1. Introduction
Quality management provides many operational and management techniques that save both time and cost to achieve the standard finished product. These techniques are used in manufacturing processes, filling processes, and in services to spot the un-natural variations that improve the quality of finished products. A process can be defined as “systematic way of a number of inputs and a required output” (cf. [
1]). For example, in semiconductor manufacturing processes, the controllable input variables are photolithography, temperature, silicon wafer, resistance, and some other process variables. The flow width of the resistance is a running process that has some quality characteristics. These quality characteristics are monitored using various statistical tools to improve and produce quality products. For this purpose, statistical process control (SPC) plays an important role. SPC tools are frequently used in the fields of industries, natural sciences, environmental sciences, medicines, and services to monitor the processes. A control chart has a vital role in SPC that is often used in process monitoring to identify any nonconformity in the process parameter. Variation is an unavoidable part of every process output that reduces quality. Common causes of variation do not affect the conformance of the process, while special causes of variation influence the process outputs (cf. [
1]). Control charts help us in the early detection of assignable variations that reduce the financial penalties. These abilities of SPC are helpful in producing the quality products in the aforementioned fields.
The concept of control chart, to improve the quality of the process output, was first initiated by Walter A. Shewhart in the early 1920s. Since then, the control chart has a vital role in modern SPC. The Shewhart control charts proposed by [
2] are frequently used where large changes occur in process target(s). However, these charts are unable to spot the small and persistent shifts in the process location sharply. Today, advance process monitoring schemes include the cumulative sum (CUSUM) and the exponentially weighted moving average (EWMA) control charts. These are memory-type control charts, as their charting structures are a combination of current and past information in the sample. Although their performance toward small and moderate shift detection is almost same, but most of the quality practitioners prefer the EWMA chart due to its simplicity. The use of these control charts is increasing due to their sensitive nature toward sustained changes in the process parameters. For this purpose, these are regularly used to spot small changes in the process location and scale parameter(s), where small changes may cause serious quality issues.
In practice, it is assumed that the quality characteristic of interest has some distributional assumptions, usually normality assumptions. In real situations, we negate this assumption, and so the usual parametric control charts cannot be implemented with certainty, as these produce more false alarms and invalid comparison. In modern process monitoring schemes, non-parametric (NP) control charts (or distribution-free charts) are a more robust alternative to the parametric charts when the actual process distribution is doubtful. The main advantages of these NP charts is that they do not have any distributional assumption (i.e., normality assumption), while their in-control (IC) run length (RL) distribution remains constant for all the continuous process distributions (cf. [
3]). This characteristic of NP charts provides full knowledge and stability of IC and out-of-control (OOC) properties that are compulsory to implement a control chart. According to these limitations, a number of studies were proposed on NP design structures for process monitoring. Some of the studies have used sign (SN) and signed rank (SR) tests, which are the simplest ones under the NP structure (cf. [
4]). Using a SN test, Ref. [
5] proposed an NP Shewhart-type chart for monitoring the process target. A distribution-free Shewhart chart based on an SR test was proposed by [
6] using grouped observations to monitor the process mean. These Shewhart-type charts under NP structure were also discussed in [
7] and provided sufficient knowledge about their IC RL properties. A phase-II Shewhart-type NP chart based on Mann–Whitney statistic, for monitoring the process target, was proposed by [
8]. Using runs rules, two NP Shewhart-type control charts were proposed by [
9] for monitoring the unknown process location. However, these charts efficiently performed for large shifts only and failed to detect small and persistent shifts in the process target. The SN and SR tests were also used in a CUSUM control chart and have shown better performance than the Shewhart-type control chart to monitor the process location (cf. [
10]). An NPCUSUM chart based on exceedance statistic was proposed by [
11] for monitoring the process location. The only drawback of the CUSUM chart is that it involves very complicated calculations making it difficult to use, whereas the EWMA charts have the same level of performance with a simple design structure and are easy to apply in practical scenarios. For more details on NP control charts, readers and practitioners are referred to [
4,
12].
