1. Introduction
A main objective for a process is to continuously improve its quality, which can be statistically expressed as variation reduction. Chance and assignable causes exist and lead to variation in a process. The variation caused by change is unavoidable and always exists in a process, even if the operation is carried out using standardized raw material and methods. It is not practical to eliminate the chance cause technically and economically, while variation caused by assignable causes indicates that there exist some unwanted factors to be detected.
Statistical Process Monitoring (SPM) provides a large set of tools to help practitioners in monitoring manufacturing or service processes to quickly detect assignable causes. Among these, control charts are widely used online and can be implemented with a charting statistic related to the process mean or/and dispersion. The aim of a control chart is to detect abnormal changes in the process as soon as possible. Many univariate mean (
) charts, such as the Shewhart
chart, Cumulative Sum (CUSUM)
chart, and Exponentially Weighted Moving Average (EWMA)
chart were investigated by researchers; see Brook and Evans [
1], Nelson [
2], Lucas and Saccucci [
3], and Hawkins and Olwell [
4]. More recent works on control charts can refer to Li et al. [
5], Mukherjee and Rakitzis [
6], Zwetsloot et al. [
7], and Perry [
8], to name a few. The Shewhart charts are known to be effective when the shift size in the process is large. For the detection of small to moderate shifts, both CUSUM and EWMA charts using the current and former samples information perform much better than Shewhart type charts, see Montgomery [
9].
Since the primary works of Crowder [
10], Lucas and Saccucci [
3], and Domangue and Patch [
11], EWMA type charts have received much attention. For example, for non-normal and autocorrelated processes, the properties of EWMA
charts were first investigated by Borror et al. [
12] and Lu and Jr. [
13], respectively. The performance of the EWMA
chart was investigated by Jones et al. [
14] when the process parameters are estimated. Recently, Celano et al. [
15], Calzada and Scariano [
16], and Haq et al. [
17] studied the run length performance of the EWMA
t charts. To summarise, only a two-sided EWMA chart was used in the above researches. In practice, the direction of the out-of-control shift is usually known in advance, which implies that it is possible to tune the upward and downward parts of EWMA charts separately [
18]. Then two separate one-sided EWMA charts were studied by some researchers. For instance, Tran et al. [
19] and Tran and Knoth [
20] studied the properties of two one-sided EWMA charts to monitor the ratio of two variables. Zhang et al. [
21] and Muhammad et al. [
22] investigated the performance of two one-sided EWMA charts for monitoring the coefficient of variation (CV).
In this paper, the work in Zhang et al. [
21] is highlighted for the new resetting model of the Modified One-sided EWMA (MOEWMA) charting statistic. In the EWMA charting statistic, information of former and current samples are both used and the charting statistic is reset to the target if it is smaller than the target. While their work studied the EWMA chart for monitoring the CV, as far as we know, there is no research on the proposed scheme for monitoring the mean of a normally distributed process. In fact, a normally distributed quality characteristic usually exists in some industrial processes. To fill this gap, we investigate the properties of the MOEWMA
charts. In addition, it is known that control charts with the variable sampling interval (VSI) features are more efficient than the corresponding fixed sampling interval (FSI) charts in the detection of shifts. In the past decades, much research has been conducted on VSI control charts. For instance, Nguyen et al. [
23] suggested a VSI CUSUM chart to monitor the ratio of two normal variables and showed that the proposed chart had some advantages over the corresponding FSI CUSUM chart. Using extensive Monte-Carlo simulations, Haq [
24] studied the performance of the weighted adaptive multivariate CUSUM chart with VSI feature. It was shown that the proposed charts perform uniformly better than the corresponding FSI charts in terms of the ATS (Average Time to Signal) and AATS (Average Adjusted Time to Signal) performances. Coelho et al. [
25] proposed a VSI nonparametric Shewhart type control chart, which was shown to be better than the existing FSI chart. For more research works, we direct readers to the works [
26,
27,
28,
29,
30,
31] and the references cited therein. To further increase the sensitivity of the MOEWMA
charts and gain motivation from the above works on the VSI charts, the VSI MOEWMA
charts are proposed, and it is expected that the VSI MOEWMA charts perform better than the corresponding FSI one-sided charts.
