply means adjusting the grouping of the inspected units. Since every unit is inspected in the 100% inspection, the
number of inspected units is always equal to the number of produced units. A larger sample size is associated with a smal-
ler number of samples (or less frequent sampling), and vice versa. Suppose the number of units produced in a process per
working day is 10,000. Then, if the sample size n is set as 5, 2000 samples of size 5 will be taken every day. On the other
hand, if n is equal to 1000, then 10 samples of size 1000 will be taken for the same time period.
A very small sample size n of a CUSUM chart may incur some difficulties in implementation. When n is small, it is
quite difficult to handle the frequent sampling because of the troublesome updating of the statistic Ct in Equation (1)
required by each sample. A large sample size will contrarily require less frequent updating of Ct (Bourke 2001). For
instance, if 2000 samples of size 5 have to be taken per day as in the above example, Ct has to be updated 2000 times
per day. This must be mentally exhausting and prone to error, and may encounter strong resistance from operators even
if they only have to enter the number d of non-conforming units of each sample from the keyboard of a computer.
Moreover, there is usually an overhead cost for taking a sample. This may include interrupting a process, walking a dis-
tance, cleaning hands and platform, switching the computer and so on.
The Bernoulli CUSUM chart is a special case of the binomial CUSUM chart in which the sample size n = 1.
Although this minimum possible sample size (n = 1) makes the Bernoulli CUSUM chart effective for detecting large p
shifts (Bourke 2001), it may be neither applicable nor preferable in many applications. Suppose a Bernoulli CUSUM
chart is used to monitor a process in which 20 units are produced per minute. This means that the user has to key in
the reading and update the statistic Ct 20 times per minute. This must be very exhausting and troublesome as aforemen-
tioned even with the help of an on-site computer. A possible approach to overcome this problem is to use an integrated
automatic system which is able to measure and transfer the data directly into a computer. However, the design, installa-
tion and maintenance of such integrated automatic systems are very costly, especially for the vast majority of small and
medium organisations.
On the contrary, if a Curt_CUSUM chart with a sample size of (say, n = 100) is used to monitor the same process,
the threshold Gt needs to be updated only once for every 5 min. In view of this, a minimum allowable sample size nmin
should be specified based on the operational and managerial conditions.
The optimal design of a Curt_CUSUM chart can be formulated by the following model:
Objective: Minimise AND (16)

Constraint: ATS0  s (17)

Design variables: n; k; H
4652 S. Haridy et al.

where the sample size n and reference parameter k are treated as independent design variables. The upper control limit
H is dependent on n, k and the specified τ.
The objective of the optimal design is to identify the optimal values of the independent design variables n and k that
minimise the objective function AND, and meanwhile H is adjusted so that ATS0 ≥ τ. The minimisation of AND will
in turn shorten the ATS values for different values of δ over the p shift range, or reduce the average number of
non-conforming units incurred in the out-of-control cases.
Because of the variability of the simulation results and the discrete nature of attributes, it is quite difficult to make
the value of ATS0 exactly equal to τ. However, ATS0 should be as close as possible to τ.
The optimal design is implemented by a two-level search as outlined below:
(1) Specify the design specifications τ, p0, μδ and nmin.
(2) Initialise a variable ANDmin as a very large number, say 107 (ANDmin is used to store the minimum value of
AND).
(3) At the first level, search the optimal value of n by increasing it from nmin with a step size of one. The search
at this top level will be terminated when AND cannot be further reduced.
(4) At the second level, with the given value of n determined at the first level, search the optimal value of k. For
Downloaded by [Nanyang Technological University] at 19:36 18 July 2014

