Journal of Quality in Maintenance Engineering

2
A new CUSUM-S control chart for monitoring the process variance
P. Castagliola G. Celano S. Fichera
Article information:
To cite this document:
2
P. Castagliola G. Celano S. Fichera, (2009),"A new CUSUM-S control chart for monitoring the process
variance", Journal of Quality in Maintenance Engineering, Vol. 15 Iss 4 pp. 344 - 357
Permanent link to this document:
http://dx.doi.org/10.1108/13552510910997724
Downloaded on: 05 April 2016, At: 08:34 (PT)
References: this document contains references to 15 other documents.
Downloaded by University of Georgia At 08:34 05 April 2016 (PT)

To copy this document: [email protected]

The fulltext of this document has been downloaded 488 times since 2009*
Users who downloaded this article also downloaded:
(2013),"Food safety in the US catering industry: empirical findings", International Journal of Contemporary
Hospitality Management, Vol. 25 Iss 3 pp. 365-382 http://dx.doi.org/10.1108/09596111311311026
(2013),"The impacts of human capital in enhancing new venture's performance: Competence, motivation
and creativity", Journal of Knowledge-based Innovation in China, Vol. 5 Iss 2 pp. 146-168 http://
dx.doi.org/10.1108/JKIC-06-2013-0011
(2015),"How to transform research cases into teaching cases", The CASE Journal, Vol. 11 Iss 1 pp.
133-143 http://dx.doi.org/10.1108/TCJ-04-2014-0027

Access to this document was granted through an Emerald subscription provided by emerald-srm:403714 []
For Authors
If you would like to write for this, or any other Emerald publication, then please use our Emerald for
Authors service information about how to choose which publication to write for and submission guidelines
are available for all. Please visit www.emeraldinsight.com/authors for more information.
About Emerald www.emeraldinsight.com
Emerald is a global publisher linking research and practice to the benefit of society. The company
manages a portfolio of more than 290 journals and over 2,350 books and book series volumes, as well as
providing an extensive range of online products and additional customer resources and services.
Emerald is both COUNTER 4 and TRANSFER compliant. The organization is a partner of the Committee
on Publication Ethics (COPE) and also works with Portico and the LOCKSS initiative for digital archive
preservation.

*Related content and download information correct at time of download.

 The current issue and full text archive of this journal is available at
www.emeraldinsight.com/1355-2511.htm

JQME
15,4 A new CUSUM-S 2 control chart
for monitoring the process
variance
344
P. Castagliola
Université de Nantes & IRCCyN UMR CNRS 6597, Carquefou, France, and
G. Celano and S. Fichera
Università di Catania, Catania, Italy

Abstract
Downloaded by University of Georgia At 08:34 05 April 2016 (PT)

Purpose – The purpose of this paper is to introduce and investigate the performances of a new
CUSUM-S 2 control chart designed to monitor the sample variance of samples from a normally
distributed population.
Design/methodology/approach – The proposed chart monitors a statistic computed as a
logarithmic transformation of the sample variance; the introduction of the sample variance
logarithmic transformation has a twofold effect: to quickly detect the occurrence of an “out-of-control”
condition; to deal with a quasi-standard normal statistic.
Findings – A design strategy trying to minimize the “out-of-control” average run length (ARL) of the
chart is presented and the statistical performance of the CUSUM-S 2 chart has been assessed through a
comparison with an EWMA-S 2 control chart proposed in the literature to monitor the process
dispersion.
Research limitations/implications – The paper only deals with uncorrelated normally
distributed data.
Practical implications – The obtained results show how the CUSUM-S 2 chart is particularly
suitable when reduction in the process dispersion should be detected by means of subgroups having
limited sample sizes.
Originality/value – The paper shows the new CUSUM-S 2 control chart allows a decreasing of the
variability to be detected faster than the corresponding EWMA-S 2 control chart proposed earlier in
the literature.
Keywords Cumulative sum techniques, Average run length, Statistical process control, Control charts
Paper type Research paper

1. Introduction
Manufacturing processes are characterized by a certain amount of variability, which in
some cases can strongly affect the quality of the outcome. Generally, process
variability consists of an “inherent” or “natural” variability, which cannot be
eliminated and can be considered as a background noise. Sometimes a further source of
variability is present in the process due to the occurrence of a special cause: this
variability sums to the natural variability and can lead to an unacceptable level of
process performance. Several causes can give raise to an excessive amount of process
Journal of Quality in Maintenance variability: defective raw materials, incorrect methods of working, erroneous machine
Engineering setups or operating conditions, operator errors, and so on. The increase in process
Vol. 15 No. 4, 2009
pp. 344-357 variability is generally reflected on the process mean and/or process dispersion: a shift
q Emerald Group Publishing Limited
1355-2511
in process mean causes the outcome be far from the target with a consequent
DOI 10.1108/13552510910997724 production of an excessive number of nonconformities; an increase in process
 dispersion leads to a larger spread of data, which corresponds to a lower uniformity in A new
process output: the result is once again an excessive number of produced CUSUM-S 2
nonconforming items. Processes characterized exclusively by the presence of the
natural variability are said to operate in the “in-control” condition; when a special control chart
cause occurs, the process state is said to be “out-of-control”.
Statistical process control SPC provides a large set of techniques designed to help
practitioners in monitoring the state of control of a manufacturing process and quickly 345
detecting special causes when they occur: control charts are a widely used on-line
control technique and can be implemented on sample statistics related to both process
mean and dispersion. An efficient monitoring procedure based on control charts should
include the contemporary implementation of two charts for monitoring both the
process mean and dispersion: to do this, the sample mean Shewhart X
chart is widely
used to monitor the process position and the R, S, S 2 control charts are implemented by
practitioners to get a continuous monitoring about the process dispersion,
Downloaded by University of Georgia At 08:34 05 April 2016 (PT)

