Accepted Manuscript

An EWMA monitoring scheme with a single auxiliary variable for industrial

processes

Ridwan A. Sanusi, Muhammad Riaz, Nurudeen A. Adegoke, Min Xie

PII: S0360-8352(17)30469-2
DOI: https://doi.org/10.1016/j.cie.2017.10.001
Reference: CAIE 4938

To appear in: Computers & Industrial Engineering

Received Date: 20 April 2017

Revised Date: 30 September 2017
Accepted Date: 2 October 2017

Please cite this article as: Sanusi, R.A., Riaz, M., Adegoke, N.A., Xie, M., An EWMA monitoring scheme with a
single auxiliary variable for industrial processes, Computers & Industrial Engineering (2017), doi: https://doi.org/
10.1016/j.cie.2017.10.001

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers
we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and
review of the resulting proof before it is published in its final form. Please note that during the production process
errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
 AN EWMA MONITORING SCHEME WITH A SINGLE AUXILIARY VARIABLE FOR

INDUSTRIAL PROCESSES

Ridwan A. Sanusi1*, Muhammad Riaz2, Nurudeen A. Adegoke3, and Min Xie1

1
Department of Systems Engineering and Engineering Management, City University of Hong Kong,

Kowloon, Hong Kong, SAR.

2
Department of Mathematics and Statistics, King Fahd University of Petroleum and Minerals, Saudi

Arabia.

3
Institute of Natural and Mathematical Sciences, Massey University Auckland, New Zealand.

Corresponding author: Ridwan A. Sanusi; email: [email protected]

ABSTRACT

When using control charts to monitor manufacturing processes, Shewhart control chart is known to be

useful for detecting transient shifts, while the EWMA and CUSUM charts are useful for detecting

persistent shifts. The efficiency of EWMA chart in monitoring location parameter can be improved by

using an auxiliary variable that is closely related to the variable of interest. In this paper, an EWMA-type

scheme using ratio estimator is developed to further increase the effectiveness of the classical EWMA

chart in monitoring the location parameter. The proposed procedure outperforms the classical EWMA and

even the mixed EWMA-CUSUM chart, especially when there is a strong positive relationship between

the variable of interest and the auxiliary variable. Finally, a real data set is used to show the

implementation procedures of the proposed chart.

Keywords: ARL; auxiliary variable; control chart; EWMA; location parameter; ratio-estimator.

1 Introduction

Statistical process control (SPC) is a data-driven statistical method for monitoring and controlling a

process in manufacturing industry. The Shewhart control chart (Shewhart 1924) is an example of the

memoryless-type control chart. The two most commonly used memory type control charts are the

exponentially weighted moving average (EWMA) control chart and the cumulative sum (CUSUM)

control chart, proposed by Page (1954) and Roberts (1959), respectively. The memory control charts

utilize previous information with present information to yield a better result for detecting small to

moderate shifts, unlike Shewhart-type charts that use only the current information. Research works

related to the Shewhart-type charts and CUSUM-type charts can be found in Hawkins and Olwell (1998),

Mukherjee and Sen (2015), Chong et al. (2016), Li et al. (2016), Sanusi et al. (2016), and Lombard et al.
(2017). Different EWMA-type schemes have also been proposed in the literature. Noorossana et al. (2015)

investigated the joint effect of non-normality and parameter estimation on EWMA chart. Also, Tamirat

and Wang (2015) introduced an acceptance sampling plan scheme based on an EWMA statistic. In the

case of unknown parameters, EWMA median control chart with estimated parameters was introduced to

monitor the location parameter of a normal process (Castagliola et al. 2016). Also, Zhou et al. (2016)

suggested a one-sided EWMA scheme with varying sample sizes for monitoring rare events. Furthermore,

EWMA control chart found early applications in economics (Muth 1960) and in inventory control and

forecasting (Dushman et al. 1962). For more works on the improvement of EWMA chart, interested

readers can see Liu et al. (2007), Sheu et al. (2009), Teh et al. (2011), Xie et al. (2011), Nishimura et al.

(2015), and Zwetsloot et al. (2016).

The traditional EWMA chart monitors the process mean (say Y ) of a process distribution. The Y is

computed using the famous simple random sampling (SRS) approach. However, in the presence of an

auxiliary variable (X) that is closely related to the study variable (Y), Y can be estimated more efficiently.

Consequently, Cochran (1940) used the advantage of auxiliary information to propose a ratio mean

estimator  y r  for estimating the population mean of Y. Choudhury and Singh (2012) noted that y r is

most effective when there is a positive linear relationship (which passes through the origin) between Y

and X and the mean square error (MSE) of Y is proportional to X. Murthy (1964) suggested the use of y r

when YX CY CX  0.5 , where YX , CY , and C X are, respectively, the correlation coefficient between

Y and X, the coefficient of variation of the study variable (Y), and the coefficient of variation of X. Also,

Adebola et al. (2015) introduced an efficient estimator, with cum-dual ratio estimator as intercept, for

estimating the population mean.

Riaz (2008a) popularized the idea of using an auxiliary variable at the stage of estimating the plotting

statistic of a monitoring chart. He presented a Shewhart-type scheme based on a regression-type estimator

for detecting shifts in process variability and also showed the superiority of the scheme over the

traditional Shewhart control chart. In addition, a Shewhart-type control chart was suggested by Riaz

(2008b) for monitoring process location. The chart is based on the regression mean estimator and it was

shown that it outperforms the Shewhart’s X - bar chart and the cause-selecting charts. The work was later

extended to the EWMA set-up for detecting small to moderate shifts in the process mean (Abbas et al.

2014). It was revealed that the chart outperforms other existing univariate and bivariate charts Also, some

efficient estimators with an auxiliary variable are used to improve the performance of the combined

Shewhart-CUSUM chart for detecting both small and large shifts in the location parameter of a process

(Sanusi et al. 2017).