The remainder of this paper is organized as follows:
Section 2 reviews several types of one-sided EWMA
charts and presents the MOEWMA
chart. The zero-state (ZS) and steady-state (SS) Average Run Length (
) performances of the proposed MOEWMA
charts are presented in
Section 3 and are compared with other competing charts.
Section 4 presents the detailed construction of the MOEWMA
charts with the VSI feature and, moreover, both the ZS and SS performances of the proposed VSI MOEWMA
charts are investigated. A real data example is used to illustrate the implementation of the MOEWMA
charts in
Section 5. Finally, some conclusions and recommendations are made in the last section.
2. One-Sided EWMA Type Charts
Assume that , is a sample of size from an independent normal distribution, i.e., , where and are the in-control mean and standard deviation, respectively, and is the magnitude of the mean shift. When , the process is considered to be in-control. Otherwise, the process is out-of-control. At each sample point , the sample mean is computed for the process monitoring, where . Without loss of generality, we assume and in this paper.
2.1. Traditional One-Sided EWMA Charts
The traditional two-sided EWMA chart construct the monitoring statistic with a fixed smoothing constant and the initial value . The upper () and lower () control limits of the EWMA chart are generally selected based on the constraint of the desired in-control . If , the process is considered to be in-control. Otherwise, if , the process is deemed to be out-of-control. Instead of using a single two-sided EWMA chart, when the direction of the shift is known, three types of one-sided EWMA charts were suggested by some researchers. These charts are summarized as follows:
- (1)
A simple use of the one-sided EWMA
chart is to set only an upper control limit (
) or a lower control limit (
) with the traditional charting statistic
and the initial value
. This chart is denoted as SEWMA
chart. That is to say, the upper-sided SEWMA
chart declares an alarm when
and the lower-sided SEWMA
chart declares an alarm when
. More details of the SEWMA
chart can be seen in Robinson and Ho [
32].
- (2)
A second use of the one-sided EWMA
chart is to reset the traditional EWMA statistic to the target whenever it is smaller than the target (for the upper-sided chart) or whenever it is larger than the target (for the lower-sided chart). This chart is denoted as REWMA
chart. The charting statistics
and
of the upper- and lower-sided REWMA
charts are given as follows,
and
with the initial value
. An out-of-control signal is triggered as soon as
(for the upper-sided REWMA
chart) or
(for the lower-sided REWMA
chart), respectively. More details of REWMA type charts can be seen in Hamilton and Crowder [
33] and Gan [
34].
- (3)
A third use of the one-sided EWMA
chart is first truncate the sample mean
below the target to the target value (for the upper-sided chart) or above the target to the target value (for the lower-sided chart), and then apply the EWMA recursion to these truncated values. This chart is denoted as IEWMA
chart. The charting statistic
of the IEWMA chart is given as follows:
where
=
is the standardized value of
(for the upper-sided chart), and
=
is the standardized value of
= min(
,
) (for the lower-sided chart). The initial value
is set as 0. An out-of-control signal is given when
in the upper-sided chart or
in the lower-sided chart. More details of this chart can be seen in Shu and Jiang [
35] and Shu et al. [
36].
2.2. The Proposed MOEWMA Charts
In this section, the MOEWMA
charts with a new resetting model are investigated. As it will be shown in
Section 3, the proposed charts outperform the traditional one-sided EWMA
charts presented in
Section 2.1.
It can be seen from Equation (
1) that, when
is smaller than
, then
and
. All the samples information collected before time
are lost. As the main advantage of EWMA type charts is to use both current and former samples information, the charting statistic of the upper-sided MOEWMA
chart is constructed as,
where
and the initial value
. It can be noted that the charting statistic
in Equation (
4) uses all samples information collected before. The chart triggers an out-of-control signal if
is larger than the
. Similarly, a lower-sided MOEWMA
is suggested with the following charting statistic,
where the initial value
. An out-of-control signal is triggered if
is smaller than the
.
