2011 International Conference on Circuits, System and Simulation

IPCSIT vol.7 (2011) © (2011) IACSIT Press, Singapore

A Study on the False Alarm Rates of X , EWMA and CUSUM Control

Charts when Parameters are Estimated

Michael Boon Chong Khoo +, Sin Yin Teh, Li Fen Ang and Xi Wah Ng
School of Mathematical Sciences, Unversiti Sains Malaysia, 11800 Minden, Penang, Malaysia.

Abstract. Nowadays, quality is an important issue, especially in manufacturing industries. Producing a

quality product that meets customers’ requirements is the main objective of manufacturing industries. A
control chart is a powerful tool in statistical process control (SPC) which is widely used in manufacturing
industries. In order to construct a control chart, the estimates of the parameters mean, μ and standard
deviation, σ have to be computed first. The Shewhart X chart is slow in detecting small shifts in the process
mean. Thus, usually the EWMA (exponentially weighted moving average) and CUSUM (cumulative sum)
charts are used in the detection of small shifts. The aim of this study is to compare the performances of the
Shewhart X , EWMA and CUSUM charts, in terms of their false alarm rates, when parameters are estimated.
A Monte Carlo simulation is conducted using the Statistical Analysis Software (SAS) to compute the average
run length (ARL) profiles of each of the charts. Overall, the results show that the false alarm rates when the
parameters are estimated is inversely proportional to the number of in-control samples and the size of each of
these samples, used in the estimation of the process parameters.
Keywords: ARL; X chart; EWMA chart; CUSUM chart; false alarm rates; parameter estimation

1. Introduction
Statistical process control (SPC) is a collection of problem-solving tools that is useful in achieving
process stability and improving capability through the reduction of variability [1]. A control chart is a
powerful tool in SPC. It is used in the monitoring of the quality of a process. A control chart displays the
quality characteristic of a process in a graphical way. It consists of the lower control limit (LCL), center line
(CL) and the upper control limit (UCL). The stability of a process can be determined from a control chart. If
all the sample points fall within the control limits without showing any systematic behavior, it shows that the
process is stable or in statistical control [2].
It is sometimes difficult to compute the control limits because the process parameters, such as μ and σ
are usually unknown. Therefore, these parameters are estimated from a preliminary sample taken from an in-
control process. The estimated sample mean and sample variance should be unbiased of the population mean
and population variance, respectively.
There are two distinct phases in the implementation of a control chart. In Phase I, parameters are
estimated from the dataset given by the preliminary samples, followed by computing the trial control limits.
After a Phase I control chart is constructed, the process is checked to ensure that it stays in-control before
control charting for a Phase II process is initiated. In a Phase II process, a control chart is used to monitor a
future process. Studying the false alarm rates is one of the ways to evaluate the performance of a control
chart. A false alarm, a.k.a., a Type I error occurs when a control chart declares a process to be out-of-control,
when in fact it is in-control [3]. An efficient control chart is able to detect instantaneously a process shift,
stop sampling as quickly as possible when an out-of-control signal is detected and adopts essential corrective
actions to improve the process quality [4].
+
Corresponding author. Tel.: + (6)04-6533941; fax: +(6)04-6570910.
E-mail address: [email protected]
61
 In this study, we are interested to study the false alarm rates of the Shewhart X , EWMA and CUSUM
charts, when parameters are estimated. The false alarm rates are computed using a Monte Carlo simulation.
The remainder of this paper is organized as follows: Section 2 discusses the unbiased estimators for
estimating the process mean and variance of the Shewhart X , CUSUM and EWMA control charts. In
Section 3, a simulation study is carried out to compare the performances of the Shewhart X , CUSUM and
EWMA charts, based on the false alarm rates computed. Some useful conclusions and recommendations of
potential future works are summarized in Section 4. Section 5 provides some acknowledgements while
Section 6 contains the references that are used in this paper.