In this article, an auxiliary variable is introduced at the estimation stage to monitor the location parameter

of a process distribution. The proposed chart, denoted as MrEWMA, is an EWMA-type control chart

based on the ratio mean estimator. This motivation further enhances the sensitivity of the control chart,

especially in detecting shifts of small magnitudes. The average run length (ARL) approach is used to

evaluate the performance of the chart. Also, other performance measures such as the standard deviation of

the run length (SDRL), the extra quadratic loss (EQL), the relative average run length (RARL), and the

performance comparison index (PCI) are considered. The management perspective of the proposed

scheme is also briefly discussed.

One of the three activities for the successful execution of an effective management of a process is quality

control. This requirement ensures that products are up to standard through continuous improvement. The

proposed MrEWMA scheme would further helps to constantly improve the performance of a product,

which will leads to a long-term reward for an industry. In addition, the early detection of shifts would

avoid mass inspection in controlling quality, since a good quality is achieved by preventing defective

items, instead of inspecting the items for bad products. Moreover, industry with modern method of

improving products quality, and can demonstrate process capability and control, has an edge over other

competitors. These are in agreement with the Deming philosophy in improving management strategies

(Montgomery 2009).
The rest of this article is arranged as follows: The statistical preliminaries of the proposed scheme; the

structural framework of the classical EWMA chart including its plotting statistic, control limits, and ARL;

and the design of the proposed chart are presented in Section 2. The performance evaluations and the

major findings of the proposed MrEWMA chart are provided in Section 3. The comparison of the

MrEWMA control chart with its existing counterparts is given in Section 4. A real-life illustrative

example is given in Section 5. Finally, the summary and conclusion of the findings are provided in

Section 6.

2.0 The proposed monitoring scheme

The motivation of this work is to enhance the sensitivity of EWMA control chart in detecting shifts in the

location parameter of a control process. This is achieved by introducing an auxiliary variable, in the form

of a ratio estimator, to the charting scheme. In the next subsections, the mathematical preliminaries of the

proposed scheme are explained, followed by a brief description of the classical EWMA, and then, the

construction of the proposed MrEWMA chart.

2.1 Statistical Preliminaries

Let X represents an auxiliary variable which is positively correlated with the study variable (Y) of a

control process. The extra information provided by the auxiliary variable can increase the efficiency in

 
estimating the population mean Y . One of the ways of integrating the extra information is to use the

ratio mean estimator defined as:

y r   y x X , (2.1)

The estimator is biased, that is, E( yr )  Y  B( yr ) , where E ( y r ) is the expectation of the ratio mean

estimator ( yr ) and B( yr ) is the bias of y r . The B( yr ) is defined as:

 
B( yr )  Y C X2  YX CY C X  n, (2.2)

For the Case-U, B( yr ) can be estimated as:

 
Bˆ ( yr )  y c X2  rYX cY c X n , (2.3)

where C X   X X , C X   Y Y , c X  s X x , and cY  sY y .

Furthermore, the mean squared error of y r is given as

 
MSE ( yr )  Y 2 CY2  CX2  2YX CY CX n , (2.4)

which can be estimated as

mse( yr )  y 2 c y2  c x2  2ryx c y c x  n . (Singh and Mangat 1996) (2.5)

Note that sampling with replacement and approximation up to the first order are assumed. Also, the

choice of  YX that gives an efficient estimation is determined by the condition  YX  C X 2CY (Singh

and Mangat 1996). More often, C X  CY , which implies that C X CY  1 . Consequently, we have

YX  0.5 . As a result, in the presence of an auxiliary variable, it is more efficient to apply the ratio mean

estimator instead of the usual SRS mean estimator if YX  0.5 (Murthy 1964). Murthy (1964) noted that

in the case of large samples and when both Y and X are both positive or negative, the product estimator,

the usual mean estimator, and the ratio estimator should be used when  1  YX   C X 2CY ,

 C X 2CY  YX  C X 2CY , and C X 2CY  YX  1 , respectively. If either Y or X is negative,

the conditions for the product and ratio estimators will be reversed. In Figure 1, the broken horizontal line

is the standard error (SE) of the usual SRS mean, while the curve is the SE of the ratio estimator which

varies

PLEASE INSERT FIGURE 1

with different YX values. The two lines meet at point C X 2CY . It is shown that the SE of the ratio

estimator is less than the SE of the usual SRS mean when C X 2CY  YX  1 . In this research, it is

assumed that both Y and X are greater than zero. Hence, only positive values of  YX (  YX  0.25,

0.50, 0.75, and 0.95) are considered. Furthermore, the usual SRS estimator may be better than the ratio

estimator when near proportionality between Y and X does not exist (Cochran 1940). This happens when

the regression line of Y on X passes through a point on Y-axis that is not close to the origin. Before

constructing the proposed control chart, the classical EWMA will be introduced briefly.

2.2 The classical EWMA chart

The classical EWMA control chart is a memory-type chart that detects persistent shifts in a control

process. It is frequently used for monitoring single observations, though it can also be used for monitoring

process mean of rational subgroups of size n  1 (Montgomery 2009). Consequently, the procedures for

constructing an EWMA chart with rational subgroup is explained. Assume that the study variable (Y) of a

control process is from a normal distribution with population mean Y and population standard deviation

 Y , that is, Y ~ N Y ,  Y2  . For the case of monitoring the process mean y  i 1Yi n , where
n

 
y ~ N Y ,  Y2 n , the plotting statistic of the classical EWMA scheme for the tth rational subgroup is

given as

Wt  yt  1   Wt 1 , t  1, 2,  (2.6)

where  is the smoothing parameter such that 0    1 . The structure becomes a Shewhart X -bar

chart whenever  = 1. Also, the sensitivity of the chart to small shifts is inversely related to the

smoothing parameter. Since the aim of the EWMA-type charts is to detect small shifts,  = 0.1 or 0.2 is

often recommended in the literature. The quantity Wt 1 denotes the past value, while W0 is the initial

value. If the process parameters are known (often referred to as the Case-K), W0 is taken to be the target
or in-control mean (say Y0 ). For an in-control process, the mean and the variance of the EWMA statistic

(Roberts 1959) are, respectively,

E Wt   Y0 and V (Wt )   Y2 n 2   1    .