(Montgomery, 2004). As an alternative to the Shewhart charts, many other charting

schemes have been proposed in literature to detect the presence of unnatural variability
within a manufacturing process. Since the mid-twentieth century the use of the
CUSUM (cumulative sum) control chart as a process-monitoring tool has become more
and more popular in the statistical process control field. Implementing a CUSUM chart
to control a manufacturing process requires the computation and plotting of a random
variable which is a function of the current sample statistic and of the past samples
collected from the process: this allows the CUSUM to prevail over the traditional
Shewhart chart in terms of statistical sensitivity when small shifts in the process
position and/or dispersion are expected. The first CUSUM chart monitoring the sample
mean was introduced by Page (1954). More recently, the CUSUM charts were also
employed to perform the on-line control of the process data dispersion: a CUSUM chart
monitoring the sample range and aimed at detecting increases in process dispersion
was proposed by Page (1963); Tuprah and Ncube (1987) studied the performance of a
CUSUM chart monitoring the sample standard deviation; a logarithmic transformation
of the sample variance was used by Chang and Gan (1995) as monitored statistic of a
LnS 2 CUSUM chart. Several models of CUSUM control charts were compared each
other and with EWMA and Shewhart charts monitoring the process dispersion in a
paper by Acosta-Mejia et al. (1999): the best performing chart was found to be a Cp
CUSUM derived from a likelihood ratio test of the change point problem for the process
variance.
In this paper a CUSUM-S 2 chart monitoring a statistic based on a proper
logarithmic transformation of the sample variance is presented. The proposed
transformation extend that reported in Chang and Gan (1995) and has a twofold effect:
(1) to quickly detect the occurrence of an “out-of-control” condition;
(2) to deal with a quasi-standard normal statistic.

To test the performance of the proposed chart, a numerical analysis was carried out by
computing the “out-of-control” average run length (ARL), i.e. the expected number of
samples to be taken between the occurrence of a special cause and its detection on the
chart. The statistical efficiency of the CUSUM-S 2 chart has been compared with that of
the corresponding EWMA-S 2 developed by Castagliola (2005a). No comparison was
performed with the Cp CUSUM: it is worth noting that the Cp CUSUM presented in
 JQME Acosta-Mejia et al. (1999) works with a hybrid statistic which is a function of both the
15,4 standard deviation and the square of the standardized observations within a sample;
the aim of this paper is dealing with and comparing charts monitoring statistics which
are exclusively a function of the sample variance, thus avoiding to work with too much
complicated random variables which could result not easy to be managed in practice.
The rest of the paper is organized as follows: in the next paragraph the scheme of
346 the CUSUM-S 2 chart is presented together with the properties of the monitored
statistic and a numerical example illustrating how the proposed chart works on
process data is reported; then, the optimization procedure implemented to compute the
optimal statistical design of the CUSUM-S 2 chart is detailed; finally, the computational
results and the tables reporting the optimal design parameters of the CUSUM-S 2 chart
are presented and compared with those of other control charting schemes. In the
Appendix, a brief description of the Markov chain model related to the average run
length computation of the CUSUM-S 2 chart is reported.
Downloaded by University of Georgia At 08:34 05 April 2016 (PT)

2. A new CUSUM chart for monitoring the sample variance

Let Xk,1, . . . ,Xk,n be a sample of n independent normal (m0, s0) random variables, where
m0 is the nominal process mean, s0 is the nominal process standard-deviation and k is
the subgroup number. The investigated chart is used to monitor the process dispersion,
therefore the “out-of-control” condition for the process corresponds to the occurrence of
a special cause, which leaves unchanged the process mean m and shifts the standard
deviation from s0 to s1: the entity of this shift is quantified through the parameter
t ¼ s1 =s0 . Let S 2k be the sample variance of subgroup k, i.e.