a given set of values of (n, k),


(i) Determine the control
 limit  H so that the constraint of (ATS0 ≥ τ) is satisfied within a predetermined
tolerance such that ATSs0 s\0:05.
(ii) When the values of all three charting parameters, n, k and H, are preliminarily determined, the
objective function AND is calculated by Equation (12).
(iii) If the calculated AND is smaller than the current ANDmin, replace the latter by the former and the
current values of n, k and H are stored as a temporary optimal solution.
(5) At the end of the entire two-level search, the optimal Curt_CUSUM chart that produces the minimum AND
and satisfies the constraint (ATS0 ≥ τ) is identified. The corresponding optimal values of n, k and H are also
finalised.
The above grid search approach is quite feasible, because one of the two independent design variables, n, is an
integer variable, and therefore all its possible values can be examined. For every given value of n, the optimal value of
the only remaining independent design variable k can be easily determined by a search algorithm for a single variable.
A computer programme has been coded in C language based on the procedure detailed above to carry out the design of
a Curt_CUSUM chart.
In this article, the values of the in-control ATS0 and out-of-control ATS of the conventional CUSUM and
Curt_CUSUM charts are evaluated by simulation. By using a simulation sample size (or the number of iterations) of
100,000, the coefficients of variations of ATS0 and ATS for both the conventional CUSUM chart and the Curt_CUSUM
chart are below 1%.

4. Comparative studies
In this section, the detection effectiveness of two control charts (the conventional CUSUM chart and the Curt_CUSUM
chart) is studied and compared. Both charts are designed by using the design model in Equations (16) and (17). The
design of a conventional CUSUM chart is also to find the best combination of the sample size n, reference parameter k
and upper control limit H so that the chart produces the minimum AND (Equation (12)) and meanwhile has an ATS0
equal to or larger than τ.

4.1 Comparison in a general case


The two charts are first studied under a general case in which the design specifications are as below:

s ¼ 5000; p0 ¼ 0:01; ld ¼ 5; nmin ¼ 50 (18)

4.1.1 Using AND as an objective function


The two charts are designed using AND as the objective function and the results are shown below:
International Journal of Production Research 4653

Conv.CUSUM chart: n ¼ 60; k ¼ 0:950; H ¼ 2:599; AND ¼ 6:7371:

Curt CUSUM chart: n ¼ 62; k ¼ 1:240; H ¼ 1:796; AND ¼ 5:5796:


As mentioned before, the value of δmax can be calculated by Equation (14) based on μδ so that the probability of
(δ > δmax) is below 0.0001 and negligible. Hence, the value of δmax is approximately equal to 12 for this case. The
values of the in-control ATS0 (where δ = 1) and out-of-control ATS (where 1 < δ ≤ 12) of the two charts are calculated
within the process shift range, and the results are displayed in Table 1. The curves of the normalised ATS (i.e. ATS/
ATSCurt_CUSUM) of the two charts are illustrated in Figure 2. It is interesting to observe the following from Table 1 and
Figure 2:
(1) Firstly, each of the two charts generates an ATS0 values larger than but close to τ (constraint (17)) when the
process is in control. This suggests that both charts have nearly identical false alarm rates which provide a
common ground for the comparison.
(2) The Curt_CUSUM chart is always more effective (has smaller ATS values) than the conventional CUSUM
chart over the whole p shift range, except for δ = 2 where the latter marginally outperforms the former. It is
Downloaded by [Nanyang Technological University] at 19:36 18 July 2014

clear that the curtailment feature of the Curt_CUSUM chart makes this scheme very effective from an overall
viewpoint.
(3) It can be observed that the superiority of the Curt_CUSUM chart over the conventional CUSUM chart
increases with the increase in δ. When δ = 12, the ATS value of Curt_CUSUM chart is larger than that of
the conventional CUSUM chart by 68%.

Table 1. ATS values of the two charts using AND as an objective function.