2t
(2.7)

For the case of unknown process parameters (often referred to as the Case-U), the parameters Y0 and  Y2

are estimated from preliminary samples.

The center line (CL), the upper control limit (UCL), and the lower control limit (LCL) of the EWMA

chart at time t can be defined as follow:


LCLt  Y0  L  Y n   2   1  1     ,
2t
(2.8)

CL  Y0 , (2.9)


UCLt  Y0  L  Y n   2   1  1     .
2t
(2.10)

The LCLt and UCLt have subscript t because the chart is time-varying. Asymptotically, the limits

become constant as time increases. Hence, the asymptotic limits are given in equations (2.11) and (2.12).

The constant L regulates the width of the control limits. Its value is selected to fix a particular in-control

ARL (ARL0).


LCL  Y0  L  Y n   2    , (2.11)


UCL  Y0  L  Y n   2    . (2.12)

Table 1 gives the ARL values of the classical EWMA control chart with time-varying limits (Steiner

1999). In the Table, δ is the amount of shift defined as   Y1  Y0  Y 

n , where Y0 and Y1 are the

in-control and the out-of-control mean, respectively.

 PLEASE INSERT TABLE 1

Many authors have proposed different modifications of the memory-type control charts that give smaller

out-of-control ARL values for a particular ARL0. Some of the modifications are the use of auxiliary

variable and the introduction of first initial response (Lucas and Crosier 1982; Rhoads et al. 1996; Sanusi

et al. 2016), among others.

2.3 Design of the proposed MrEWMA control chart

In this study, the values of Y and X are gotten in paired form for each sample observation and the

population mean and variance of X are assumed to be known. More so, Y and X are assumed to follow a

bivariate normal distribution, that is, (Y , X ) ~ N 2 (Y , X ,  Y ,  X , YX ) where N2 represents the bivariate

normal distribution. Based on the ratio estimator in equation (2.1), the plotting statistic of the proposed

MrEWMA chart is given as:

Wt  y rt  1   Wt 1 , t  1, 2,  (2.13)

Based on the statistical properties of yr , the control limits for MrEWMA chart are defined as:


LCLt  Y0  L MSE y r  2    1  1   
2t
 (2.14)

CL  Y0 (2.15)


UCLt  Y0  L MSE y r  2    1  1   
2t
 (2.16)

Equations (2.14, 2.15 and 2.16) represent the time-varying control limits. For constant limits, they can be

written as:

LCL  Y0  L MSE y r  2    (2.17)

UCL  Y0  L MSE y r  2    . (2.18)

More argument about the efficiency of the proposed chart over the classical EWMA chart is given in

Appendix A, where it is proved that the MSE of the plotting statistic of MrEWMA is less than that of the

plotting statistic of the classical EWMA when C X 2CY  YX  1 .

3.0 Performance measures

Following Steiner (1999), a comprehensive assessment of the proposed MrEWMA chart in terms of ARL

and SDRL is performed with the aid of R software (R core team 2016). ARL is the average number of

samples plotted until the first out-of-control signal (Montgomery 2009), while SDRL is the standard

deviation of the number of samples plotted until the first out-of-control signal. ARL0 is the ARL when a

process is in-control, while ARL1 is the ARL when a process is out of control. It is expected that ARL0

has a large value while ARL1 has a small value (Abbasi et al., 2012; Ahmad et al., 2014). Knoth (2007)

gave an accurate ARL calculation of EWMA chart for the simultaneous monitoring of normal mean and

variance. To show that the sensitivity of EWMA-type chart decreases with increase in  , small and large

values of  (   0.05, 0.25, and 0.50) are considered, and the corresponding values of L that fix ARL0

to be 500 are searched through a Monte Carlo search algorithm. This is to ensure a fair comparison of the

MrEWMA chart with other existing charts of same ARL0. In addition, the run length properties of the

proposed chart, such as average, standard deviation and median, are investigated for different shift sizes,

say δ = 0, 0.25, 0.50, 0.75, 1, 1.50, 2, 2.50, 3, 4, and 5 (Table 2). The results are computed using the

time-varying limits.

Some authors have raised concern about the use of ARL for comparing different charts because it is based

on the effectiveness of charts at a specific shift point (Wu et al. 2008). Consequently, we further measure

the efficiency of the proposed chart using performance measures that give the performance of a chart over

all the domain of shifts considered. The performance measures are:

 Extra Quadratic Loss (EQL): This measures the effectiveness of a chart over all the domain of the

process shifts. It is the weighted average ARL over the entire shifts of a charting structure. It

makes use of the square of shift (δ2) as weight.

Let  be the loss subject to an out-of-control state,

   f  M   (3.1)

where f = EQL per unit

= (quadratic loss with respect to  ) – (quadratic loss when process is in-control).

M = number of units produced when the process is in an out-of-control state.

 
f    Qc  02     0   Qc 02  Qc 02 2
2
(3.2)

where Qc = constant cost for an individual process

M    b  j  ARL    0.5, b = rate of production, and j = regularly spaced time interval.

M   assumes that the random time of a mean shift follows a uniform distribution (Reynolds et
al., 1990). Considering all shifts in the process domain,

  d
1
EQL    max
 max   min min (3.3)
 b  j  Qc   2
0  max   min    min  ARL  d  
1  max 2
 2
max   max  min   2
min  6
Without loss of generalization and for simplicity purpose, b  j  Qc   02 and

 2
max   max  min   min
2

6 are constants and can be eliminated from equation (3.3) since they do

not affect the performance comparison and the optimization solution (Wu et al. 2009). Hence, we

have:

EQL   max   min    max  2 ARL  d

1
min
(3.4)

Note that a chart with a high EQL value is said to be inferior to a chart with a low EQL value.
  Relative Average Run Length (RARL): It uses ARL to determine the global performance of a

chart with respect to a benchmark chart. The benchmark chart has a RARL value of one.

Apparently, a chart with RARL greater than one is said to be inferior to the benchmark chart (Wu

et al. 2009). The classical EWMA chart is used as the benchmark chart in this research.