1 X n
S 2k ¼
2
ðX k;j 2 XkÞ
n 2 1 j¼1

is the sample mean of subgroup k. In order to monitor the process variability,

where Xk
(Castagliola, 2005a, b) and (Castagliola et al., 2006a,b; Castagliola et al., 2006c) proposed
new EWMA type control charts (for the sample variance S 2, the sample standard
deviation S and the range R), mainly inspired from the work of (Crowder and Hamilton,
1992; Hamilton and Crowder, 1992). In order to monitor the process variance,
Castagliola (2005a) suggested to apply the following three parameters (a,b,c)
logarithmic transformation (which is Johnson’s type transformation) to S 2k :

T k ¼ a þ blnðS 2k þ cÞ
where a, b and c . 0 are three constants, and then, to use the classical EWMA
approach on the Tk:
Z k ¼ ð1 2 lÞZ k21 þ lT k
where l is a smoothing constant satisfying 0 , l # 1. The main expectation of this
approach is that if parameters a, b and c are judiciously selected then this
transformation may result in approximate normality to Tk better than the approach in
Crowder and Hamilton (1992). Trying to make Tk more normally distributed is related
to making the distribution of Tk more symmetrical. The (fixed) control limits of the
EWMA-S 2 control chart proposed by Castagliola (2005a) (corresponding to the Zk) are:
 rffiffiffiffiffiffiffiffiffiffiffiffi
l A new
LCL ¼ EðT k Þ 2 K sðT k Þ ð1Þ
22l CUSUM-S 2
control chart
rffiffiffiffiffiffiffiffiffiffiffiffi
l
UCL ¼ EðT k Þ þ K sðT k Þ ð2Þ
22l
347
where K is a positive constant (for example K ¼ 3) and where E(Tk) and s(Tk) are the
theoretical expectance and standard deviation of Tk respectively. Castagliola (2005a)
proved that the constants a, b and c are necessarily equal to:
b ¼ BðnÞ
c ¼ CðnÞs20
Downloaded by University of Georgia At 08:34 05 April 2016 (PT)

a ¼ AðnÞ 2 2BðnÞlnðs0 Þ

where A(n), B(n), and C(n) are three functions depending only on the sample size n.
The closed form of these functions are derived in Appendix of Castagliola (2005a) and
their values can be found in Table I for n ¼ 3; . . . ; 15. In the same Appendix,
Castagliola also derived the distribution f T k ðtjnÞ of Tk and showed that this
distribution depends only on n. This property is important since it allows the
calculation of the values of EðT k Þ and sðT k Þ independently of the value of s0 . The
computation of EðT k Þ and sðT k Þ has been achieved by numerical quadrature, for
n ¼ 3; . . . ; 15, and the results are also shown in Table I. One can notice that values for
EðT k Þ are very close to 0. In fact they are so close to 0 that assuming EðT k Þ ¼ 0 is a
very good approximation. Finally, Castagliola (2005a) demonstrated that a reasonable
value for Z0 is:
Z 0 ¼ AðnÞ þ BðnÞlnð1 þ CðnÞÞ
As can be noticed, this value depends only on n and not on s0 . The values for Z0 are in
Table I. One can notice that these values are also close to 0 and can be replaced by 0 in

n A(n) B(n) C(n) Z0 E(Tk) s(Tk)

3 20.6627 1.8136 0.6777 0.276 0.02472 0.9165

4 20.7882 2.1089 0.6261 0.237 0.01266 0.9502
5 20.8969 2.3647 0.5979 0.211 0.00748 0.9670
6 20.9940 2.5941 0.5801 0.193 0.00485 0.9765
7 21.0827 2.8042 0.5678 0.178 0.00335 0.9825
8 21.1647 2.9992 0.5588 0.167 0.00243 0.9864
9 21.2413 3.1820 0.5519 0.157 0.00182 0.9892 Table I.
10 21.3135 3.3548 0.5465 0.149 0.00141 0.9912 Constants A(n), B(n),
11 21.3820 3.5189 0.5421 0.142 0.00112 0.9927 C(n), Z 0 , EðT k Þ, sðT k Þ
12 21.4473 3.6757 0.5384 0.136 0.00090 0.9938 defining the logarithmic
13 21.5097 3.8260 0.5354 0.131 0.00074 0.9947 transformation of the
14 21.5697 3.9705 0.5327 0.126 0.00062 0.9955 sample variance S 2k , for
15 21.6275 4.1100 0.5305 0.122 0.00052 0.9960 n [ {3; . . . ; 15}
 JQME practice with little practical effect. The goal of this paper is to extend the EWMA
approach presented above to a CUSUM type approach, i.e. to monitor the statistic
15,4
þ
Z k ¼ maxðZ 2
k ; Zk Þ

with
348 Z2 2
k ¼ maxð0; Z k21 þ EðT k Þ 2 T k 2 LÞ

Zþ þ
k ¼ maxð0; Z k21 þ T k 2 EðT k Þ 2 LÞ

where the slack value L $ 0 is a constant. The upper control limit of the CUSUM-S 2
control chart (the lower control limit is LCL ¼ 0) is
UCL ¼ K
Downloaded by University of Georgia At 08:34 05 April 2016 (PT)

where the decision interval K $ 0 is a constant.

