Documente Academic
Documente Profesional
Documente Cultură
net/publication/323258596
CITATIONS READS
0 83
3 authors, including:
Nasir Abbas
King Fahd University of Petroleum and Minerals
39 PUBLICATIONS 445 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
An adaptive EWMA control chart based on CUSUM accumulation error. View project
All content following this page was uploaded by Nasir Abbas on 19 February 2018.
To cite this article: Ishaq Adeyanju Raji, Muhammad Riaz & Nasir Abbas (2018): Robust dual-
CUSUM control charts for contaminated processes, Communications in Statistics - Simulation and
Computation, DOI: 10.1080/03610918.2018.1435804
1. Introduction
Statistical Process Control (SPC) is a set of techniques used to control and monitor the spe-
cial causes of variation in manufacturing or service processes, since the natural variations
are inevitable part of the processes. These techniques are called the magnificent seven of SPC,
namely: cause and effect diagram, Pareto charts, histogram, flow charts, probability plots, scat-
ter diagram and control chart Montgomery (2012). It is truism that amongst these techniques,
control chart is the most efficient and widely used of all. Processes are said to be statistically
In Control (IC) with the presence of natural variations only, but they are considered Out of
Control (OC) due to the existence of unnatural variations. Control charts are classified, (with
respect to the magnitude of variation they are designed to monitor) into memory-less charts
and the memory type. The memory-less chart uses the present information of the process,
truncating the past information in constructing the chart statistic. This makes it less sensitive
to detect small and moderate shift but highly sensitive to large shifts in the process. While
the memory chart makes use of both present and past information of the process in con-
structing the chart statistics and by so doing makes it more sensitive to small and moderate
shifts in the process. An example of the memory-less chart is the Shewhart chart named after
Shewhart (1931). Examples of the memory chart are Cumulative Sum (CUSUM) by Page
(1954) and Exponential Weighted Moving Average (EWMA) by Robert (1959). Both memory
and memory-less charts are used to monitor and control spread and/or location parameters
in any process.
CONTACT Nasir Abbas nasirabbas@yahoo.com Department of Mathematics and Statistics, King Fahad University
of Petroleum and Minerals, Dhahran , Saudi Arabia.
Color versions of one or more of the figures in the article can be found online at www.tandfonline.com/lssp.
© Taylor & Francis Group, LLC
2 I. A. RAJI ET AL.
The importance and effectiveness of the memory charts have attracted the interest of many
researchers, leading to many modifications and improvements in the field. In the quest of
increasing the efficiency and performance of these charts, various type of charts are available
in literature. To refer a few of them; Estimating common-cause sigma proposed by Boyels
(1997), to control both large and small shifts with a single chart. Lucas and Crosier (1990)
introduced a Fast Initial Response (FIR) feature to the classical CUSUM in order to detect
small and moderate shifts faster than the ordinary CUSUM.
In the same vain, Zhao, Tsung, and Wang (2005) designed a dual CUSUM (DCUSUM)
scheme that combines two single-sided CUSUM statistics and monitors them simultaneously
to detect a range of small shifts in a process. This idea is an adjustment and improvement
to that pioneered by Lordens (1971) where a combined set of infinite number of classical
CUSUM statistics are plotted simultaneously against their respective limits. Lordens (1971)
proved the asymptotical optimality feature of this chart by claiming the smallest OC Average
Run Length (ARL) when the IC ARL tends to infinity. Unfortunately, there was no empirical
or simulation illustration in literature to buttress his claim.
The underlying assumptions of all the aforementioned charts are that the environment of
these processes are normal, uncontaminated and free from outliers. In a situation where any
of these assumptions is not met, the efficiency of the corresponding chart becomes question-
able. One of the ways to address this matter is to use some robust charts that can not only
resist the violation of the assumptions but also maintain the efficiency of their performances.
Series of publications are available in literature that incorporate robustness in the charts. Most
of these techniques are based on transforming the data into a non-parametric statistics and
then using the transformed statistics in the CUSUM schemes. (see, Ion and Klaassen (2006);
Li (2015) and Haq (2018)). To mention a few: a robust control chart for estimating disper-
sion is designed by Tatum (1997); Moustafa and Mokhtar (1999) constructed another robust
control chart using the Hodges-Lehman and Shamos-Bickel-Lehman estimators for location
and dispersion parameters, simultaneously. Lately, Riaz, Abbas, and Does (2011) presented a
comprehensive study on control charts whose ideal assumptions are violated. Their properties
and effects on their performances were discussed in details.