ATS
δ Conv.CUSUM chart Curt_CUSUM chart

1 5013.608 5026.318
2 507.718 516.896
3 232.990 212.022
4 151.714 127.852
5 117.856 91.488
6 96.670 70.961
7 83.098 59.058
8 74.098 49.744
9 67.432 43.318
10 61.528 39.063
11 56.260 34.552
12 52.942 31.451

ATS/ATSCurt_CUSUM
1.7

1.6

1.5

1.4

1.3

1.2

1.1

0.9
0 1 2 3 4 5 6 7 8 9 10 11 12

Figure 2. Normalised ATS of the two charts using AND as an objective function.
4654 S. Haridy et al.

The AND values (Equation (12)) of the two charts are also calculated. The ratio of (ANDCUSUM/ANDCurt_CUSUM) =
6.7371/5.5796 = 1.21. This value indicates that, for this general case, the Curt_CUSUM chart reduces the AND by 21%
compared with the conventional CUSUM chart over the range of p shifts.

4.1.2 Using ATS as an objective function


In this general case, a Curt_CUSUM chart is designed using the out-of-control ATSδ as the objective function to
be minimised and is named as Curt_CUSUMObj_ATS chart. Here, ATSδ is the out-of-control ATS at a specified shift δ
(δ > 1).
The Curt_CUSUMObj_ATS chart is designed under the same specifications in Section 4.1.1 (τ = 5000, p0 = 0.01 and
nmin = 50), with ATSδ being evaluated at δ = 2. The results of the Curt_CUSUM chart (designed in Section 4.1.1) and
the Curt_CUSUMObj_ATS chart are listed below:
Curt CUSUM chart: n ¼ 62; k ¼ 1:240; H ¼ 1:796; ATSd ¼ 516:896; AND ¼ 5:5796:

Curt CUSUMObj chart: n ¼ 63; k ¼ 0:970; H ¼ 3:080; ATSd ¼ 488:59; AND ¼ 5:8698:
Downloaded by [Nanyang Technological University] at 19:36 18 July 2014

ATS

The values of the in-control ATS0 (when δ = 1) and out-of-control ATS (where 1 < δ ≤ 12) of the Curt_CUSUMObj_ATS
chart and Curt_CUSUM chart are shown in Table 2. Figure 3 illustrates the normalised ATS (i.e. ATS/ATSCurt_CUSUM).
The following can be deduced from Table 2 and Figure 3:

Table 2. ATS values of the two charts using ATSδ as an objective function.

ATS
δ Curt_CUSUMObj_ATS chart Curt_CUSUM chart

1 5078.715 5026.318
2 488.587 516.896
3 215.222 212.022
4 134.599 127.852
5 99.916 91.488
6 78.955 70.961
7 66.923 59.058
8 58.162 49.744
9 50.198 43.318
10 44.959 39.063
11 40.529 34.552
12 36.850 31.451

ATS/ATSCurt_CUSUM
1.2

1.15

1.1

1.05

0.95

0.9
0 1 2 3 4 5 6 7 8 9 10 11 12

Figure 3. Normalised ATS of the two charts using ATSδ as an objective function.
International Journal of Production Research 4655

(1) While the Curt_CUSUM chart has a smaller AND than the Curt_CUSUMObj_ATS chart, the latter has a
smaller ATS than the former at δ = 2. The reason is that if ATSδ at a shift δ is used as the objective function,
the design algorithm will minimise the ATS without considering any other shifts over the whole range of
shifts. As a result, the ATS value at this particular δ will be smaller than that using AND as the objective
function to be minimised.
(2) The ATS values of the Curt_CUSUM chart are smaller than those of the Curt_CUSUMObj_ATS chart over the
whole range of shifts except for δ = 2. This indicates that the overall performance of a control chart using
AND as an objective function is usually better than that of a chart using ATSδ as the objective function.
(3) The above discussions indicates that a design algorithm that minimises the out-of-control ATS value at a
specified process shift will not guarantee the best overall detection effectiveness over the whole domain of
process shifts.