1 ARL  
RARL   max d (3.5)
 max   min min
ARLbmk  

Where ARLbmk   is the ARL of the benchmark chart when there is a shift of size  .

Both EQL and RARL can be calculated using numerical integration.

 Performance Comparison Index (PCI): This is the ratio of the average extra quadratic loss (AEQL)

of a particular chart to the AEQL of a benchmark chart. Ou et al. (2012) used the best chart as the

benchmark. For consistency, the classical EWMA is used as the benchmark in our case. A chart

of PCI equals to one performs equally as the benchmark chart, while a chart with PCI lesser

(greater) than one performs better (worse) than the benchmark chart. The mathematical

expression of the PCI is given as:

PCI  AEQL AEQLbmk

But AEQL   max   min    max  2 ATS  d and ATS  ARL  h , where ATS is the
1 
min

average time to signal and h is a regularly spaced time interval. If h = 1, ATS  ARL and

AEQL  EQL . Hence,

PCI  EQL EQLbmk (3.6)

3.1 Step-by-step algorithm of the MrEWMA

The illustrative procedures of the proposed chart for the Case-K are listed.

1. Determine the in-control parameters (Y , X ,  Y ,  X , YX ) and the appropriate constants  and

L.
 2. Construct the control limits using equations (2.14, 2.16) for the time-varying limits, or equations

(2.17, 2.18) for the constant limits.

3. Obtain the t th test sample xti , yti  of size n in the course of monitoring, where t  1, 2,

4. Compute the ratio mean estimate yrt using equation (2.1).

5. Compute the plotting statistic Wt  y rt  1   Wt 1 .

6. Plot Wt against the control limits.

7. If Wt falls within the control limits, the process is declared to be in-control, then the process goes

back to Step 3 for the monitoring of the next test sample. Alternatively, if Wt falls outside the

control limits, the process is said to have shifted to an out-of-control state. Consequently, the

monitoring process terminates and the cause of the shift in the process mean is investigated.

For the Case-U, the in-control parameters can be determined from a phase-1 analysis.

3.2 Major findings of the proposed MrEWMA chart

The findings of the proposed MrEWMA chart are summarized as follow:

(1) The introduction of auxiliary variable (in the form of ratio estimator) to the plotting statistic of the

classical EWMA control chart improves the performance of the chart for values of  XY  0.50 (Table 2).

Also, for  XY = 0.25, the ARL1 values of the MrEWMA chart are slightly greater than ARL1 values of

the classical EWMA chart. Hence, it is less effective in monitoring the location parameter of a control

process. In addition, for  XY  0.50 , the MrEWMA chart performs better than the classical EWMA

chart. In fact, the performance of the proposed chart increases as  XY increases. This is evident from the

lower ARL1 values of the MrEWMA chart (Figure 2).

PLEASE INSERT FIGURE 2

(2) For different sensitivity parameter values (that is,  0.05, 0.25, and 0.50), it is shown that the

MrEWMA chart is more sensitive to small and moderate shifts than the classical EWMA chart (Table 2).
(3) The MrEWMA chart is ARL unbiased for all values of  , L and  XY , i.e., ARL0 is always greater
than ARL1 for any choice of δ (Table 2). Discussion on biased ARL charts can be found in Knoth and

Morais (2015).

(4) The out-of-control ARL and SDRL values decrease quickly for all δ and the decrease is more

significant for  XY = 0.95 (Table 2).

(5) When  = 0, there is no significant difference between the ARL and the SDRL values. This is in

agreement with the geometric distribution property of run length. Also, when  > 0, the ARL and the

SDRL values are directly proportional but indirectly proportional to the shift   in the location

parameter (Table 2).

(6) The chart detects very large shifts at the first out-of-control sample. Thus, the out-of-control ARL and

SDRL values tend to 1 and 0, respectively, as the shift increases (Table 2).

(7) The ARL values are always greater than the median of the run length (except for some sampling error).

This implies that the run length distribution is skewed to the right.

PLEASE INSERT TABLES 2 – 3

PLEASE INSERT FIGURE 3

(8) The EQL, RARL, and PCI, which explain the overall performance of a chart over all range of shifts,

show that the classical EWMA is inferior to the proposed chart for  XY  0.50 (See Table 3). The

classical EWMA chart is used as the benchmark chart. A chart with lower EQL, RARL, and PCI values

than the benchmark chart is said to be superior to the benchmark chart. It is shown in Figure 3 that the

proposed chart possesses all these attributes when  XY  0.50 . These further show the supremacy of the

proposed chart over the classical EWMA chart when the correlation between the study variable and the

auxiliary variable is not weak.

4.0 Comparisons with other location monitoring charts

A general comparison of the proposed scheme with the classical EWMA, MxEWMA, and the Mixed

EWMA-CUSUM charts is provided in this section, by presenting a one by one comparison of MrEWMA

chart with its existing counterparts.

4.1 MrEWMA vs. classical EWMA

Section 2 broadly explains the classical EWMA chart and Table 1 shows the ARL of the classical EWMA

chart with time-varying limits. For all the values of  considered, it is observed that the classical EWMA

chart outperform the proposed chart when there is a weak correlation between the study variable and the

auxiliary variable. Contrarily, when  XY  0.50 , the proposed MrEWMA chart outperforms the classical

EWMA chart for all the values of the process mean shift. This is as a result of the lower ARL1 values of

the MrEWMA chart for all the  and δ considered. In fact, the supremacy of the proposed chart is more

significant when  XY = 0.95 (Table1 vs. Table 2). This is expected since the proposed chart will perform

best for a very high positive correlation coefficient between the study variable and the auxiliary variable.

Furthermore, it is worthy to note that a special case of the MrEWMA chart is the classical EWMA chart

when  XY = 0. Also, the L values of the MrEWMA chart are slightly different from that of the classical

EWMA chart because the values are adjusted for the bias in the ratio estimator.