3. Numerical Results and Comparisons
In this section, some
measures, including the
and the Standard Deviation of Run Length (
) are used to investigate the performance of the one-sided EWMA charts in
Section 2. The
is defined as the expected number of samples on the chart until a signal occurs. A control chart is desirable when the in-control
is large and at the same time, the out-of-control
is as small as possible. In addition, the
determines the variability of the
distribution. The smaller the
value, the better the
performance of a control chart, see Haq [
37]. In addition, the subscripts 0 and 1 are used with
and
to denote the in-control and out-of-control properties, respectively. To obtain these
properties of the proposed chart, the Monte-Carlo method is adopted in this paper. Under each simulation run,
iterations of
values are used to calculate the values of
and
.
3.1. Comparisons with Some Competing Charts
In this section, to provide some direct insight into the performance of the proposed charts, the (
,
) of the MOEWMA
charts are compared with the ones of the SEWMA, REWMA, and IEWMA
charts. The properties of the SEWMA, REWMA, and IEWMA
charts can be obtained using the Markov chain approach. For the EWMA type charts, values of
were recommended by Montgomery [
9]. Moreover, a relatively small smoothing parameter
is usually suggested for monitoring small shifts while larger values of
are suggested for larger shifts. In this paper,
are selected for illustration and the corresponding control limits of EWMA type charts can be obtained with the constraint on the desired
. For simplicity, the
is set to be 200. For the proposed one-sided MOEWMA
chart, a bisection algorithm similar to Dickinson et al. [
38] is used to find the control limit. The algorithm stops when the in-control
falls within the interval
.
Table 1 presents the (
,
) values of these EWMA control charts for different shifts
varying from 0.1 to 3 when
. It can be noted from this table that, for the upper-sided MOEWMA
chart, a small value of
is relatively effective for small shifts
and vice verse. For instance, when
and
, the (
,
) = (54.06, 45.94) of the upper-sided MOEWMA
chart when
is smaller than the (
,
) = (75.26, 71.98) of the chart when
. Compared with the competing EWMA (SEWMA, REWMA, and IEWMA) charts, some conclusions are made as follows:
Irrespective of the values of and n, the and values of the upper-sided MOEWMA chart are generally smaller than the ones of the upper-sided REWMA chart, especially for small shifts. This fact clearly demonstrates the advantage of the proposed chart. For instance, when , , and , the of the upper-sided MOEWMA chart are smaller than the of the upper-sided REWMA chart.
The proposed chart always has a little smaller value than the one of the upper-sided SEWMA chart. For example, for the same values of n, and presented above, the of the upper-sided SEWMA chart are close to the ones of the upper-sided MOEWMA chart.
Compared with the upper-sided IEWMA chart, the proposed chart performs better for small shifts and worse for moderate to large shifts. For instance, when and , the upper-sided MOEWMA chart with () = (60.12, 54.07) is better than the upper-sided IEWMA chart with () = (68.17, 63.64) for the detection of . However, for the detection of , the upper-sided MOEWMA chart with () = (3.81, 1.29) is worse than the upper-sided IEWMA chart with () = (3.38, 1.43).
For a large shift, for instance when or larger than 3, all the charts perform similarly, as the is close to 1 and the value converges to 0 with increasing.
As the symmetry of the normal distribution, similar conclusions are drawn for the lower-sided MOEWMA chart. For simplicity, these results are not presented here.