4.2 Comparison in a factorial experiment


Downloaded by [Nanyang Technological University] at 19:36 18 July 2014

Next, the two charts are further studied under more different circumstances through a 23 factorial experiment in which
the three specifications (τ, p0 and ldl ) are used as the input factors and each of them is varied at two levels as shown
below:
s : 1000; 10000:
p0 : 0:005; 0:03:
ld : 3; 7:
The levels of the factors are determined with reference to those commonly used by many authors (Gan 1993; Wu,
Yeo, and Spedding 2001; Haridy et al. 2012). The fourth specification nmin is fixed as 50 for all the cases in this
factorial experiment.
This 2k experiment results in eight different cases or combinations of τ, p0 and ldl as shown in Table 3 (in cases 1
to 8). Case 0 represents the general case conducted in the last section. For each case, the two control charts are designed
and each of them produces an ATS0 no smaller than τ. In all these eight cases, the relative detection effectiveness of the
charts is similar to that revealed in Table 1. Namely, the Curt_CUSUM usually produces smaller out-of-control ATS
values than the conventional CUSUM chart.
The charting parameters of the two charts and the overall performance, as reflected by AND, in this factorial
experiment are listed in Table 3. The value of ANDConv.CUSUM/ANDCurt_CUSUM is always larger than one. This indicates
that the Curt_CUSUM chart always outperforms the conventional CUSUM chart from an overall viewpoint. The former

Table 3. Comparison of the two charts in the 23 factorial design.

AND/
Case τ p0 μδ Chart n k H AND ANDCurt_CUSUM

0 5000 0.01 5 Conv.CUSUM 60 0.950 2.559 6.7371 1.2074


Curt_CUSUM 62 1.240 1.796 5.5796 1.0000
1 1000 0.005 3 Conv.CUSUM 84 0.880 0.242 2.9912 1.1206
Curt_CUSUM 83 0.840 0.368 2.6691 1.0000
2 1000 0.005 7 Conv.CUSUM 64 0.490 1.021 3.0203 1.4049
Curt_CUSUM 81 0.820 0.375 2.1497 1.0000
3 1000 0.03 3 Conv.CUSUM 60 2.850 1.151 8.3159 1.2771
Curt_CUSUM 60 2.850 1.245 6.5113 1.0000
4 1000 0.03 7 Conv.CUSUM 62 2.970 1.060 8.0588 1.8145
Curt_CUSUM 69 3.140 0.861 4.4412 1.0000
5 10,000 0.005 3 Conv.CUSUM 76 0.660 2.7016 8.5704 1.0897
Curt_CUSUM 112 0.870 2.9097 7.8645 1.0000
6 10,000 0.005 7 Conv.CUSUM 56 0.620 2.379 5.5274 1.1683
Curt_CUSUM 66 0.660 2.377 4.7310 1.0000
7 10,000 0.03 3 Conv.CUSUM 56 2.660 3.743 16.9628 1.1145
Curt_CUSUM 56 2.600 4.055 15.2199 1.0000
8 10,000 0.03 7 Conv.CUSUM 66 3.020 3.961 12.1120 1.4843
Curt_CUSUM 71 3.760 3.239 8.1597 1.0000
4656 S. Haridy et al.

overwhelmingly excels the latter when τ is small and p0 is large. For example, in case 4, when τ = 1000 and p0 = 0.03,
the ratio of ANDConv.CUSUM/ANDCurt_CUSUM has its maximum value of 1.81.
Finally, a grand average ANDConv:CUSUM =ANDCurt CUSUM is calculated. It indicates the average of the
ANDConv.CUSUM/ANDCurt_CUSUM values encompassing all the nine cases in Table 3. The result is
ANDConv:CUSUM =ANDCurt CUSUM = 1.30. This reveals that, from the most comprehensive viewpoint (covering all
different values of τ, p0 and ldl ), the Curt_CUSUM chart is more effective than the conventional CUSUM chart by
30%. This reflects the unique contribution of the curtailment mechanism to the improvement of overall detection effec-
tiveness. The curtailment allows the Curt_CUSUM chart to use a fairly large sample size so that it is sensitive to small
p shifts. On the other hand, when a large p shift takes place, the curtailment enables the Curt_CUSUM chart to signal
the large p shift before all of the n units in a sample are inspected. It suggests that the curtailment mechanism is able to
improve the detection effectiveness against both small and large p shifts.