4.2 MrEWMA vs. MxEWMA

The MxEWMA proposed by Abbas et al. (2014) uses auxiliary information in the form of regression

estimator to improve the classical EWMA chart. It slightly outperforms the proposed chart. However, the

MrEWMA chart competes closely for higher  XY (Table 2 vs. Table 4). The supremacy of the

MXEWMA chart is not surprising because the regression estimator has higher efficiency (or lower MSE)

than the ratio

PLEASE INSERT TABLE 4

estimator. Nevertheless, the ratio estimator is often preferred to the regression estimator because of the

simplicity of its structure and its merits when the estimates of total (of the study variable) or the ratio of

totals (of the study variable to the auxiliary variable) are required (Särndal et al. 2003).

4.3 MrEWMA vs. mixed EWMA-CUSUM (MEC)

The concept of the mixed EWMA-CUSUM chart (denoted as MEC hereafter) was proposed to further

increase the sensitivity of the EWMA and CUSUM charts to small shifts. The design structure of the

scheme can be found in Abbas et al. (2013). When  = 0.25, the MEC chart outperforms the MrEWMA

chart in detecting small shifts   1 in process mean when the correlation between X and Y is not strong

(that is,  XY  0.50), but inferior for detecting large shifts   1 in the location parameter of a process.

However, the proposed MrEWMA chart outperforms the MEC chart for all shifts when  XY  0.50

(Figure 4(a)).

PLEASE INSERT FIGURE 4

When  = 0.25, the MEC only performs better than the MrEWMA chart in detecting shifts of   1.50,

  1, and   0.50 when  XY  0.25,  XY  0.50, and  XY  0.75, respectively, but inferior to the

MrEMWA chart in detecting all sizes of shift when  XY  0.95 (Figure 4(b)). Hence, it can be generally

inferred that the MEC chart only performs better than the proposed MrEWMA chart in detecting small

shifts when the correlation between X and Y is not strong.

5.0 Illustrative example

In this section, we provide an illustrative example to show the implementation of our proposed chart

using a practical example. For this purpose, we have considered the dataset based on the nonisothermal

continuous stirred tank chemical reactor model (namely CSTR process) originally proposed by Marlin

(1995), and has been widely used as benchmark in fault detection and diagnosis (Yoon and MacGregor

2001; Shi et al. 2013). The CSTR process comprises of nine process variables, among which we have
considered Outlet temperature as the study variable (Y) and cooling water temperature as the auxiliary

variable (X). Details of other variables may be seen in Yoon and MacGregor (2001) and Shi et al. (2013).

The said data set originally contains 1024 values that have been collected on sampling interval of half

minute. The first 512 data points which are in-control are used to estimate the population parameters.

These estimates come out to be y , x , sy , sx and ryx .

Considering these estimates as the known parameters, we have generated 15 paired observations of size n

= 5 from a bivariate normal distribution and disturbance is introduced to the mean of the study variable

after the sixth observation, which makes the mean to be shifted to a new value . The

PLEASE INSERT FIGURES 5 - 6

inspiration of generating dataset in such manner is taken from Singh and Mangat (1996).We have applied

the classical EWMA (with   0.03 and L  2.483) and the MrEWMA (with   0.03 and

L  2.4841) to the generated datasets. Figure 5 shows that the classical EWMA detects the shift after the

12th sample, while the proposed chart detects the shift earlier at the 9th sample (Figure 6). These results are

in agreement with the results of the comparison of the proposed chart with the classical EWMA chart in

section 4.

6 Summary and conclusion

Control charts are used in monitoring process to detect any special cause variations. Primarily, they are

classified into memoryless control charts and memory control charts. Memoryless control charts aimed at

detecting large shifts in a process parameter, while memory control charts have design structures that are

efficient in detecting small to moderate shifts in a process parameter. We proposed an EWMA-type chart

that uses an auxiliary variable in the form of ratio estimator. The auxiliary variable must be strongly

positively related to the study variable for the proposed MrEWMA chart to perform better than its
existing counterparts. Except for the case of weak correlation value, it is shown that the proposed

MrEWMA chart outperforms the classical EWMA chart in detecting small to moderate shifts in the

location parameter of a control process. It gives the best performance when the correlation is 0.95. The

proposed MrEWMA chart also performs better than the mixed EWMA-CUSUM chart when  XY  0.50.

Even when there is a weak correlation between Y and X, the proposed chart outperforms the MEC chart in

detecting large shifts in process mean. In addition, the proposed chart performs closely with the EWMA-

type chart that estimates the process mean using the regression estimator. Furthermore, the performance

of the proposed chart is assessed in terms of ARL, EQL, RARL, and PCI for different positive values of

the correlation coefficient between the study variable and the auxiliary variable. At last, the proposed chat

is applied to a real dataset and it detects the disturbance in the data earlier than the classical EWMA.

In future, the runs rule schemes of the proposed chart may also be studied. The scope of the study may be

extended to the multivariate EWMA control chart using multivariate auxiliary information. Also, the

study may be extended to monitor shifts in processes variability using ARL and the overall performance

measures.

APPENDIX

Theorem: The MrEWMA control chart is more efficient then the classical EWMA control chart when

C X 2CY  YX  1 .

Proof: MSE(Plotting statistic of MrEWMA) < MSE(Plotting statistic of the classical EWMA)


  MSE  yr      2     1  1   
2t
   MSE  y    2    1    2t
Y 2  CY2  C X2  2 YX CY C X  n   SY2 n 
Y 2  CY2  C X2  2 YX CY C X  n  CY2Y 2 1 n 
Y 2  C X2  2 YX CY C X  n  0
C X2  2 YX CY C X
YX  C X 2CY

But 0  YX  1

Then, C X 2CY  YX  1 .

Generally, if CY  C X , we say, the MrEWMA control chart is more efficient then the classical EWMA

control chart when 0.5  YX  1 .

References

Abbas, N., Riaz, M., Does, R.J.M.M. (2013). Mixed exponentially weighted moving average–cumulative

sum charts for process monitoring. Quality and Reliability Engineering International. 29, 345–356.