3. Illustrative example
The goal of this example is to show how the CUSUM-S 2 control chart behaves in the
case of an increase and a decrease in the nominal process variability. The first 100 data
points plotted in Figure 1 (top and bottom) consist of 20 identical subgroups of n ¼ 5
observations randomly generated from a normal (20, 0.1) distribution (corresponding
to an in-control process), while the last 50 data points of Figure 1 (top) consist of 10
subgroups of n ¼ 5 observations randomly generated from a normal (20, 0.2)
distribution (the nominal process standard-deviation s0 has increased by a factor of 2),
and the last 50 data points of Figure 1 (bottom) consist of ten subgroups of n ¼ 5
observations randomly generated from a normal (20, 0.05) distribution (the nominal
process standard-deviation s0 has decreased by a factor of 2). The corresponding 30
sample variances are plotted in Figure 2, for the two cases (increasing and decreasing).
At this step, the asymmetry between increasing and decreasing sample variances
is particularly noticeable. If n ¼ 5 and s0 ¼ 0:1, then b ¼ 2:3647,
c ¼ 0.5979 £ 0.12 ¼ 0.005979 and a ¼ 20:8969 2 2 £ 2:3647 £ lnð0:1Þ ¼ 9:9929. The
30 transformed sample variances are plotted in Figure 3. The CUSUM-S 2 control chart
ðK * ¼ 2:2409; L * ¼ 0:98Þ plotted in Figure 4 clearly detects an out of control signal at
the 22th subgroup (in the increasing case) and at the 26th subgroup (in the decreasing
case), pointing out that an increase/decrease of the process variability occurred.

4. ARL optimization for the CUSUM-S 2 control chart

In this paper the used statistical measures of performance is the ARL (average run
length), defined as the average number of samples before the control chart signals an
“out-of-control” condition or issues a false alarm. Once the control chart parameters (in
our case K and L) are defined, the ARL of the CUSUM-S 2 control chart can be
numerically evaluated for a particular shift t ¼ s1 =s0 of the process variability using
an approach based on a Markov-chain approximation. This procedure, described in the
Appendix, is similar to that introduced by Brook and Evans (1972), and used by Lucas
and Saccucci (1990).
In practice, it is sometimes interesting for the quality practitioner to find out the
optimal couples ðK * ; L * Þ veryfing:
 A new
CUSUM-S 2
control chart

349
Downloaded by University of Georgia At 08:34 05 April 2016 (PT)

Figure 1.
Data with an increasing
variance (top), with a
decreasing variance
(bottom)
 JQME
15,4

350
Downloaded by University of Georgia At 08:34 05 April 2016 (PT)

Figure 2.
Sample variances S 2k
corresponding to data of
Figure 1

Figure 3.
Transformed sample
variances T k ¼
a þ blnðS 2k þ cÞ
corresponding to data of
Figure 1
 A new
CUSUM-S 2
control chart

351
Downloaded by University of Georgia At 08:34 05 April 2016 (PT)

Figure 4.
CUSUM-S 2 control chart
ðK * ¼ 2:2409; L * ¼ 0:98Þ
corresponding to data of
Figure 1

ARLð1; K * ; L * ; nÞ ¼ ARL0

ARLðt; K * ; L * ; nÞ ¼ minARLðt; K; L; nÞ
K;L

i.e. the optimal couples ðK * ; L * Þ such that:

.
when the process is functioning at the nominal variability s ¼ s0 (or
equivalently t ¼ 1), then ARL ¼ ARL0 where ARL0 is the “in-control” ARL.
.
for a specified value of the shift t, the unique couple ðK * ; L * Þ yields the smallest
possible “out-of-control” ARL ¼ ARL * .

These optimal couples ðK * ; L * Þ and the corresponding minimal “out-of-control”

ARL *, presented in Table II, have been computed using a non-linear optimization
algorithm for n ¼ 3; 5; 7; 9, for t ¼ 0:5; 0:6; . . . ; 0:9; 0:95; 1:05; 1:1; 1:2; . . . ; 2 and
for ARL0 ¼ 370:4. For example, the optimal couple ðK * ; L * Þ ensuring the smallest
ARL for a shift t ¼ 1:2 and n ¼ 7 is ðK * ¼ 5:3898; L * ¼ 0:41Þ and the
corresponding minimal ARL is ARL * ¼ 18:0. The values within parentheses in
Table II are the minimal out-of-control ARL * of the EWMA-S 2 control chart
presented in Castagliola (2005a). Table II shows some interesting results
concerning the optimal couples of design parameters ðK * ; L * Þ of the CUSUM-S 2
chart and their statistical performances compared to those of the corresponding
EWMA-S 2.
The optimal couples of design parameters ðK * ; L * Þ and the comparison of the
statistical performances of the two charts suggest some interesting findings:
 Downloaded by University of Georgia At 08:34 05 April 2016 (PT)

15,4

352
JQME

Table II.