The most famous and frequently used EWMA chart under NP structure suggested by [
13] for monitoring shifts in the process target/median was based on a simple sign test statistic. Due to the sensitive nature of the EWMA chart toward small and persistent shifts, different modifications have already been proposed and tested in the non-parametric exponentially weighted moving average (NPEWMA) charting structure. The authors of [
14] designed an NPEWMA chart that was based on an SR test (NPEWMA-SR) for monitoring small and persistent shifts in the process target/median. A NP synthetic EWMA (SynEWMA) sign chart was proposed by [
15] that improved the sensitivity of NPEWMA with the conforming RL chart. For more information on NP extended EWMA charts, readers and practitioners are referred to [
16,
17] and the references therein. All the above-mentioned extensions were based on simple random sampling (SRS) scheme, whereas the charts based on the ranked set sampling (RSS) scheme are more sensitive in nature and are often more economical. A lot of work needs to be done in this area to offer more sensitive and economical design structures.
In the modern monitoring schemes, financial control has become a key point for quality practitioners that are used for reducing the quality cost. Quality engineers are always trying to minimize the difference between actual and budgeted cost. The SRS technique is frequently used in SPC but sometimes it has a high cost and provides less efficient results. In such situations, practitioners need a better alternative to SRS. The RSS technique provides efficient results and is also helpful for reducing the prevention cost.
The idea of RSS was initiated by [
18] in situations when measurements are very expensive or of a distractive nature. In real life, it is apparent in many situations that the actual measurements of datasets are very expensive, while the visual inspection done by an expert (perfect ranking) is easy and inexpensive. As an example, in different manufacturing processes, the actual measurements of each quantity are costly and time consuming, while the visual inspection can save time and is relatively inexpensive. In control charting, RSS schemes are mostly used to increase the sensitivity of the monitoring process. RSS estimators are more efficient and unbiased as compared to the SRS scheme. The sample selection criteria under RSS schemes are not problematic. The criteria have some similarities to stratified random sampling, in which we divide the whole population into different subgroups where within groups objects are homogeneous, and between, they are heterogeneous. The following steps are involved to draw a random sample of size
using the RSS scheme. (i) Draw an
size of samples from the targeted population/process. (ii) Divide these samples into
subgroups in an arbitrary way where each sample has size
. (iii) Rank each observation given in samples in ascending order. (iv) Choose the first observation from the first sample, the second observation of the second sample, and the last observation of the last sample. (v) It provides a ranked set sample of size
with
cycle.
The efficiency of the control chart is mostly enhanced by using Fast Initial Response (FIR) (cf. [
19]) Auxiliary information (cf. [
20,
21]) and RSS scheme (cf. [
22,
23,
24,
25]).
This study intends to increase the sensitivity of the NPEWMA-SN chart based on the SN test statistic developed by [
26] using an RSS scheme (NPREWMA-SN chart hereafter). The proposed NPREWMA-SN chart is developed under a process of standard normal, student’s
t, logistic, Laplace, and contaminated normal distributions. The main purpose of this study is to develop a simple, sensitive and inexpensive chart for monitoring the small deviations in the process location. In this article, the median is used as a target value because the mean has great deficiencies in the presence of outliers (cf. [
27,
28]). The key advantages of the proposed NPREWMA-SN chart over its counterparts can be described as follows. (i) There is no need of symmetry assumption for the proposed scheme (as is required for some others). (ii) The proposed scheme can be used to monitor the known mean, median, or percentiles of the distribution. (iii) The proposed scheme can also be performed efficiently for heavy tailed distributions. (iv) There is no need for actual measurements of observations, just a need for the information that the actual measurement is greater or less than the targeted value. (v) There is no need for process variance specification. (vi) It is easy to find the IC and OOC distribution of the sign (SN) statistic that is binomial. (vii) The efficiency of the designed NPREWMA-SN method is compared with its existing counterparts and it is found to be superior.