2. Unbiased Estimators for Estimating the Process Mean and Variance

Most studies in evaluating the performance of a control chart assume that the in-control process
parameters are known. However, there are many cases where the process parameters are unknown. Woodall
and Montgomery [5] studied the effects of parameter estimation on the properties of control charts. The most
important point in parameter estimation is selecting the most suitable estimator. We are concerned with
estimating the process average, μ and the process standard deviation, σ from preliminary samples or samples
taken from the process which is assumed to be in-control and follows a normal distribution. Because of a
high sampling and inspection costs when the sample size is large, an appropriate sample size that is
sufficiently large would be used in estimating the process parameters.
The performances of the Shewhart X , EWMA and CUSUM charts are compared, in terms of the false
alarm rates when parameters are estimated. The following are some descriptions of these charts and their
respective unbiased estimators for estimating the process mean and variance.

2.1. The Shewhart X Chart

Suppose that a process follows a normal distribution and a sample of size n, consisting of measurements
X 1 , X 2 ,..., X n is drawn from this process, where the average of this sample is
1 n
X= ∑ Xi .
n i =1
(1)

Let X 1 , X 2 ,..., X m be m sample averages computed from an in-control process, then the best estimator of μ, i.e.
the process grand average, X is computed as
1 m
X= ∑Xj .
m j =1
(2)

The process standard deviation, σ is another unknown parameter in most of the applications of control
charts. It is a measurement showing the spread of the data and is estimated from either the standard
deviations or ranges of the m samples. Here, the estimation of σ using the sample range method is used.
Let R1 , R2 ,..., Rm be the ranges of the m samples, where R = X max − X min is the difference between the
largest sample observation X max and the smallest sample observation X min . Then, the average sample range,
R is calculated by
1 m
R= ∑ Rk .
m k =1
(3)

The control limits of the Shewhart X chart can be computed as follows [1]:
a
LCL = X − R, (4)
d2 n
CL = X (5)
and
a
UCL = X + R, (6)
d2 n
where “a” in Equations (4) and (6) are measured in multiples of standard deviation, whose value depends on
the desired in-control average run length, ARL0. The value of the constant d2 depends on the sample size and
is given in most statistical quality control textbooks.

62
 The main drawback of the Shewhart X control chart [6] is that it only considers the information
given in the last plotted point and ignores all the information in the sequence of points. Thus, it is less
effective in detecting small changes in the process. Page [7] proposed the CUSUM chart while Roberts [8]
introduced the EWMA chart to enable a quick detection of small shifts. These 2 charts complement the
Shewhart X chart when the detection of a small shift is of interest. This is because the CUSUM and EWMA
charts take into account the information contain in the sequence of points. Thus, they perform better in
signalling small changes compared to the Shewhart X chart.

2.2. The CUSUM Chart

The CUSUM chart, developed by Page [7] incorporates past information into each individually plotted
observation. This increases the sensitivity of the CUSUM chart in detecting small shifts in the process. The
CUSUM chart plots the cumulative sums of deviations of the sample values from a target value against time.
There are two types of CUSUM charts, i.e., the tabular CUSUM and the V-mask CUSUM charts. A CUSUM
chart is a two-sided chart. It plots the cumulative sum of the sample average versus the sample number j. The
tabular CUSUM employs two sample statistics ( C +j and C −j ), where one is the one-sided upper CUSUM that
accumulates positive deviations above the target and the other is the one-sided lower CUSUM that
accumulates negative deviations below the target [9]:
C −j = min {0, C −j −1 + (Y j + k )} (7)
and
C +j = max {0, C +j −1 + (Y j − k )} , (8)
δ X j − μ0
where k = is the chart’s parameter, with k ≥ 0 and δ is the shift in the mean. Note that Y j = .
2 σ
n
The lower and upper control limits of a CUSUM chart are as follows:
LCL = − H (9)
and
UCL = H . (10)

Similar to the Shewhart X chart, the center line of the tabular CUSUM chart represents the target value, μ0 .
If either C +j or C −j exceeds the predetermined decision interval, H, the process is considered to be out-of-
control.