1.05,1.1,1.2,. . .,2
and corresponding

and for ARL0 ¼ 370.4

t ¼ 0.5,0.6,. . .,0.9,0.95,
minimal out-of-control
Optimal couples (K *,L *)

ARL * for n ¼ 3,5,7,9, for

n¼3 n¼5 n¼7 n¼9
* * * * * * * *
t K L* ARL K L* ARL K L* ARL K L* ARL

0.50 4.9853 0.39 10.8 3.4158 0.66 5.6 2.6512 0.87 3.8 2.2147 1.04 2.9
0.60 5.9450 0.31 15.4 (17.8) 4.3440 0.51 8.3 (9.0) 3.4818 0.67 5.7 (6.2) 2.9276 0.81 4.4 (4.7)
0.70 7.3463 0.23 24.1 (28.5) 5.6793 0.37 13.4 (14.6) 4.6433 0.49 9.4 (10.0) 4.0612 0.58 7.3 (7.7)
0.80 9.6080 0.15 44.0 (54.3) 7.7933 0.24 25.4 (28.2) 6.6862 0.31 18.2 (19.5) 5.9259 0.37 14.3 (15.1)
0.90 13.998 0.07 108.9 (155.8) 11.785 0.12 68.4 (81.0) 10.699 0.15 51.1 (56.8) 9.7217 0.18 41.3 (44.7)
0.95 17.001 0.04 216.9 (325.7) 15.978 0.06 154.8 (202.9) 15.401 0.07 122.9 (150.6) 13.989 0.09 103.1 (120.9)
1.05 3.4322 0.61 183.3 (153.6) 12.902 0.10 145.7 (121.4) 13.201 0.10 117.5 (99.3) 12.732 0.11 99.5 (84.0)
1.10 8.9175 0.17 98.6 (64.9) 9.7068 0.17 64.6 (44.8) 9.0117 0.20 49.2 (34.9) 8.5740 0.22 40.3 (29.0)
1.20 5.6732 0.33 39.5 (21.6) 5.9309 0.35 24.3 (15.3) 5.3898 0.41 18.0 (12.4) 4.9728 0.46 14.5 (10.7)
1.30 4.2860 0.47 21.7 (11.6) 3.9280 0.57 13.1 (8.8) 3.6405 0.64 9.6 (7.4) 3.3501 0.71 7.7 (6.6)
1.40 3.5864 0.58 14.1 (7.8) 2.7317 0.82 8.3 (6.2) 2.5530 0.90 6.1 (5.3) 2.4018 0.97 4.9 (4.7)
1.50 2.9582 0.72 10.2 (5.9) 2.2409 0.98 5.9 (4.8) 2.0653 1.08 4.3 (4.2) 1.9183 1.17 3.5 (3.4)
1.60 2.4164 0.89 7.8 (4.8) 2.0032 1.08 4.5 (4.0) 1.7944 1.21 3.3 (3.2) 1.6730 1.30 2.7 (2.7)
1.70 2.1455 1.00 6.2 (4.0) 1.8413 1.16 3.6 (3.5) 1.6346 1.30 2.7 (2.6) 1.6730 1.30 2.2 (2.2)
1.80 1.9764 1.08 5.2 (3.5) 1.7156 1.23 3.0 (2.9) 1.6346 1.30 2.3 (2.2) 1.6730 1.30 1.9 (1.9)
1.90 1.8619 1.14 4.4 (3.1) 1.6172 1.29 2.6 (2.5) 1.6346 1.30 2.0 (2.0) 1.6730 1.30 1.7 (1.6)
2.00 1.7732 1.19 3.8 (2.9) 1.6015 1.30 2.3 (2.3) 1.6346 1.30 1.8 (1.8) 1.6730 1.30 1.5 (1.5)
 .
as usual for a CUSUM chart, when the slack value L reduces the decision interval A new
K is forced to enlarge to guarantee for an “in-control” ARL equal to 370; CUSUM-S 2
therefore, when small shifts in the variance are expected, i.e. 0:9 # t # 1:1, K is
always larger than 8.5 similarly to what happens for CUSUM charts monitoring control chart
the sample mean.
.
when dealing with CUSUM control charts monitoring the sample mean, a
quasi-optimal decision for the slack value L is suggested to be equal to one half 353
the shift of the process mean, (Montgomery, 2004). In the case of the transformed
variable T k it is worth noting, however, that even if the logarithmic
transformation allows the T k distribution to be quasi – standard normal for
both the “in-control” and “out-of-control” operating conditions, the shifting of the
transformed statistic to the “out-of-control” condition should not be interpreted
as the shift of the sample mean of a process parameter having mean equal to 0
and standard deviation equal to 1. When a shift in the process dispersion occurs,
Downloaded by University of Georgia At 08:34 05 April 2016 (PT)

the T k distribution shifts on the right (left) for t . 1 ðt , 1Þ changing

corresponding value of EðT k Þ and modifies its shape resulting in a variance
V ðT k Þ increase (reduction). For this reason the optimal values of the slack
parameter L cannot be suggested to be equal to one half the shift in the expected
value EðT k Þ.
.
Statistically, no chart always performs the better: in fact, the statistical
performance, expressed through the “out-of-control” ARL, is influenced by how
the variation in process dispersion is headed for: for scenarios characterized by
an increase in process dispersion ðt . 1Þ the CUSUM-S 2 shows a poorer
statistical performance than the corresponding EWMA-S 2. On the other hand,
when the process dispersion reduces ðt , 1Þ an opposite trend occurs. This
result is due to the different sensitivity to the process dispersion shifts showed
by the EMWA and CUSUM optimized charts: the former is more sensitive to the
increase of V ðT k Þ; the latter works better when V ðT k Þ reduces. The selected
sample partially influence this behaviour: as n increases, the difference in the
statistical performances becomes tighter and control charts designed for large
subgroup sizes tend to have the same “out-of-control” ARL, whichever the
process dispersion variation.