5. Illustrative example
This example illustrates the application of the Curt_CUSUM chart in a company producing transformers. This company
Downloaded by [Nanyang Technological University] at 19:36 18 July 2014

produces 200 transformers per hour (N = 200). A transformer is classified as non-conforming, if any of the three defects
(wrong orientation, misalignment and spot) are detected. The current in-control p0 is estimated as 0.0125 from the
records of pilot runs. Based on some investigation records of the out-of-control cases, the mean value μδ of the random
shift δ is estimated as 4.35. The allowable minimum τ is set as 2000 h. A minimum value of 50 is used as a lower
bound for the sample size n. The specifications are summarised as follows:
p0 = 0.0125, in-control fraction non-conforming.
μδ = 4.35, mean value of the random shift δ.
τ = 2000 h, allowable minimum value of the in-control ATS0.
nmin = 50 units, allowable minimum value of the sample size n.
In this example, the random shift δ is assumed to follow a Rayleigh distribution characterised by μδ, hence the value
of δmax can be determined so that the probability of (δ > δmax) is equal to 0.0001. When μδ = 4.35, the value of δmax is
approximately equal to 10.
The conventional CUSUM chart has also been considered for monitoring the process. The charting parameters of
the two charts as well as their AND values and the ANDConv.CUSUM/ANDCurt_CUSUM ratio are listed below:
Conv.CUSUM chart: n ¼ 65; k ¼ 1:315; H ¼ 1:695; AND ¼ 6:2475  200 ¼ 1249:50;
ANDConv:CUSUM =ANDCurt CUSUM ¼ 1:2917:
Curt CUSUM chart : n ¼ 76; k ¼ 1:650; H ¼ 1:351; AND ¼ 4:8366  200 ¼ 967:32:
As shown, in this example, the Curt_CUSUM chart reduces the average number of non-conforming units by 29%
compared with the conventional CUSUM chart over the tested range of p shifts. The values of the in-control ATS0 and
out-of-control ATS of the two charts are shown in Table 4. It can be observed that the Curt_CUSUM chart always
produces the smallest out-of-control ATS over the entire p shift range. Figure 4 shows the ATS curves of the two charts
for (δ > 2).

Table 4. ATS values of the two charts in the example.

ATS
δ Conv.CUSUM chart Curt_CUSUM chart

1 2097.482 2022.332
2 312.446 297.945
3 160.138 136.611
4 112.627 87.447
5 90.869 64.174
6 77.707 50.569
7 69.698 42.542
8 64.307 36.017
9 60.833 31.495
10 58.510 27.951
International Journal of Production Research 4657

ATS
180

160 Conv. CUSUM chart


140
Curt_CUSUM chart
120

100

80

60

40

20

0
0 1 2 3 4 5 6 7 8 9 10

Figure 4. ATS of the two control charts in the example.


Downloaded by [Nanyang Technological University] at 19:36 18 July 2014

If the Curt_CUSUM chart is selected for this application, 76 units in each sample are inspected one by one. If the
count gti of non-conforming units in a sample is larger than the threshold Gt (Equation (7)) at any time, the inspection
is terminated immediately and the process is signalled as out of control. If this does not happen till the end of the
sample, the process is thought currently in control.