Abbas, N., Riaz, M., Does, R.J.M.M. (2014). An EWMA-Type Control Chart for Monitoring the Process

Mean Using Auxiliary Information. Communication in Statistics - Theory Methods. 43, 3485–3498.

Abbasi, S.A., Riaz, M., Miller, A. (2012). Enhancing the performance of CUSUM scale chart. Computers

and Industrial Engineering. 63, 400–409.

Adebola, F.B., Adegoke, N.A., Sanusi, R.A. (2015). A Class of Regression Estimator with Cum-Dual

Ratio Estimator as Intercept. International Journal of Probability and Statistics. 4, 42–50.

Ahmad, S., Abbasi, S.A., Riaz, M., Abbas, N. (2014). On efficient use of auxiliary information for

control charting in SPC. Computers and Industrial Engineering. 67, 173–184.

Castagliola, P., Maravelakis, P.E., Figueiredo, F.O. (2016). The EWMA median chart with estimated

parameters. IIE Transaction. 48, 66–74.

Chong, Z.L., Mukherjee, A., Khoo, M.B.C. (2016). Distribution-Free Shewhart-Lepage Type Premier

Control Schemes for Simultaneous Monitoring of Location and Scale. Computers and Industrial

Engineering.

Choudhury, S., Singh, B.K. (2012). An efficient class of dual to product-cum-dual to ratio estimators of

finite population mean in sample surveys. Global Journal of Science Frontier Research.

Cochran, W.G. (1940). The estimation of the yields of cereal experiments by sampling for the ratio of

grain to total produce. The Journal of Agricultural Science. 30, 262–275.

Dushman, S., Lafferty, J.M., Brown, S.C. (1962). Scientific foundations of vacuum technique. American

Journal of Physics. 30, 612.

Hawkins, D.M., Olwell, D.H. (1998). Cumulative sum charts and charting for quality improvement.

Springer Science & Business Media.

Knoth, S. (2007). Accurate ARL Calculation for EWMA Control Charts Monitoring Normal Mean and

Variance Simultaneously. Sequential Analysis. 26, 251–263.

Knoth, S., Morais, M.C. (2015). On ARL-unbiased control charts. In: Frontiers in Statistical Quality

Control 11. pp. 95–117. Springer.

Li, C., Mukherjee, A., Su, Q., Xie, M. (2016). Optimal design of a distribution-free quality control

scheme for cost-efficient monitoring of unknown location. Internatinal Journal of Production Research.

54, 7259–7273.

Liu, J.Y., Xie, M., Goh, T.N., Chan, L.Y. (2007). A study of EWMA chart with transformed exponential

data. Internatinal Journal of Production Research. 45, 743–763.

Lombard, F., Hawkins, D.M., Potgieter, C.J. (2017). Sequential rank CUSUM charts for angular data.

Computational Statistics and Data Analysis. 105, 268–279.

Lucas, J.M., Crosier, R.B. (1982). Fast initial response for CUSUM quality-control schemes: give your

CUSUM a head start. Technometrics. 24, 199–205.

Marlin, T.E. (1995). Process Control: Designing Processes and Control Systems for Dynamic

Performance. McGraw-Hill Inc.

Montgomery, D. (2009). Introduction to statistical quality control. John Wiley & Sons Inc.

Mukherjee, A., Sen, R. (2015). Comparisons of Shewhart-type rank based control charts for monitoring

location parameters of univariate processes. Internatinal Journal of Production Research. 53, 4414–4445.

Murthy, M.N. (1964). Product method of estimation. Sankhyā: The Indian Journal of Statistics, Series A.

69–74.

Muth, J.F. (1960). Optimal properties of exponentially weighted forecasts. Journal of the American

Statistical Association. 55, 299–306.

Nishimura, K., Matsuura, S., Suzuki, H. (2015). Multivariate EWMA control chart based on a variable

selection using AIC for multivariate statistical process monitoring. Statistics and Probability Letters. 104,

7–13.

Noorossana, R., Fathizadan, S., Nayebpour, M.R. (2015). EWMA Control Chart Performance with

Estimated Parameters under Non‐ normality. Quality and Reliability Engineering International.

Ou, Y., Wu, Z., Tsung, F. (2012). A comparison study of effectiveness and robustness of control charts

for monitoring process mean. International Journal of Production Economics. 135, 479–490.

Page, E.S. (1954). Continuous inspection schemes. Biometrika. 100–115.

R Core Team (2016). R: A language and environment for statistical computing. R Foundation for
Statistical Computing, Vienna, Austria. 2015. URL h ttp. www. R-project. org.

Reynolds, M.R., Amin, R.W., Arnold, J.C. (1990). CUSUM charts with variable sampling intervals.

Technometrics. 32, 371–384.

Rhoads, T.R., Montgomery, D.C., Mastrangelo, C.M. (1996). A fast initial response scheme for the

exponentially weighted moving average control chart. Quality Engineering. 9, 317–327.

Riaz, M. (2008a). Monitoring process variability using auxiliary information. Computational Statistics.

23, 253–276.

Riaz, M. (2008b). Monitoring process mean level using auxiliary information. Statistica Neerlandica. 62,

458–481.

Roberts, S.W. (1959). Control chart tests based on geometric moving averages. Technometrics. 1, 239–

250.

Sanusi, R.A., Riaz, M., Abbas, N. (2017). Combined Shewhart CUSUM charts using auxiliary variable.

Computers and Industrial Engineering. 105, 329–337.

Sanusi, R.A., Riaz, M., Abbas, N., Abujiya, M.R. (2016). Using FIR to Improve CUSUM Charts for

Monitoring Process Dispersion. Quality and Reliability Engineering International.

Särndal, C.-E., Swensson, B., Wretman, J.: Model assisted survey sampling. Springer Science and

Business Media (2003).

Sheu, S.-H., Tai, S.-H., Hsieh, Y.-T., Lin, T.-C.: Monitoring process mean and variability with generally

weighted moving average control charts. Computers and Industrial Engineering. 57, 401–407 (2009).

Shewhart, W.A. (1924). Some applications of statistical methods to the analysis of physical and

engineering data. Bell System Technical Journal. 3, 43–87.