2.3. The EWMA Chart

The EWMA statistics [8] is defined as follows:
Z j = λX j + (1 − λ ) Z j −1 , j = 1, 2,..., m (11)
Here, 0 < λ ≤ 1 is the smoothing constant while Z 0 = μ0 is the starting value of Zj, when the process is in-
control.
The control limits and center line of an EWMA chart are as follows:
λ ⎡1 − (1 − λ ) ⎤
2i
σ ⎣ ⎦ ,
LCL = μ0 − L (12)
n 2−λ
CL = μ 0 (13)
and
λ ⎡1 − (1 − λ ) ⎤
2i
σ ⎣ ⎦ .
UCL = μ 0 + L (14)
n 2−λ
The factor L controls the width of the control limits. As i increases, the term ⎡1 − (1 − λ ) ⎤ approaches unity.
2i
⎣ ⎦
When the parameters μ0 and σ are estimated, the control limits of the EWMA chart in Equations (12) and (14)
become:

63
 R λ
LCL = X − L (15)
d2 n 2−λ
and
R λ
UCL = X + L . (16)
d2 n 2−λ

Note that the EWMA chart behaves like the Shewhart X chart when λ = 1 because for this case, the EWMA
statistics depend on only the most recent sample.

3. A Comparison of the Performances of the X , EWMA and CUSUM Charts

When Parameters are Estimated
The main objective of statistical process control (SPC) is to detect as early as possible the presence of
assignable causes of variation that affects the quality of a process. This study compares the performances of
the Shewhart X , CUSUM and EWMA charts, in terms of false alarm rates, when parameters are estimated.
When the process is in-control, the false alarm rates should be sufficiently small or close to the target value.
The in-control ARLs of the Shewhart X , CUSUM and EWMA charts are set as ARL0 = 250 and their false
alarm rates are compared, for a given combination of the number of in-control samples, m and the size of
each sample, n. The number of samples, m = 20, 50, 100 and 1000 and the sample sizes n = 3, 5, 10, 15, 20,
25, 30, 40, 50 and 100 are considered.

Table 1. False alarm rates for the Shewhart X , CUSUM and EWMA charts for ARL0=250.
CUSUM EWMA
Shewhart X
k = 0.50 H = 4.360 λ = 0.15 L = 2.659
m 20 50 100 1000 20 50 100 1000 20 50 100 1000
n
3 .00736 .00526 .00463 .00404 .01393 .00818 .00633 .00479 .01160 .00709 .00549 .00406
5 .00611 .00476 .00435 .00401 .01274 .00758 .00604 .00477 .01104 .00676 .00532 .00407
10 .00555 .00460 .00430 .00402 .01201 .00733 .00595 .00482 .01062 .00659 .00529 .00409
15 .00531 .00447 .00422 .00400 .01198 .00728 .00595 .00476 .01067 .00661 .00528 .00404
20 .00524 .00446 .00422 .00396 .01171 .00724 .00586 .00472 .01048 .00658 .00521 .00402
25 .00527 .00446 .00419 .00402 .01215 .00726 .00587 .00477 .01080 .00658 .00519 .00407
30 .00520 .00444 .00421 .00398 .01181 .00715 .00587 .00478 .01055 .00650 .00523 .00406
40 .00513 .00443 .00419 .00401 .01185 .00724 .00580 .00481 .01060 .00661 .00519 .00406
50 .00516 .00438 .00417 .00397 .01191 .00717 .00585 .00478 .01066 .00652 .00521 .00405
100 .00511 .00442 .00420 .00401 .01169 .00720 .00583 .00479 .01047 .00652 .00519 .00404
Table 1 presents the false alarm rates for the Shewhart X , CUSUM and EWMA charts when parameters
are estimated, based on ARL0 = 250. The false alarm rate corresponding to ARL0 = 250 when parameters are
known is 1/250 = 0.004. The results for the Shewhart X chart in Table 1 show that when m and (or) n
increase(s), the false alarm rates when parameters are estimated approach 0.004, i.e., close to the value when
parameters are known. For example, when n = 3 and m varies from {20 → 50 → 100 → 1000}, the false
alarm rates decrease from {0.00736 → 0.00526 → 0.00463 → 0.00404}. However, when m = 50 and n
increases from {3 → 5 → 10 → 15 → 20 → 25 → 30 → 40 → 50 → 100}, the false alarm rates decrease
from {0.00526 → 0.00476 → 0.00460 → 0.00447 → 0.00446 → 0.00446 → 0.00444 → 0.00443 →
0.00438 → 0.00442}. We observe that when either m or n increases, the false alarm rate decreases. In other
words, the false alarm rate is inversely proportional to m and n. From the false alarm rates for the CUSUM
and EWMA charts, we can conclude that changes in m and n have the same effect on the Shewhart X ,
CUSUM and EWMA charts. The false alarm rates of the Shewhart X , CUSUM and EWMA charts when
parameters are estimated approach that of the case when parameters are known when either or both m and n
increase.
An interesting finding in Table 1 is that the effect of n is negligible when m is large. Note that the false
alarm rates are close to one another when m = 1000, where they are almost the same as those for the case
when parameters are known, i.e. 0.004. We also notice that there is no significant difference among the false