The remaining article is outlined as follows.
Section 2 provides the structure formulation of the proposed chart.
Section 3 describes the estimation procedures of the design parameters.
Section 4 deals with RL evaluations and robustness. The OOC performance of the proposed chart is discussed in
Section 5.
Section 6 provides extensive comparisons of the NPREWMA-SN chart with its existing competitors.
Section 7 illustrates a numerical example using a real dataset. The article ends with a summary and concluding remarks.
3. Evaluation
The measures used for evaluating and comparing the proposed scheme are the average RL (ARL), standard deviation RL (SDRL), P5, P25 (first quartile Q1), P50 (median RL(MRL)), P75 (3rd quartile Q3), and P95. ARL is a statistical measure used as a yardstick to assess RL properties of the NPREWMA-SN method. The total number of sample observations before a chart alarms is defined as RL and the mean of the RL distribution is called ARL. The IC ARL of a chart is denoted by ARL0 and OOC is nominated as ARL1. A chart is considered efficient that has a smaller ARL1 at a specific shift. The ARL plays a leading role in RL evaluation but, due to the skewed nature of RL distribution, important information may be missed if all the aforementioned measures have not been considered. The RL characteristics of the proposed chart are calculated using the Monte Carlo simulation method, and codes are developed in R programming language applying 50,000 iterations. The Monte Carlo simulation is used and the complete algorithm to find a nominal ARL0 is as follows:
- (i)
Using an RSS scheme, we generate subgroups of size and repeat them for cycles to acquire from continuous distributions that are under study.
- (ii)
Choose the values of smoothing parameter and for a specified ARL0 and shift size.
- (iii)
Perform the NPREWMA-SN plotting statistic given in Equation (5) at each subgroup.
- (iv)
Construct the control limits given in Equation (6) and plot the NPREWMA-SN statistic used in step (iii) on these control limits.
- (v)
Observe the process until an OOC point appears.
- (vi)
Fixed the value of the multiplier at the required IC ARL0 (370, 500).
- (vii)
Repeat these steps 50,000 times and estimate the RL profiles of the process.
Selection of Design Parameters and Implementation of the Proposed Scheme
The proposed scheme has four design parameters , where represents the number of cycles used in the RSS scheme, represents the size of the subgroups, is the smoothing parameter, and is the size of the control limits (also known as a multiplier of the control limits). These parameters play a vital role in obtaining the optimal RL profiles of the proposed scheme. As mentioned above, a chart is considered more efficient that has a smaller value of ARL1 at a specific shift using a fixed value of ARL0. According to this rule, two steps are involved in the selection of design parameters. In first phase, those combinations of the design parameters are selected that give desired fixed values of ARL0 i.e., (100, 168, 370, or 500). Then, the value of the smoothing parameter is finalized that gives a minimum ARL1 at a specific shift in the process location. It is understood that the EWMA chart performs effectively at small values of smoothing parameter , but large values can be used for the purpose of forecasting.
Using the above-mentioned criteria, all possible measures of the RL distributions are estimated for the proposed scheme with the combinations of design parameters
,
,
, and size of the limits
that are reported in
Table 1.
The first line of
Table 1 shows the evaluated values of ARL
0 and SDRL
0, while P
5, P
25, P
50 = MRL, P
75, and P
95 are reported in the 2nd line of the table, respectively, where P
k (k = 1, 2, 3, …, 100) are the percentile values of the RL distribution (cf.
Table 1). It can be seen in
Table 1 that at a fixed value of the smoothing parameter (
), the values of IC RL (reported in
Table 1) of the proposed scheme are rapidly increasing as the value of
increases. For example, the ARL
0 is 518.34 at
; when
is increased from 2 to 3, the ARL
0 becomes 4960.91 for
. This difference is very high at small values of
but relatively low at large values of
. These results confirm the theory that the EWMA chart performs better at small values of
.