On the contrary, Yang, Pai, and Wang (2010) proposed a robust CUSUM control chart
using the sample median as the chart statistic where the observed data is not transformed
to a non-parametric statistic. Muzaffar, Deng, and Rashid (2016) also designed a non-
parametric optimal service pricing. In extension, Nazir et al. (2013) designed a set of
robust CUSUM control charts with some robust location estimators and compared their
effectiveness and robustness for phase II monitoring. Zwetsloot, Schoonhoven, and Does
(2016) also designed a robust point location estimators for the EWMA control chart. (see
Abbas et al. (2017) and Riaz, Khaliq, and Gul (2017)). Nazir et al. (2016) provided a
thorough comparison of Shewhart, CUSUM and EWMA charts under different types of
contaminations. In this article, we propose a set of robust dual CUSUM control chart-
ing schemes with five different location estimators for monitoring phase II samples. We
evaluate these charts based on their run length distribution (ARL and Standard Devia-
tion Run Length (SDRL)). The performances are examined in different environments i.e.
the uncontaminated, the location contaminated and the dispersion contaminated normal
environments.
The remainder of this article is organized as follows: the next section, we discuss the clas-
sical CUSUM control chart versus the construction and design of the DCUSUM control
chart; Section 3 is a detailed description of the proposed estimators and their correspond-
ing DCUSUM chart; the performance evaluation of these proposed charts are presented in
COMMUNICATIONS IN STATISTICS—SIMULATION AND COMPUTATION ® 3
Table A. Nomenclature.
S/N Notations/Symbols Meaning
Section 4; Section 5 buttresses all the proposals with application on a real life dataset; finally,
summary and conclusions appear in Section 6. Also, Table A provides the summary of all the
symbols and notations that are to be used in the rest of this article.
where i is the subgroup number, θˆi is the estimator of the study variable under consideration.
μθˆ = μ0 is the target mean of the study variable, C0+ = C0− = 0 and K = kσθˆ is the reference
value of the CUSUM scheme. The next task is to plot each of these two statistics against H =
hσθˆ . If Ci+ > H for any value of i we conclude that the process has shifted upward and if
Ci− > H at any point i, the process mean is declared to have moved downwards. When n = 1,
θˆi is replaced by the corresponding individual observation Xi . The parameters that define this
chart are K and H, which are to be selected using a predetermined IC Average Run Length
(ARL0 ) of interest. Therefore, a careful selection of these parameters should be put in place
in order not to affect the performance of the chart. Researchers have published a number of
4 I. A. RAJI ET AL.
articles where the selection of these parameters and their effect on the chart were studied in
details.
The two statistics in Eqs (2) and (3) are characterized by different parameters K1 & H1 for
C1i+and K2 & H2 for C2i+ . If C1i+ > H1 and/or C2i+ > H2 at any time i the chart sends a signal
and concludes that process is OC at that time. C10 + +
= C20 = 0. θˆi is a location estimator for ith
subgroup of size n. It is important to note that this chart can be used to monitor spread as
well.
He (Lordens) proved that this chart has the lowest OC ARL when ARL0 tends to infinity,
meeting the restriction that the multiplication of the parameters of all individual CUSUM
charts must be equal i.e. k1 h1 = k2 h2 = k3 h3 = . . . ki hi = h. The third condition above fol-
lows from applying this restriction for the DCUSUM. Taking all these three axioms into con-
sideration, we carefully selected the parameters k1 and k2 from a Monte Carlo simulation to
be 0.12 and 0.5 respectively.
(2) Plot the C1i+ and C2i+ against their control limits as specified in Eqs (2 and 3) follow-
ing the axioms in the inequalities (4). Whenever C1i+ and C2i+ are greater than any
of their respective limits, the process is out of control (OC). Terminate the program
and record the point i which the program went OC. Apply this to each of the five
estimators.
(3) Repeat steps (1 and 2) for 105 runs. Then find the averages standard deviations of all
points i recorded for which the process sent a signal. The averages are termed ARL
and the standard deviations are SDRL. The shifts we considered in this study lie in the
range of [0, 1] both inclusive.