6. Conclusions and discussions


This article proposes a new binomial CUSUM chart with curtailment (Curt_CUSUM chart). While the general idea of
the curtailment is very simple, the new Curt_CUSUM chart outperforms the conventional CUSUM chart under different
settings. On average, the former is more effective than the latter by 30% in terms of AND.
The high overall effectiveness of the Curt_CUSUM chart is mainly attributable to the use of curtailment. The fairly
large sample size of the Curt_CUSUM chart makes it powerful for detecting p shifts of small sizes. Meanwhile, if large
p shifts occur, the curtailment mechanism comes to signal the out-of-control condition before all of the n units in a
sample are inspected.
The whole design of the Curt_CUSUM chart can be easily implemented by a computer programme by following a
well-developed procedure. Once the optimal design is carried out, the designed Curt_CUSUM chart can be used contin-
uously (until the process conditions change) and the improvement in effectiveness can be reaped on a long-term basis.
In a nutshell, the current paper studies the incorporation of the curtailment into the binomial CUSUM chart. It
investigates whether the curtailment feature can be employed to improve the detection power of the binomial CUSUM
chart as it was for the np chart (Wu, Luo, and Zhang 2006). This is not straightforward as the charting parameters, oper-
ational rules, implementation, design and formulation of the CUSUM chart are completely different from that of the
simple np chart. The results of this research reveal that the curtailment technique indeed significantly enhances the effec-
tiveness of the CUSUM chart. Moreover, this paper systematically studies the performance of the Curt_CUSUM chart
under different conditions and provides the readers with the degree of improvement in a quantitative form. Such infor-
mation is not available in the literature and will greatly help the potential users of the Curt_CUSUM chart.
As a by-product of the optimal design, the resultant ATS0 of the Curt_CUSUM chart is usually very close to the
specification τ in spite of the discrete nature of the attribute quality characteristics. This feature enables the designers to
accurately specify and control the false alarm rate and, at the same time, to make full use of the potential of the
detection effectiveness of the chart.
Although the Curt_CUSUM chart is discussed based on a 100% inspection, it can also be used to monitor a process
in which a uniform (or random) sampling inspection is adopted, as explained by Bourke (1991). In such cases, the
Curt_CUSUM chart only takes the inspected units into account and ignores any units produced during periods of non-
inspection.
Finally, the curtailment technique may be used to enhance the power of some other control charts such as the
EWMA and synthetic charts for attributes. Future research is warranted to extend the curtailment technique to cover a
broader scope of designing more effective control charts.

Acknowledgements
The authors would like to thank DAAD (German Academic Exchange Service) greatly for supporting this research work.
4658 S. Haridy et al.

References

Albers, W. 2011. “Control Charts for Health Care Monitoring under Overdispersion.” Metrika 74: 67–83.
Bounds, G. M. 1988. “Success in Implementing Statistical Process Control as a Function of Contextual Variables in 20 Manufacturing
Organizations.” PhD diss., University of Tennessee at Knoxville, TN.
Bourke, P. D. 1991. “Detecting a Shift in Fraction Nonconforming Using Run-length Control Charts with 100% Inspection.” Journal
of Quality Technology 23: 225–238.
Bourke, P. D. 2001. “Sample Size and the Binomial CUSUM Control Chart: The Case of 100% Inspection.” Metrika 53: 51–70.
British Standard Institution. 1985. Quality Management Systems: Quality Control. Vol. 24 of BSI Handbook. London: British
Standard Institution.
Cannon, J., P. A. Krokhmal, Y. Chen, and R. Murphey. 2012. “Detection of Temporal Changes in Psychophysiological Data Using
Statistical Process Control Methods.” Computer Methods and Programs in Biomedicine 107: 367–381.
Castillo, E. D. 1995. “Relations Between X-bar Control Chart Design Variables and Production Control.” International Journal of
Production Research. 33: 2709–2721.
Chen, H. C., A. B. Yeh, C. L. Yen, and L. A. Chen. 2012. “The Density Control Chart: A General Approach for Constructing a Single
Chart for Simultaneously Monitoring Multiple Parameters.” International Journal of Production Research. 50: 3904–3919.
Dondero, C. 1991. “SPC Hits the Road.” Quality Progress 24: 43–44.
Downloaded by [Nanyang Technological University] at 19:36 18 July 2014