SHI, X., LV, Y., FEI, Z., Liang, J. (2013). A Multivariable Statistical Process Monitoring Method Based
on Multiscale Analysis. International Journal of Innovative Computing, Information and Control. 9,

1781–1800.

Singh, R., Mangat, N.S. (1996). Elements of Survey Sampling. Springer Science+Business Media.

Steiner, S.H. (1999). Exponentially weighted moving average control charts with time varying control

limits and fast initial response. Journal of Quality Technology. 31, 1.

Tamirat, Y., Wang, F. (2015). Sampling Plan based on the Exponentially Weighted Moving Average

Yield Index for Autocorrelation within Linear Profiles. Quality and Reliability Engineering International.

Teh, S.Y., Khoo, M.B.C., Wu, Z. (2011). A sum of squares double exponentially weighted moving

average chart. Computers and Industrial Engineering. 61, 1173–1188.

Wu, Z., Jiao, J., Yang, M., Liu, Y., Wang, Z. (2009). An enhanced adaptive CUSUM control chart. IIE

Transactions. 41, 642–653.

Wu, Z., Yang, M., Jiang, W., Khoo, M.B.C. (2008). Optimization designs of the combined Shewhart-

CUSUM control charts. Computational Statistics and Data Analysis. 53, 496–506.

Xie, Y., Xie, M., Goh, T.N. (2011). Two MEWMA charts for gumbel’s bivariate exponential distribution.

Journal of Quality Technology. 43, 50.

Yoon, S., MacGregor, J.F. (2001). Fault diagnosis with multivariate statistical models part I: using steady

state fault signatures. Journal of process control. 11, 387–400.

Zhou, Q., Shu, L., Jiang, W. (2016). One-Sided EWMA Control Charts for Monitoring Poisson Processes

with Varying Sample Sizes. Communications in Statistics-Theory and Methods.

Zwetsloot, I.M., Schoonhoven, M., Does, R.J.M.M. (2016). Robust point location estimators for the

EWMA control chart. Quality Technology and Quantitative Management. 13, 29–38.
 APPENDIX

Symbols and acronyms

Symbols and Acronyms Meaning

Y The variable of interest

X The auxiliary variable

X and Y The population means of X and Y

x and y The sample means of X and Y

Y0 and Y1 The in-control (or target) mean and the out-of-control mean of Y

yr The ratio estimator of population mean Y

E( ) , B( ) and MSE ( ) Expectation, bias and Mean Squared Error

 X and  Y The population standard deviations of X and Y

s x and s y The sample standard deviations of X and Y

 r
The population standard deviation of ratio estimator

N The sample size

 YX and rYX The population and sample correlation values between X and Y

C X and CY The population coefficient of variations of X and Y

c x and c y The sample coefficient of variations of X and Y

 The smoothing parameter of EWMA chart

L Controls the width of the control limits of EWMA chart

 The shift in process mean

 MrEWMA The proposed EWMA-type chart

ARL and SDRL The average run length and standard deviation of run length

EQL The extra quadratic loss

RARL The relative average run length

PCI The performance comparison index.

ATS Average time to signal

AEQL Average extra quadratic loss

SRS Simple random sampling

SE Standard error

Case-K, Case-U Known case and unknown case

LIST OF TABLES

Table 1: ARL values for the classical EWMA chart with ARL0 = 500.

 0 0.25 0.5 0.75 1 1.5 2 2.5 3 4 5

  0.05 L = 2.639 500 77.75 23.71 11.87 7.31 3.77 2.43 1.77 1.41 1.09 1.01
  0.25 L = 3.001 500 169.5 47.38 19.32 10.41 4.78 2.94 2.09 1.62 1.16 1.02
  0.50 L = 3.072 500 254.7 88.47 35.59 17.18 6.27 3.39 2.26 1.70 1.18 1.03

Table 2: ARL values of the proposed MrEWMA chart with ARL0 = 500 and n = 10.