On the basis of these conclusions, the following guidelines are suggested: when the
subgroup size should be maintained small always prefer the CUSUM-S 2 if reductions
in the process variation are expected, otherwise implement the EWMA-S 2. When the
subgroup size can be settled relatively large ðn . 10Þ use the EWMA-S 2 for a major
easiness of use whichever the entity of the variation in the dispersion of the process
data.

5. Conclusions
In this paper a CUSUM-S 2 chart has been proposed to perform the on-line
monitoring of the data dispersion of a parameter which affects the output quality of
a manufacturing process. The CUSUM statistic plotted on this chart is computed
after a logarithmic transformation of the original sample variance random variable.
The CUSUM-S 2 control chart shares the same advantages with the EWMA-S 2
proposed by Castagliola (2005a), i.e. easy-to-use scheme for computing the control
 JQME limits and the parameters, improved performance in terms of normality/symmetry
15,4 of the transformed sample variances, and possibility of detecting increases and
decreases of the nominal standard deviation in a similar way. The statistical
efficiency of the proposed CUSUM-S 2 control chart has been measured by means of
its “out-of-control” average run length. The statistical efficiency of the CUSUM-S 2
chart has been compared with that of the corresponding EWMA-S 2 developed by
354 Castagliola (2005a). The obtained results suggest to prefer the CUSUM-S 2 if
reductions in the process variation are expected, otherwise implement the
EWMA-S 2 when the subgroup size should be maintained small. When the
subgroup size can be settled relatively large ðn . 10Þ always use the EWMA-S 2 for
a major easiness of use whichever the entity of the variation in the dispersion of the
process data. Future research will be devoted to the development of adaptive
versions of the CUSUM-S 2 chart which allow the sampling frequency and/or the
sample size to be changed at each sampling epoch depending on the past history of
Downloaded by University of Georgia At 08:34 05 April 2016 (PT)

the process.

References
Acosta-Mejia, C., Pigniatiello, J. and Rao, B. (1999), “A comparison of control charting
procedures for monitoring process dispersion”, IIE Transactions, Vol. 31 No. 6,
pp. 569-79.
Brook, D. and Evans, D. (1972), “An approach to the probability distribution of CUSUM run
length”, Biometrika, Vol. 59 No. 3, pp. 539-49.
Castagliola, P. (2005a), “A new S 2-EWMA control chart for monitoring the process variance”,
Quality and Reliability Engineering International, Vol. 21 No. 8, pp. 781-94.
Castagliola, P. (2005b), “A R-EWMA control chart for monitoring the process range”,
International Journal of Reliability, Quality and Safety Engineering, Vol. 12 No. 1,
pp. 31-49.
Castagliola, P., Celano, G. and Fichera, S. (2006a), “Evaluation of the statistical performance of a
variable sampling interval R EWMA control chart”, Quality Technology and Quantitative
Management, Vol. 3 No. 3, pp. 307-23.
Castagliola, P., Celano, G. and Fichera, S. (2006b), “Monitoring process variability using
EWMA”, in Pham, H. (Ed.), Springer Handbook of Engineering Statistics, Springer, New
York, NY, pp. 291-325.
Castagliola, P., Celano, G., Fichera, S. and Giuffrida, F. (2006c), “A variable sampling interval
S 2-EWMA control chart for monitoring the process variance”, International Journal of
Technology Management, Vol. 37 Nos 1/2, pp. 125-46.
Chang, T. and Gan, F. (1995), “A cumulative sum control chart for monitoring process variance”,
Journal of Quality Technology, Vol. 27 No. 2, pp. 109-19.
Crowder, S. and Hamilton, M. (1992), “An EWMA for monitoring a process standard-deviation”,
Journal of Quality Technology, Vol. 24 No. 1, pp. 12-21.
Hamilton, M. and Crowder, S. (1992), “Average run lengths of EWMA control charts for
monitoring a process standard deviation”, Journal of Quality Technology, Vol. 24 No. 1,
pp. 44-50.
Lucas, J. and Saccucci, M. (1990), “Exponentially weighted moving average control schemes:
properties and enhancements”, Technometrics, Vol. 32 No. 1, pp. 1-12.
Montgomery, D. (2004), Introduction to Statistical Quality Control, 5th ed., Wiley, New York,
NY.
 Page, E. (1954), “Continuous inspections schemes”, Biometrika, Vol. 41 Nos 1/2, pp. 100-14. A new
Page, E. (1963), “Controlling the standard deviation by CUSUM and warning lines”,
Technometrics, Vol. 5 No. 3, pp. 307-15.
CUSUM-S 2
Tuprah, K. and Ncube, M. (1987), “A comparison of dispersion quality control charts”, Sequential
control chart
Analysis, Vol. 6 No. 2, pp. 155-63.