Adopting the algorithm described, we derive the control limit coefficient H1 and H2 for
each of the five estimators under consideration for sample size n = 5, k1 = 0.12, k2 = 0.5
and a specified ARL0 = 370. For our study purposes our derived H1 and H2 is shown in the
table below:
Estimators H1 H2
To accomplish one of the aims of this article, which is examining the robustness of the
DCUSUM charts, we introduce some location and dispersion contaminations to the standard
normal environment and study their behaviours. Therefore, the simulation results we present
are in three folds, the uncontaminated normal environment, the location contaminated and
variance contaminated environments. And each of the contamination is of two levels, the 5%
and 1% contaminations.
Table . ARL and SDRL values for the DCUSUM with k1 = 0.12, k2 = 0.50, and ARL0 = 370.
S hift = . . . . . . . . . . . .
Mean
ARL . . . . . . . . . . . .
SDRL . . . . . . . . . . . .
Median
ARL . . . . . . . . . . . .
SDRL . . . . . . . . . . . .
MR
ARL . . . . . . . . . . . .
SDRL . . . . . . . . . . . .
TM
ARL . . . . . . . . . . . .
SDRL . . . . . . . . . . . .
HL
ARL . . . . . . . . . . . .
SDRL . . . . . . . . . . . .
Figure . The ARL1 curves and corresponding ARL0 values for the normal environment, LCNE and VCNE.
COMMUNICATIONS IN STATISTICS—SIMULATION AND COMPUTATION® 9
Table . ARL and SDRL values DCUSUM % LCNE k1 = 0.12, k2 = 0.50 ARL0 = 370.
S hift = . . . . . . . . . . . .
Mean
ARL . . . . . . . . . . . .
SDRL . . . . . . . . . . . .
Median
ARL . . . . . . . . . . . .
SDRL . . . . . . . . . . . .
MR
ARL . . . . . . . . . . . .
SDRL . . . . . . . . . . . .
TM
ARL . . . . . . . . . . . .
SDRL . . . . . . . . . . . .
HL
ARL . . . . . . . . . . . .
SDRL . . . . . . . . . . . .
drawn from a LCNE, if it has (ϕ)100% observations from N(μ0 , σ02 ) and (1 − ϕ)100% obser-
vations from N(μ0 + ωσ0 , σ02 ) where −∞ < ω < ∞. In this section, we study the run length
distribution of the DCUSUM charts with ϕ = 0.05 & 0.01 and ω = 4, with the proposed esti-
mators. In Table 2 for ϕ = 0.05 and Table 3 for ϕ = 0.01 respectively, it is obvious that all
the five estimators are not adequate for the target ARL0 as they are drastically lower than
ARL0 = 370. Though, despite the poor performances, we can still compare the estimators
based on their resistance and robustness to the disturbances.
In both cases, when median turns to have the highest ARL0 , followed by tri-mean and
Hodges-Lehmann. The remaining estimators are therefore not comparable with the first three
due to their low performances. It should be noted that, the substantial drop in the ARL0 values
of these estimators is because they are measures of location and the disturbance is also on
the location parameter. This can be justified by the increase in the values of ARL 0 s for all
the estimators, as the level of disturbance ϕ decreases. In Table 3, when ϕ = 0.01, the ARL 0 s
values of the estimators are although not comparable with that of the uncontaminated normal
environment, but substantially greater compared to when ϕ = 0.05. This poor performance
in the ARL0 , thwarted the aim of analysing the ARL 1 s results and comparing the estimators
based on these values (cf. Tables 2-3 and Figure 1).
Table . ARL and SDRL values DCUSUM % LCNE k1 = 0.12, k2 = 0.50 ARL0 = 370.
S hift = . . . . . . . . . . . .
Mean
ARL . . . . . . . . . . . .
SDRL . . . . . . . . . . . .
Median
ARL . . . . . . . . . . . .
SDRL . . . . . . . . . . . .
MR
ARL . . . . . . . . . . . .
SDRL . . . . . . . . . . . .
TM
ARL . . . . . . . . . . . .
SDRL . . . . . . . . . . . .
HL
ARL . . . . . . . . . . . .
SDRL . . . . . . . . . . . .
10 I. A. RAJI ET AL.
Table . ARL and SDRL values DCUSUM % VCNE k1 = 0.12, k2 = 0.50 ARL0 = 370.
S hift = . . . . . . . . . . . .
Mean
ARL . . . . . . . . . . . .
SDRL . . . . . . . . . . . .