64
alarm rates when n ≥ 15. The Shewhart X , CUSUM and EWMA charts show the same behaviour with
respect to changes in m and n, when parameters are estimated.
In addition, we also found from Table 1 that in contrast to the Shewhart X chart, the CUSUM and
EWMA charts need a larger m to give reliable estimates of their false alarm rates. Also note that the false
alarm rates of the CUSUM chart are slightly higher than that of the EWMA chart.
It is worth noting that similar trends are observed when different ARL0 values, such as 500, 1000 and
2000 are used. The results are not displayed in this paper because of space limitation.

4. Conclusions and Recommendations

The findings in this study indicate that parameter estimation is an important issue in control charting.
The choice of the values of m and n affects the false alarm rates when parameters are estimated. We found
that the false alarm rates for the three charts considered are higher than their nominal values. If a large m,
like m = 1000 is selected, the false alarm rates of the charts approach that of the nominal value, irrespective
of the sample size, n. However, meeting this requirement is not practical in most process monitoring
situations. It is very costly and time consuming to use a large m. A moderate number of samples may be
sufficient to get an estimate that does not differ much from the nominal value. Thus, for estimating
appropriate control limits, a moderately large m and n suffice. This study serves as a reference in determining
adequate number of samples and the sample size for the purpose of parameter estimation.
Future works related to this topic that is worthy of pursuing are as follows:
• To evaluate the performances of control charts in detecting shifts in the process variance when
parameters are estimated.
• To investigate the effects of parameter estimation on multivariate EWMA and multivariate CUSUM
charts.
• To study the effects of parameter estimation on attribute charts.

5. Acknowledgements
This work is funded by the Universiti Sains Malaysia (USM) Incentive Grant, no.
1001/PMATHS/822080 and supported by the Academic Staff Training Scheme (ASTS), USM.

6. References
[1] D. C. Montgomery. Statistical Quality Control: A Modern Introduction, 6th ed. Asia: John Wiley & Sons (Asia)
Pte. Ltd., 2009.
[2] J. S. Oakland. Statistical Process Control, 6th ed. Oxford, UK: Butterworth-Heinemann, 2008.
[3] D. H. Besterfield. Quality Control, 8th ed. Upper Saddle River, New Jersey: Pearson Education Inc., 2009.
[4] T. S. Bakir. A distibution-free Shewhart quality control chart based on signed-ranks. Qual. Eng. 2004, 16 (4): 613-
623.
[5] W. H. Woodall, and D. C. Montgomery. Research issues and ideas in statistical process control. J. Qual. Technol.
1999, 31: 374-386.
[6] W. A. Shewhart. Economic Control of Quality of Manufactured Products. Princeton, New Jersey: D. Van
Nostrand. Reprinted by the American Society for Quality Control, Milwaukee, Wisconsin, 1931.
[7] E. S. Page. Continuous inspection schemes. Biometrika 1954, 41 (1-2): 100-115.
[8] S. W. Roberts. Control chart tests based on geometric moving averages. Technometrics 1959, 1: 239-250.
[9] F. F. Gan. An optimal design of CUSUM quality control charts. J. Qual. Technol. 1991, 23 (4): 279-386.

65