Appendix 355
The ARL of the CUSUM-S 2 control chart can be numerically evaluated using an approach
based on a Markov-chain approximation. This Markov chain approach, originally proposed
Brook and Evans (1972), is flexible and relatively easy to use. This procedure involves
dividing the interval between LCL ¼ 0 and UCL ¼ K (see Figure A1) into p subintervals of
width 2d, where d ¼ ðUCL 2 LCLÞ=ð2pÞ ¼ K=ð2pÞ. When the number of subintervals p is
sufficiently large, the finite approach provides an effective method that allows ARL to be
accurately evaluated.
þ þ
Downloaded by University of Georgia At 08:34 05 April 2016 (PT)

The statistics Z 2 2
k ¼ maxð0; Z k21 þ EðT k Þ 2 T k 2 LÞ and Z k ¼ maxð0; Z k21 þ T k 2
EðT k Þ 2 LÞ are said to be in transient state j at time k if H j 2 d , Z 2
k , H j þ d and Hj 2 d ,
Zþk , H j þ d respectively, for j ¼ 1; 2; . . . ; p, where Hj represents the midpoint of the jth
þ þ
subinterval. The statistics Z 2 2
k and Z k are in the absorbing state at time k if Z k . K and Z k . K
2
respectively. The ARL of the CUSUM-S control chart is equal to:

1 ARL 2 ARL þ
ARL ¼ ¼
1 1
ARL 2 þ ARL þ
ARL 2 þ ARL þ
þ
where ARL 2 and ARL þ are the ARL’s corresponding to statistics Z 2
k and Z k respectively.
2 þ
Approximations for both ARL and ARL are

ARL 2 . u T Q 2 1
ARL þ . u T Q þ 1

where u ¼ ð1; 0; . . . ; 0ÞT is the ( p, 1) initial probability vector (we assume the initial state is state
1), Q 2 ¼ ðI 2 P 2 Þ21 and Q þ ¼ ðI 2 P þ Þ21 are the fundamental ( p, p) matrices, P 2 and P þ
are the (p, p) transition probabilities matrices and 1 is a ( p,1) vector of sampling frequencies
h ¼ 1. The transition probability matrices P 2 and P þ contain the one step transition
þ
probabilities. The generic elements p2 i;j and pi;j of P
2
and P þ respectively represent the

Figure A1.
Interval between LCL ¼ 0
and UCL ¼ K divided into
p ¼ 2 subintervals of
width 2d
 þ
JQME probabilities that statistics Z 2
k and Z k go from state i to state j in one step. In order to
þ
approximate these probabilities, it is assumed that statistic Z 2
k and Z k are equal to H j whenever
15,4 they are in state j.
By definition, the value of p2i;j is
.
if j ¼ 2, . . . , p then
2 2
p2
i;j ¼ PðH j 2 d , Z k # H j þ djZ k21 ¼ H i Þ
356
¼ PðH j 2 d , Z 2 2
k21 þ EðT k Þ 2 T k 2 L # H j þ djZ k21 ¼ H i Þ

¼ PðH j 2 d , H i þ EðT k Þ 2 T k 2 L # H j þ dÞ
¼ PðH i 2 H j þ EðT k Þ 2 L 2 d , T k # H i 2 H j þ EðT k Þ 2 L þ dÞ
¼ F T ðH i 2 H j þ EðT k Þ 2 L þ dÞ 2 F T ðH i 2 H j þ EðT k Þ 2 L 2 dÞ

. where F T ðtÞ is the cumulative distribution function of T k , defined for t $

Downloaded by University of Georgia At 08:34 05 April 2016 (PT)

AðnÞ þ BðnÞlnðCðnÞÞ and equal to (see Castagliola, 2005a):

     
1 t 2 AðnÞ
F T ðtjnÞ ¼ F S 2 2 exp 2 CðnÞ jn
t BðnÞ
where F S 2 ðsjnÞ is the cumulative distribution function for the sample variance.
.
if j ¼ 1 then (takes into account the fact that if Z 2 2
k21 þ EðT k Þ 2 T k 2 L # 0 then Z k ¼ 0,
i.e. in state 1)
2 2
p2
i;1 ¼ PðH 1 2 d , Z k # H 1 þ djZ k21 ¼ H i Þ

þPðZ 2 2
k21 þ EðT k Þ 2 T k 2 L # 0jZ k21 ¼ H i Þ

¼ PðH 1 2 d , H i þ EðT k Þ 2 T k 2 L # H 1 þ dÞ
þPðH i þ EðT k Þ 2 T k 2 L , 0Þ
¼ PðH i 2 H 1 þ EðT k Þ 2 L 2 d , T k # H i 2 H 1 þ EðT k Þ 2 L þ dÞ
þPðT k . H i þ EðT k Þ 2 LÞ
F T ðH i 2 H 1 þ EðT k Þ 2 L þ dÞ 2 F T ðH i 2 H 1 þ EðT k Þ 2 L 2 dÞ
þ1 2 F T ðH i þ EðT k Þ 2 LÞ