Median
ARL . . . . . . . . . . . .
SDRL . . . . . . . . . . . .
MR
ARL . . . . . . . . . . . .
SDRL . . . . . . . . . . . .
TM
ARL . . . . . . . . . . . .
SDRL . . . . . . . . . . . .
HL
ARL . . . . . . . . . . . .
SDRL . . . . . . . . . . . .
Table . ARL and SDRL values DCUSUM % VCNE k1 = 0.12, k2 = 0.50 ARL0 = 370.
S hift = . . . . . . . . . . . .
Mean
ARL . . . . . . . . . . . .
SDRL . . . . . . . . . . . .
Median
ARL . . . . . . . . . . . .
SDRL . . . . . . . . . . . .
MR
ARL . . . . . . . . . . . .
SDRL . . . . . . . . . . . .
TM
ARL . . . . . . . . . . . .
SDRL . . . . . . . . . . . .
HL
ARL . . . . . . . . . . . .
SDRL . . . . . . . . . . . .
COMMUNICATIONS IN STATISTICS—SIMULATION AND COMPUTATION ® 11
Figure . The control charts of DCUSUM scheme for the five estimators under normal environment.
set which consists of 30 sample points with the sample size n = 5, is the waiting time in min-
utes of patients for a colonoscopy procedure in a regional health center. Jones-Farmer, Jordan,
and Champ (2009). The original data set without contamination is presented in Figure 2(a),
while the subsequent Figure 2(b–f) show the DCUSUM control chart of respective charts.
For all the estimators, their respective charts behave quite well under the uncontaminated
normal environment as we can see all the points of the charts falling within the control
limits.
In the location contaminated environments, comes the survival of the fittest rule. As we can
see from Figures 3(b) and 3(d) that both mean and midrange could not withstand the location
contamination applied to the original data set. At sample point 14, the midrange chart went
out of its limit till point 17, which makes it the least robust of all the estimator. Next to it in
bad performance, is the mean. It becomes OC process at sample point 15 and maintains its
limit back immediately at point 16.
Out of the five estimators, the median chart proves to be the best, as we can see from
the chart, followed by trimean and Hodges-Lehmann estimators hierarchically. The real-life
example as just buttressed the simulation result produced in that regard.
The same scenario occurs under the variance contaminated environment sample, the
median retains its position as the best of all, followed by trimean and Hodges-Lehmann
chart respectively. Though the performance of the median and mean charts under the vari-
ance contamination seems to be better than that of location contamination, but relatively,
12 I. A. RAJI ET AL.
Figure . The control charts of DCUSUM scheme for the five estimators under LCNE.
the two charts are still in an OC state at some points, as it is obvious in their respective
charts (cf. Figure 4).
Figure . The control charts of DCUSUM scheme for the five estimators under VCNE.
because the estimators are estimating location parameters. Despite this worst performance,
the TM and HL estimators’ charts outperform others, based on ARL0 and ARL1 examination.
Though the ARL0 of the median seems to be the best, but the median chart performance is
slow compared to TM and HL performances in detecting small shifts in the process which is of
paramount. The performance of these charts under dispersion contamination environments
depicts the efficiency and robustness of the TM and HL estimators better.
In general, this study substantially proves that the DCUSUM chart is preferred to the
ordinary CUSUM, and TM estimator is the best amongst the five estimators under all
circumstances.
Acknowledgments
The authors Muhammad Riaz and Nasir Abbas would like to acknowledge the support provided by
Deanship of Scientific Research (DSR) at King Fahd University of Petroleum and Minerals for funding
this work through project No. IN161037.
Funding
King Fahd University of Petroleum and Minerals (IN161037).
14 I. A. RAJI ET AL.
References
Abbas, S., and R. Fried. 2017. Control charts for the mean based on robust two-sample tests. Journal of
Statistical Computation and Simulation 87 (1):138–55. doi:10.1080/00949655.2016.1194839.
Abbas, T., S. Ahmad, M. Riaz, and Z. Qian. 2017. A bayesian way of monitoring the linear pro-
files using CUSUM control charts. Communications in Statistics-Simulation and Computation.
doi:10.1080/03610918.2017.1375520.
Boyles, R. A. 1997. Estimating common-cause sigma in the presence of special causes. Journal of Quality
Technology 29 (4):381–95.