 0 0.25 0.5 0.75 1 1.5 2 2.5 3 4 5

 XY = 0.25 ARL 501.50 87.45 26.72 13.54 8.24 4.25 2.72 1.99 1.56 1.16 1.03
  0.05 SDRL 514.01 79.10 20.81 9.58 5.50 2.57 1.50 1 0.70 0.38 0.16
L = 2.647 Median 343 65 22 12 7 4 2 2 1 1 1
 XY = 0.25 ARL 498.11 173.02 53.20 22.47 12.09 5.44 3.33 2.35 1.80 1.27 1.06
  0.25 SDRL 499.58 168.72 49.90 19.28 9.23 3.42 1.82 1.18 0.83 0.48 0.23
L = 3.015 Median 346 122 38 17 10 5 3 2 2 1 1
 XY = 0.25 ARL 500.56 238.07 92.45 40.20 20.11 7.42 3.92 2.61 1.92 1.30 1.07
  0.50 SDRL 490.95 236.11 91.81 38.41 18.37 5.82 2.61 1.46 0.96 0.51 0.26
L = 3.093 Median 351 166 65 28 14 6 3 2 2 1 1
 XY = 0.50 ARL 501.21 64.41 19.50 9.92 6.03 3.16 2.08 1.56 1.26 1.04 1
  0.05 SDRL 519.15 56.48 14.48 6.78 3.83 1.80 1.07 0.70 0.48 0.19 0.04
L = 2.649 Median 339 49 16 8 5 3 2 1 1 1 1
 XY = 0.50 ARL 499.65 137.90 36.19 15.06 8.26 3.95 2.48 1.79 1.41 1.08 1.01
  0.25 SDRL 500.26 135.99 32.85 11.92 5.68 2.25 1.24 0.82 0.58 0.27 0.08
L = 3.0105 Median 346 97 26 12 7 4 2 2 1 1 1
 XY = 0.50 ARL 500.23 208.35 66.64 26.10 12.85 4.91 2.79 1.90 1.47 1.09 1.01
  0.50 SDRL 496.65 206.79 64.80 24.59 11.07 3.40 1.60 0.94 0.63 0.29 0.09
L = 3.0889 Median 347 145 47 19 9 4 2 2 1 1 1
 XY = 0.75 ARL 499.80 38.56 11.79 5.99 3.76 2.03 1.40 1.13 1.03 1 1
  0.05 SDRL 508.88 31.48 8.20 3.78 2.19 1.02 0.59 0.35 0.17 0.02 0
L = 2.647 Median 344 31 10 5 3 2 1 1 1 1 1
 XY = 0.75 ARL 495.57 85.40 19.22 8.21 4.73 2.42 1.60 1.23 1.06 1 1
  0.25 SDRL 497.11 82.34 15.87 5.67 2.78 1.19 0.70 0.44 0.25 0.04 0
L = 3.01 Median 343 60 15 7 4 2 1 1 1 1 1
 XY = 0.75 ARL 502.20 152.35 35.97 12.75 6.25 2.69 1.69 1.26 1.07 1 1
  0.50 SDRL 504.94 150.34 34.34 10.96 4.63 1.49 0.78 0.48 0.27 0.04 0.01
L = 3.085 Median 349 106 25.5 9 5 2 2 1 1 1 1
 XY = 0.95 ARL 500.15 15.73 4.86 2.57 1.71 1.10 1 1 1 1 1
  0.05 SDRL 512.42 11.06 2.91 1.36 0.79 0.31 0.07 0 0 0 0
L = 2.647 Median 346 13.5 4 2 2 1 1 1 1 1 1
 XY = 0.95 ARL 498.81 29.07 6.49 3.14 2.01 1.19 1.01 1 1 1 1
  0.25 SDRL 495.81 25.35 4.13 1.63 0.93 0.41 0.11 0.01 0 0 0
L = 3.013 Median 349 22 6 3 2 1 1 1 1 1 1
 XY = 0.95 ARL 504.56 63.63 9.85 3.76 2.17 1.22 1.01 1 1 1 1
  0.50 SDRL 501.93 61.70 7.92 2.33 1.11 0.44 0.12 0.02 0 0 0
L = 3.099 Median 355 45 8 3 2 1 1 1 1 1 1

Table 3: Performance comparison of the proposed MrEWMA chart with the classical EWMA
chart as the benchmark chart

CLASSICAL MrEWMA
 EWMA  XY = 0.25  XY = 0.5  XY = 0.75  XY = 0.95
EQL 0.05 9.69 10.67 8.62 6.73 5.52
0.25 12.60 13.97 10.76 7.74 5.75
0.5 16.48 18.14 13.51 9.16 6.16
RARL 0.05 1 1.11 0.87 0.63 0.46
 0.25 1 1.11 0.85 0.59 0.41
0.5 1 1.10 0.83 0.56 0.39
PCI 0.05 1 1.10 0.89 0.69 0.57
0.25 1 1.11 0.85 0.61 0.46
0.5 1 1.10 0.82 0.56 0.37

Table 4: ARL values of the MxEWMA chart with ARL0 = 500


 XY  , L  0 0.25 0.5 0.75 1 1.5 2 2.5 3 4 5
(0.05, 2.639) 500.54 73.80 22.43 11.25 6.93 3.58 2.31 1.70 1.36 1.07 1.01
(0.25, 3.001) 499.80 161.97 44.18 18 9.77 4.50 2.80 1.99 1.55 1.13 1.01
0.25
(0.50, 3.072) 499.95 245.96 83.01 32.90 15.83 5.83 3.18 2.15 1.62 1.15 1.02
(0.05, 2.639) 499.56 61.10 18.54 9.33 5.76 3.01 1.97 1.48 1.22 1.03 1
(0.25, 3.001) 499.69 135.77 34.58 14.11 7.80 3.72 2.35 1.70 1.35 1.05 1
0.50
(0.50, 3.072) 500.79 216.10 65.40 24.80 11.91 4.56 2.59 1.80 1.39 1.06 1
(0.05, 2.639) 499.59 38.46 11.71 5.93 3.72 2.03 1.40 1.13 1.03 1 1
(0.25, 3.001) 500.41 84.13 18.97 8.08 4.72 2.41 1.59 1.23 1.06 1 1
0.75
(0.50, 3.072) 499.29 146.89 34.90 12.46 6.15 2.68 1.68 1.26 1.07 1 1
(0.05, 2.639) 499.34 10.61 3.40 1.87 1.31 1.02 1 1 1 1 1
(0.25, 3.001) 500.23 16.70 4.26 2.21 1.48 1.03 1 1 1 1 1
0.95
(0.50, 3.072) 500.01 30.23 5.39 2.42 1.55 1.04 1 1 1 1 1
Acknowledgement:

The work described in this paper is partially supported by Hong Kong Research Grant Council (GRF

project No. 9042327 and a theme-based project grant T32-101/15-R).

 LIST OF FIGURES

Figure 1: Estimated SE of the ratio estimator Vs. SE of the usual SRS mean estimator.
Figure 2: Out-of-control ARL curve of the classical EWMA chart and the MrEWMA chart (for
 XY = 0.25 and 0.95).
Figure 3: EQL, RARL, and PCI of the classical EWMA chart and the MrEWMA chart.
Figure 4: Out-of-control ARL curve of the mixed EWMA-CUSUM (MEC) chart and the
MrEWMA chart (for  XY = 0.5, 0.75, and 0.95) when (a)   0.25 and (b)   0.50.
 368.35
Outlet Temperature

368.3

368.25

368.2

368.15

368.1
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Observation Number

Figure 5: Classical EWMA chart for the CSTR dataset.

 368.34
368.32
Outlet Temperature

368.3
368.28
368.26
368.24
368.22
368.2
368.18
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Observation Number

Figure 6: MrEWMA chart for the CSTR dataset.

 HIGHLIGHTS

 An EWMA-type scheme, with a single auxiliary variable, is proposed.

 The ratio mean estimator is used to improve the performance of the classical EWMA chart for

monitoring location parameter.

 The study shows that the proposed chart outperform its existing counterparts.

 A real-life industrial example is provided.