Duncan, A. J. 1956. “The Economic Design of X Charts Used to Maintain Current Control of a Process.” Journal of American
Statistical Association 51: 228–242.
Duncan, A. J. 1986. Quality Control and Industrial Statistics. Homewood, IL: Richard, D. Irwin.
Epprecht, E. K., A. F. B. Costa, and F. C. T. Mendes. 2003. “Adaptive Control Charts for Attributes.” IIE Transactions 35: 567–582.
Epprecht, E. K., B. F. T. Simões, and F. C. T. Mendes. 2010. “A Variable Sampling Interval EWMA Chart for Attributes.” The
International Journal of Advanced Manufacturing Technology 49: 281–292.
Gan, F. F. 1993. “An Optimal Design of CUSUM Control Charts for Binomial Counts.” Journal of Applied Statistics 20: 445–460.
Haridy, S., Z. Wu, and J. Flaig. 2012. “Chi-squared Control Chart for Multiple Attributes.” International Journal of Industrial and
Systems Engineering 12: 316–330.
Haridy, S., Z. Wu, M. B. C. Khoo, and F. J. Yu. 2012. “A Combined Synthetic and np Scheme for Detecting Increases in Fraction
Nonconforming.” Computers and Industrial Engineering 62: 979–988.
Haridy, S., Z. Wu, C. K. Lee, and N. Bhuiyan. 2013. “Optimal Average Sample Number of the SPRT Chart for Monitoring Fraction
Nonconforming.” European Journal of Operational Research 229: 411–421.
Haridy, S., Z. Wu, F.-J. Yu, and M. Shamsuzzaman. 2013. “An Optimisation Design of the Combined np-CUSUM Scheme for
Attributes.” European Journal of Industrial Engineering 7: 16–37.
Hawkins, D. M., and D. H. Olwell. 1998. Cumulative Sum Charts and Charting for Quality Improvement. New York:
Springer-Verlag.
Kemp, K. W. 1962. “The Use of Cumulative Sums for Sampling Inspection Schemes.” Applied Statistics 11: 16–31.
Knoth, S., and S. Steinmetz. 2013. “EWMA p Charts Under Sampling by Variables.” International Journal of Production Research
51: 3795–3807.
Lee, J., Y. Hur, S. H. Kim, and J. R. Wilson. 2012. “Monitoring Nonlinear Profiles Using a Wavelet-based Distribution-free CUSUM
Chart.” International Journal of Production Research 50: 6574–6594.
Lim, T. O., A. Soraya, L. M. Ding, and Z. Morad. 2002. “Assessing Doctors’ Competence: Application of CUSUM Technique in
Monitoring Doctors’ Performance.” International Journal for Quality in Health Care 14: 251–258.
Lucas, J. M. 1985a. “Counted Data CUSUM’s.” Technometrics 27: 129–144.
Lucas, J. M. 1985b. “Cumulative Sum (CUSUM) Control Schemes.” Communications in Statistics – Theory and Methods 14:
2689–2704.
Lucas, J. M. 1989. “Control Schemes for Low Count Levels.” Journal of Quality Technology 21: 199–201.
Montgomery, D. C. 2009. Introduction to Statistical Quality Control. 6th ed. New York: Wiley.
Mousavi, S., and M. R. Reynolds Jr. 2009. “A CUSUM Chart for Monitoring a Proportion with Autocorrelated Binary Observations.”
Journal of Quality Technology 41: 401–414.
Page, E. S. 1954. “Continuous Inspection Schemes.” Biometrika 41: 100–115.
Perry, M. B., and J. J. Pignatiello. 2011. “Estimating the Time of Step Change with Poisson CUSUM and EWMA Control Charts.”
International Journal of Production Research 49: 2857–2871.
Pignatiello Jr, J. J., and T. R. Samuel. 2001. “Estimation of the Change Point of a Normal Process Mean in SPC Applications.”
Journal of Quality Technology 33: 82–95.
Radaelli, G. 1994. “Poisson and Negative Binomial Dynamics for Counted Data under CUSUM-type Charts.” Journal of Applied
Statistics 21: 347–356.
Reynolds Jr., M. R., and Z. G. Stoumbos. 1999. “A CUSUM Chart for Monitoring a Proportion When Inspecting Continuously.”
Journal of Quality Technology 31: 87–108.
Reynolds Jr., M. R., and Z. G. Stoumbos. 2004. “Should Observations be Grounded for Effective Process Monitoring.” Journal of
Quality Technology 36: 343–366.
International Journal of Production Research 4659