3.2. Optimal Performance of the Proposed MOEWMA Charts
The results in
Section 3.1 show the advantage of the proposed chart over the SEWMA, REWMA, and IEWMA
charts. All of the simulations above are for a fixed value of
, which is not optimal for the specified shift size
. To provide a fare comparison, the optimal performances of different charts for the intended shift size are compared in this section. The optimal design of the upper-sided MOEWMA
chart involves determining the chart parameters
to minimize the
at a specified mean shift
, at the same time, satisfying the constraint on the desired
. The procedure can be concluded as a constrained nonlinear minimization problem:
subject to
By using this model, extensive computation works are then performed to numerically find the nearly optimal parameters
of the upper-sided MOEWMA
chart.
Table 2 presents the optimal chart parameters
of the proposed chart for
and the (
) values of the chart at shift
varying from 0.1 to 3. As a comparison, the nearly optimal parameters and performances of the REWMA, SEWMA, and IEWMA
charts are also presented. All charts are designed to maintain
. For example, if the specified shift size
, the
s of the upper-sided MOEWMA
chart are first determined for
to obtain
. The
values are then computed for all the combinations of
. The parameters
leading to the smallest
are considered to be the nearly optimal parameters of the control chart.
It can be concluded from
Table 2 that:
If the specified shift is small (), the optimal upper-sided MOEWMA chart performs better than the optimal REWMA, SEWMA, and IEWMA charts. For instance, if , the optimal parameters of the upper-sided MOEWMA chart is (0.05, 0.17) and the corresponding is the smallest one among these charts.
The upper-sided MOEWMA chart provides a good sensitivity against shifts smaller than the specified and the upper-sided IEWMA chart performs better than other charts for shifts larger than the specified . For instance, if , while the actual shift size in the process is not the specified one and is (smaller than ), the upper-sided MOEWMA chart with is better than other charts. If the actual shift size is (larger than ), the upper-sided IEWMA chart with performs better than other charts.
If the specified shift is moderate (), the upper-sided IEWMA chart has better sensitivity than the REWMA, SEWMA, and MOEWMA charts for all the shift sizes. For example, when , the optimal () = (1.74, 0.87) of the upper-sided IEWMA chart is smaller than the ones of these charts, and if the actual shift sizes is smaller or larger than 1.5, the upper-sided IEWMA chart still performs better than these charts.
If the actual shift size is large (), the upper-sided MOEWMA chart has the best performance among all the charts. For instance, when , the optimal is the same for all the charts. If the actual shift size is smaller than , it can be seen that the upper-sided MOEWMA chart has the smallest () value among these EWMA type charts.
The above results also indicate that both the upper-sided IEWMA and MOEWMA
charts have a practical property of good performance over a wide range of shifts rather than a scheme to optimize the control charts at a specified shift
. This property was considered to be important, as in applications, the value of shift size is seldom known, and therefore a robust monitoring procedure that efficiently signals a range of shifts is useful [
39].
3.3. The Steady-State Performance of the Proposed Chart
The results presented in the previous section are for the case in which the shift occurs from the beginning of the process or the charting statistic is at its initial starting value when the shift occurs. The computed
in this way is referred as the zero-state
. The steady-state
is based on the assumption that the process remains in-control for a long time and a shift occurs later in the process. The steady-state
of control chart is considered to be more realistic than the zero-state
, see Zwetsloot et al. [
7]. For the steady-state case,
Monte-Carlo simulations are used to estimate the steady-state
values of control charts and the shift is assumed to happen in the process after 50 in-control samples, see Dickinson et al. [
38], Xu and Jeske [
40], and Haq [
24].
The out-of-control steady-state
and
of the proposed chart together with the ones of the REWMA, SEWMA, and IEWMA
charts are presented in
Table 3 for different combinations of
n,
, and
. The in-control
is set to be 200. It can be noted from
Table 3 that the steady-state performance of the upper-sided MOEWMA
chart is almost the same as the upper-sided SEWMA
chart. Moreover, for a small shift (
), both the upper-sided MOEWMA
chart and the upper-sided SEWMA
chart generally perform better than the upper-sided IEWMA and REWMA
charts. For instance, when
,
, and
, the steady-state (
) = (58.66, 54.58) and (
) = (58.63, 54.38) of the upper-sided SEWMA and MOEWMA
charts are smaller than the steady-state (
) = (67.22, 63.92) and (
) = (65.35, 62.84) of the upper-sided IEWMA and REWMA
charts. Moreover, for the shifts larger than 0.5, we can note that the upper-sided IEWMA
chart generally performs best among these charts. The upper-sided REWMA
chart is preferred only when
.