By definition, the value of pþ

i;j is:
.
if j ¼ 2; . . . ; p then
þ þ
pþ
i;j ¼ PðH j 2 d , Z k # H j þ djZ k21 ¼ H i Þ
¼ PðH j 2 d , H i þ T k 2 EðT k Þ 2 L # H j þ dÞ
¼ PðH j 2 H i þ EðT k Þ þ L 2 d , T k # H j 2 H i þ EðT k Þ þ L þ dÞ
¼ F T ðH j 2 H i þ EðT k Þ þ L þ dÞ 2 F T ðH j 2 H i þ EðT k Þ þ L 2 dÞ

.
if j ¼ 1 then (takes into account the fact that if Z þ þ
k21 þ T k 2 EðT k Þ 2 L # 0 then Z k ¼ 0,
i.e. in state 1)
 p2 þ þ
i;1 ¼ PðH 1 2 d , Z k # H 1 þ djZ k21 ¼ H i Þ
A new
þPðZ þ þ CUSUM-S 2
k21 þ T k 2 EðT k Þ 2 L # 0jZ k21 ¼ H i Þ
control chart
¼ PðH 1 2 d , H i þ T k 2 EðT k Þ 2 L # H 1 þ dÞ
þPðH i þ T k 2 EðT k Þ 2 L , 0Þ
¼ PðH 1 2 H i þ EðT k Þ þ L 2 d , T k # H 1 2 H i þ EðT k Þ þ L þ dÞ
357
þPðT k , EðT k Þ 2 H i þ LÞ
¼ F T ðH 1 2 H i þ EðT k Þ þ L þ dÞ 2 F T ðH 1 2 H i þ EðT k Þ þ L 2 dÞ
þF T ðEðT k Þ 2 H i þ LÞ
Downloaded by University of Georgia At 08:34 05 April 2016 (PT)

Corresponding author
P. Castagliola can be contacted at: [email protected]

To purchase reprints of this article please e-mail: [email protected]

Or visit our web site for further details: www.emeraldinsight.com/reprints
 This article has been cited by:

1. Babar Zaman, Nasir Abbas, Muhammed Riaz, Muhammad Hisyam Lee. 2016. Mixed CUSUM-EWMA
chart for monitoring process dispersion. The International Journal of Advanced Manufacturing Technology
. [CrossRef]
2. Abdul Haq, Jennifer Brown, Elena Moltchanova. 2015. A New Exponentially Weighted Moving Average
Control Chart for Monitoring Process Dispersion. Quality and Reliability Engineering International 31,
1337-1357. [CrossRef]
3. Abdul Haq, Jennifer Brown, Elena Moltchanova. 2015. New Exponentially Weighted Moving Average
Control Charts for Monitoring Process Mean and Process Dispersion. Quality and Reliability Engineering
International 31, 877-901. [CrossRef]
4. Abdul Haq, Jennifer Brown, Elena Moltchanova, Amer Ibrahim Al-Omari. 2015. Improved Exponentially
Weighted Moving Average Control Charts for Monitoring Process Mean and Dispersion. Quality and
Reliability Engineering International 31:10.1002/qre.v31.2, 217-237. [CrossRef]
Downloaded by University of Georgia At 08:34 05 April 2016 (PT)

5. Jiujun Zhang, Chuan He, Zhonghua Li, Zhaojun Wang. 2015. Likelihood Ratio Test-Based Chart for
Monitoring the Process Variability. Communications in Statistics - Simulation and Computation 00-00.
[CrossRef]
6. Nasir Abbas, Muhammad Riaz, Ronald J. M. M. Does. 2014. Memory-Type Control Charts for
Monitoring the Process Dispersion. Quality and Reliability Engineering International 30:10.1002/
qre.v30.5, 623-632. [CrossRef]
7. Guoyi Zhang. 2014. Improved R and s control charts for monitoring the process variance. Journal of
Applied Statistics 41, 1260-1273. [CrossRef]
8. Nasir Abbas, Muhammad Riaz, Ronald J. M. M. Does. 2013. CS-EWMA Chart for Monitoring Process
Dispersion. Quality and Reliability Engineering International 29:10.1002/qre.v29.5, 653-663. [CrossRef]
9. D.R. PrajapatiDepartment of Mechanical Engineering, PEC University of Technology (Formerly Punjab
Engineering College), Chandigarh, India. 2011. A new approach to monitor the process dispersion.
International Journal of Quality & Reliability Management 28:3, 280-297. [Abstract] [Full Text] [PDF]
10. E. Trovato, P. Castagliola, G. Celano, S. Fichera. 2010. Economic design of inspection strategies to monitor
dispersion in short production runs. Computers & Industrial Engineering 59, 887-897. [CrossRef]