Table 1 is also helpful in identifying the shape of the RL distributions, where ARL
0 MRL
0, which indicates that the RL distributions are positive skewed. If a quality practitioner want to fix the desired ARL
0, these values are very helpful to select the values of design parameters
. For example, to fix ARL
0 370, when
, the value of
should lie between 2.4
2.5 at
(cf.
Table 1).
The values of ARL
0 are plotted in
Figure 1 at five small and moderate values of
that give an attractive visual look to easily understand the effect of
on ARL
0. It can be observed from
Figure 1 that as we increase the value of
, the values of ARL
0 increase and give the highest ARL
0 at
.
7. Real Data Application of the Proposed Scheme
For the illustration of the NPREWMA-SN chart, a real-life application is presented to observe the shift detection ability in the process monitoring. When manufacturing semiconductors, an important hard bake process is used with the combination of photolithography. The flow width of the resist in manufacturing substrate process is taken as a quality characteristic of interest. The different steps involved in this manufacturing process are displayed in
Figure 6. To monitor this process, a sample of five wafers was taken every hour, and the average of flow width was computed. The dataset provided by [
1] consists of the first 25 ICs and next 20 OOC sample observations of size 5. We mixed 125 IC and 100 OOC observations to get 25 ranked set samples from IC and 20 from OOC observations of size
with
(cycles). From the resulting 45 samples, the first 25 samples are IC with a target value
, while the remaining 20 samples are from the OOC process. From the Phase-I samples, the control limits of the proposed chart and its alternatives are computed. In the monitoring phase, the proposed and the competitors are also constructed.
For the implementation of the proposed and competitors, the design parameters for various charts are given as for NPEWMA-SN, for classical EWMA-, and for the proposed charts at fixed value of and IC ARL ≅ 370.
All the aforementioned charts are displayed in
Figure 7,
Figure 8 and
Figure 9. The proposed chart triggers the first OOC signal at the 38th measurement, while the NPEWMA-SN and classical EWMA-
control charts detect the first OOC signal at the 45th and 41st sample numbers, respectively (cf.
Figure 7,
Figure 8 and
Figure 9). The proposed scheme’s quicker detection of an OOC signal provides valuable confirmation of its superiority over the other existing alternatives.
8. Summary and Conclusions
SPC tools are used to detect aberrant changes in the process parameters, and the control chart is one of them. Control charts are designed under the normality assumptions of the process, but in many situations, normality assumption does not meet or process distribution is unknown. In such situations, the need for distribution-free control charts arises, as these charts are IC robust for all the continuous distributions. To make the charting structure more efficient, different extensions and modifications have been made in the literature, and the RSS scheme is one of them. In this study, a new NP monitoring scheme namely, NPREWMA-SN, based on a sign statistic under an RSS scheme has been proposed for monitoring small and persistent shifts in the location parameter. It is observed that the proposed chart provides a sensitive design structure for efficient process monitoring. Moreover, the RSS scheme provides an unbiased, efficient, and inexpensive estimate of the process parameter(s). The comparisons of the proposed method with NPEWMA-SN and classical EWMA- schemes advocate its dominance over its competitors. An illustrative example on the substrate manufacturing process is also provided in support of the proposed study. It concludes that the proposed scheme has performed better toward small and persistent shifts detection at small values of smoothing parameters. The OOC performance of the proposed chart in Laplace (sharped peak) and logistic distributions is far better. The OOC performance of the proposed chart under the normal, heavy-tailed ‘t’ and contaminated processes is comparatively less effective. However, the proposed chart at small and moderate shifts in the process location performs uniformly better than existing counterparts under all the process distributions incorporated in this study. This work can be extended for estimation of the process parameters, multivariate scenarios, dynamic profiling analysis, etc.