5. A Real Data Application
To show the application of the REWMA, SEWMA, IEWMA, and the proposed MOEWMA
charts, in what follows, a real dataset of semiconductor manufacturing in Montgomery [
9] is used to illustrate the charts’ implementation. The photolithography process is important in semiconductor manufacturing. It transfers a geometric pattern from a mask to the surface of a silicon wafer using light-sensitive photoresist materials. This process is complex as it involves many engineering steps, for instance, chemical cleaning of the wafers, formation of barrier layer using silicon dioxide, and hard-baking process to increase photoresist adherence to the wafer surface. During the hard-baking process, the flow width of the photoresist is an important quality characteristic that needs to be monitored, as a minor variation (10 nm) in the thickness of photoresist will change the interference color and discolor the photoresist film.
Suppose that flow width can be controlled at a mean microns and the standard deviation microns of a normally distributed process and the quality practitioner anticipates an upward shift size = 0.3 in the process when the process is out-of-control. Then the upper-sided MOEWMA chart is implemented for the process monitoring at each sampling point. For the FSI (VSI) chart, the desired () is maintained as 200, and () = (0.1, 1.6) are selected.
In
Table 6, 20 samples, each with size
, are generated from an out-of-control normal distribution of the flow width with the mean
and the standard deviation
. All sample mean values
and the corresponding values of the different EWMA charting statistics are listed in the table. As a comparison, the upper-sided SEWMA, REWMA, and IEWMA
charts together with the MOEWMA
chart are plotted in
Figure 1. It can be noted from
Figure 1 that all control charts give an out-of-control signal at the 8th sample point, except for the upper-sided REWMA
chart, where the chart gives an out-of-control signal at the 9th sample point (see the bolded values in
Table 6). This example shows that these FSI EWMA charts take about 8 or 9 time units to detect the assignable cause while, on average, we can note from
Table 1 that the MOEWMA chart detect the shift
= 0.3 more quick than the SEWMA, REWMA, and IEWMA
charts.
Moreover, for the VSI-MOEWMA chart, the charting statistics and fall in the central region [0, ], which leads to a large sampling interval to find the subsequent samples. For the charting statistic at other sampling time point, the corresponding sampling interval is . This leads to a time unit of the VSI-MOEWMA chart to detect the assignable cause. Thus, it is better to adopt the VSI-MOEWMA chart to monitor the process.
6. Conclusions and Recommendations
In this paper, we study the performance of one-sided MOEWMA chart without- and with VSI features. Both the zero-state and steady-state performances of the FSI and VSI MOEWMA chart are investigated by using extensive Monte-Carlo simulations. Through a comprehensive comparison with the SEWMA, REWMA, and IEWMA charts, it is found that the MOEWMA chart is shown to perform better than the REWMA chart, especially for small shifts and it performs better than the IEWMA chart for small shifts and worse for moderate to large shifts. Moreover, the MOEWMA chart is always a little better than the SEWMA chart. In addition, by investigating the optimal performance of the MOEWMA chart, it can be concluded that the optimal MOEWMA chart has a good performance over a wide range of shifts rather than a scheme to optimize the control charts at a specified shift. Finally, by adding the VSI feature to the MOEWMA chart, it is shown that the VSI MOEWMA chart is uniformly better than its counterpart with FSI, especially for small shifts.
As the current research works are based on the assumption of known process parameters, the manner in which the control chart performs with estimated process parameters remains an issue. Future works could be extended to this aspect. Moreover, this research is focused on the monitoring of the process mean. The methodology can also be extended to monitor the process variance, the ratio of two distributions, and so on.