Saniga, E. M., D. J. Davis, and T. P. Mcwilliams. 1995. “Economic, Statistical, and Economic-statistical Design of Attribute Charts.”
Journal of Quality Technology 27: 56–73.
Serel, D. A., and H. Moskowitz. 2008. “Joint Economic Design of EWMA Control Charts for Mean and Variance.” European
Journal of Operational Research 184: 157–168.
Shu, M. H., B. M. Hsu, and M. C. Hu. 2012. “Optimal Combination of Soldering Conditions of BGA for Halogen-free and Lead-free
SMT-green Processes.” Microelectronics Reliability 52: 2690–2700.
Shu, L., W. Jiang, and K. L. Tsui. 2008. “A Weighted CUSUM Chart for Detecting Patterned Means Shifts.” Journal of Quality
Technology 40: 194–213.
Tannock, J. D. T. 2003. “A Fuzzy Control Charting Method for Individuals.” International Journal of Production Research 41:
1017–1032.
Tao, Z., F. Liu, F. Shen, M. Suh, and D. Booth. 2012. “A New Control Chart Based on the Loess Smooth Applied to Information
System Quality Performance.” International Journal of Operational Research 15: 74–93.
Wang, S., T. Wu, S.-J. Weng, and J. Fowler. 2012. “A Control Chart Based Approach to Monitoring Supply Network Dynamics
Using Kalman Filtering.” International Journal of Production Research 50: 3137–3151.
White, C. L., J. B. Keats, and J. Stanley. 1997. “Poisson CUSUM vs. c Chart for Defect Data.” Quality Engineering 9: 673–679.
Woodall, W. H. 1997. “Control Charts Based on Attribute Data: Bibliography and Review.” Journal of Quality Technology 29:
172–183.
Downloaded by [Nanyang Technological University] at 19:36 18 July 2014

Wu, Z., J. Jiao, and Y. Liu. 2008. “A Binomial CUSUM Chart for Detecting Large Shifts in Fraction Nonconforming.” Journal of
Applied Statistics 35: 1267–1276.
Wu, Z., H. Luo, and X. Zhang. 2006. “Optimal np Control Chart with Curtailment.” European Journal of Operational Research 174:
1723–1741.
Wu, Z., and Y. Tian. 2005. “Weighted-loss-function CUSUM Chart for Monitoring Mean and Variance of a Production Process.”
International Journal of Production Research 43: 3027–3044.
Wu, Z., M. Xie, and Y. Tian. 2002. “Optimization Design of the X & S Charts for Monitoring Process Capability.” Journal of
Manufacturing Systems 21: 83–92.
Wu, Z., S. H. Yeo, and T. A. Spedding. 2001. “A Synthetic Control Chart for Detecting Fraction Nonconforming Increases.” Journal
of Quality Technology 33: 104–111.
Yang, M., Z. Wu, K. M. Lee, and M. B. C. Khoo. 2012. “The X Control Chart for Monitoring Process Shifts in Mean and Variance.”
International Journal of Production Research 50: 893–907.
Yu, J., and J. Liu. 2011. “LRProb Control Chart Based on Logistic Regression for Monitoring Mean Shifts of Auto-correlated
Manufacturing Processes.” International Journal of Production Research 49: 2301–2326.

Das könnte Ihnen auch gefallen