Haq, A. 2018. A new nonparametric synthetic EWMA control chart for monitoring process mean. Com-
munications in Statistics-Simulation and Computation. doi:10.1080/03610918.2017.1422750.
Hodges, J. L. Jr., and E. L. Lehmann. 1963. Estimates of location based on rank tests. Annals of Mathe-
matical Statistics 34:598–611. doi:10.1214/aoms/1177704172.
Ion, R. A., and C. A. J. Klaassen. 2006. Nonparametric shewhart control charts. Journal of Nonparametric
Statistics 17 (8):971–88. doi:10.1080/10485250500393271.
Jones-Farmer, L. A., V. Jordan, and C. W. Champ. 2009. Distribution-free phase I charts for subgroup
location. Journal of Quality Technology 41 (3):304–16. doi:10.1080/00224065.2009.11917784.
Li, J. 2015. Nonparametric multivariate statistical process control: A hypothesis testing-based approach.
Journal of Nonparametric Statistics 27 (3):384–400. doi:10.1080/10485252.2015.1062889.
Lorden, G. 1971. Procedures for reacting to a change in distribution. Annals of Mathematical Statistics
42 (6):1897–908. doi:10.1214/aoms/1177693055.
Lucas, J. M., and M. S. Saccucci. 1990. Exponentially weighted moving average control schemes: Prop-
erties and enhancements. Technometrics 32 (1):1–12. doi:10.1080/00401706.1990.10484583.
Montgomery, D. C. 2012. Introduction to statistical quality control. (7th ed.), New York: Wiley.
Moustafa, B. A., and A. Mokhtar. 1999. New robust statistical process control chart for location. Quality
Engineering 12 (2):149–59. doi:10.1080/08982119908962572.
Muzaffar, A., S. Deng, and A. Rashid. 2016. Nonparametric optimal service pricing: A
simulation study. Quality Technology and Quatitative Management 14 (2):142–55.
doi:10.1080/16843703.2016.1208487.
Nazir, H. Z., N. Abbas, M. Riaz, and R. J. M. M. Does. 2016. A comparative study of memory-type
control charts under normal, contaminated and non-normal environments. Quality and Reliability
Engineering International 32 (4):1347–56. doi:10.1002/qre.1835.
Nazir, H. Z., M. Riaz, R. J. M. M. Does, and N. Abbas. 2013. Robust CUSUM control charting. Quality
Engineering 25 (3):211–24. doi:10.1080/08982112.2013.769057.
Page, E. S. 1954. Continuous inspection schemes. Biometrika 41:100–15. doi:10.1093/biomet/41.1-
2.100.
Riaz, M., N. Abbas, and R. J. M. M. Does. 2011. Improving the performance of CUSUM charts. Quality
and Reliability Engineering International 27 (4):415–24. doi:10.1002/qre.1124.
Riaz, M., Q. U. A. Khaliq, and S. Gul. 2017. Mixed tukey EWMA-CUSUM control chart
and its applications. Quality Technology & Quantitative Management 14 (4):1–34.
doi:10.1080/16843703.2017.1304034.
Roberts, S. W. 1959. Control chart tests based on geometric moving averages. Technometrics 1 (3):
239–50. doi:10.1080/00401706.1959.10489860.
Schoonhoven, M., H. Z. Nazir, M. Riaz, and R. J. M. M. Does. 2011. Robust location estimators for the
X control chart. Journal of Quality Technology 43:363–79. doi:10.1080/00224065.2011.11917870.
Shewhart, W. A. 1931. Economic control of quality of manufactured product. New York, NY: D. Van
Nostrand Company.
Tatum, L. G. 1997. Robust estimation of the process standard deviation for control charts. Technometrics
39:127–41. doi:10.1080/00401706.1997.10485078.
Yang, Y. L., S. Pai, and Y. Wang. 2010. A novel CUSUM median control chart. In International multi-
conference of engineers and computer scientists proceedings, Hong Kong, March 17–19, 2010.
Zhao, Y., F. Tsung, and Z. Wang. 2005. Dual CUSUM control schemes for detecting a range of mean
shifts. IIE Transactions 37 (11):1047–57. doi:10.1080/07408170500232321.
Zwetsloot, I. M., M. Schoonhoven, and R. J. M. M. Does. 2016. Robust point location estima-
tors for the EWMA control chart. Quality Technology & Quantitative Management 13 (1):29–38.
doi:10.1080/16843703.2016.1139845.