Generally weighted moving average monitoring schemes: Overview and perspectives
Summary (5 min read)
1. Introduction
- In any field, a literature review is essential as it gives an overview of what investigations have been conducted and what are some of the possible research ideas that can be pursued.
- Thus, the tradition emphasized by the latter articles is continued with the generally weighted moving average (GWMA) monitoring schemes in this paper.
- The basic properties of the GWMA scheme and its existing enhancements as well as a detailed outline of how the review is structured are provided in Section 2.
2.1 Operation of the GWMA scheme and its enhancements
- (independent and identically distributed) normally distributed random variables with in-control (IC) mean and standard deviation √ .
- The operation of the GWMA scheme and its enhancements are illustrated for the subgroup mean.
- Thus, the asymptotic upper and lower control limits (denoted as and ): EQUATION Following a similar line of argument, it follows that the charting statistics, design parameters and control limits of the DGWMA, Shewhart-GWMA, GWMA-CUSUM and CUSUM-GWMA ̅ schemes are as shown in Table 1 .
- An illustration of different charting statistics, design parameters and control limits of the GWMA scheme and its existing enhancements when the characteristic of interest is the process mean, also known as Table 1.
Asymptotic
- Note that when the underlying process parameters are known (i.e. Case K), control limits can easily be calculated and the process monitoring can take place immediately; however, when process parameters are unknown (i.e. Case U), a two phases approach needs to be implemented, see Jensen et al 18 and Psarakis et al 19 .
- In Phase II, monitoring schemes are implemented prospectively to continuously monitor any departures from an IC state using the parameters estimated in Phase I. Even though Table 1 and Figure 1 provide properties pertaining to GWMA-related schemes for the process mean only, other types of statistics can be done in a similar manner.
.2 Outline of the review
- The journals or proceedings that have published researches on GWMA-related schemes and / or their enhancements are provided in Table 4 (as well as the total number of publications in each journal or proceedings).
- In Table 5 , the list of researchers with at least two publications on GWMA-related schemes and / or their enhancements have been provided, including their corresponding affiliations.
Process mean
- Apart from introducing the GWMA scheme to the SPM literature, Sheu and Lin 9 showed that when =1, the GWMA ̅ scheme reduces to the EWMA ̅ scheme; whereas when =1 and =0 it reduces to the Shewhart ̅ scheme.
- Apart from showing that the GWMA ̅ scheme outperforms the corresponding EWMA ̅ scheme in many situations, Sheu and Lin 9 demonstrated that it also outperforms the Shewhart ̅ scheme with and without runs-rules.
- For processes with start-up problems, a FIR approach is also incorporated and it is shown that it uniformly outperforms the basic GWMA scheme; however, it outperforms the EWMA ̅ scheme with AIB in most situations.
- The abovementioned articles in this subsection are based on i.i.d. observations used to compute the process mean.
- Using ARL and SDRL metrics, this extension is shown to be superior to the GWMA ̅ scheme in detecting medium-to-large shifts; but for small shifts, the converse is true.
Process median
- Unlike the process mean, the process median tends to be preferred in some situations due to its outlier-resistant property.
- Sheu and Yang 29 showed that the GWMA median scheme outperforms the corresponding EWMA (including the one with FIR) and the Shewhart median schemes in many situations.
- Note though, the GWMA ̅ scheme has a better OOC detection ability while the GWMA median scheme performs best when the process is subjected to outliers.
- Moreover, the quality cost model implemented with contaminated normal distribution shows that the GWMA median scheme is more economical with respect to the average quality cost as compared to the GWMA ̅ scheme.
Time-between-events
- To ensure ARL-unbiased OOC performance, Chakraborty et al 30 proposed a one-sided GWMA scheme for monitoring time-between-events (TBE) based on the gamma distribution for Cases K and U.
- It is observed that the estimation of the unknown parameter from an IC Phase I sample affects the GWMA TBE scheme"s performance in Phase II.
- Also, it has a better steady-state ARL performance than the DEWMA TBE scheme for smallto-moderate shifts; but the converse is true for other shifts.
- In some situations, the GWMA TBE scheme has a slight advantage over the DGWMA TBE scheme for (very small)-to-small shifts, but the converse is true for other shifts.
Attributes data
- When the characteristic to be monitored is in the form of count data and under the assumption of a Poisson distribution; using simulation, Sheu and Chiu 31 showed that while the GWMA scheme outperforms the corresponding Shewhart and EWMA schemes in detecting all shifts values, it outperforms the DEWMA scheme in detecting small-to-moderate shifts only.
- Moreover, the simulation approach tends to be time consuming in terms of CPU time as compared to Markov chain approach, i.e. takes 10-to-60 minutes per case study for simulation as compared to 1-to-2 minutes for Markov chain approach.
- When the characteristic to be monitored is in the form of number nonconforming and under the assumption of a binomial distribution; using both Markov chain and simulation approaches, Sukparungsee 35 proposed the GWMA scheme and showed that it outperforms the corresponding EWMA scheme for small shifts.
- In the case of high quality processes, where a large number of zero nonconforming observations exist, the and schemes result in excessive false alarms, hence a zero-inflated Poisson (ZIP) distribution is often used to model and monitor count data with such a large number of zeros.
- A comparison with the corresponding upper-sided Shewhart, CUSUM and DEWMA schemes indicates that, in many situations, the GWMA ZIP scheme is more effective when both parameters are monitored.
3.1.2 Variability
- Sheu and Tai 40 proposed the GWMA scheme and showed that it outperforms the corresponding EWMA scheme in detecting OOC observations especially for small shifts.
- Next, Sheu and Lu 41 used an unbiased estimator of the process variance to propose the one-sided GWMA scheme and showed that it yields better ARLs than the basic EWMA and exponential weighted mean square deviation schemes.
- Next, Ali and Haq 42 proposed a one-and two-sided GWMA schemes which are based on applying the normal approximation to the distributions of the logarithms of the weighted sum of chi-squared random variables.
- It is observed that a properly designed one-or two-sided GWMA scheme has a better OOC performance than the corresponding EWMA schemes in detecting small shifts.
- Finally, it is worth mentioning that Chakraborty et al 30 briefly discussed how the GWMA TBE scheme can be used to monitor downwards shifts in the variance for normally distributed data in Case K and they commented on how this can be extended to the Case U scenario.
Mean and variance using separate charting statistics
- Next, the corresponding studies to jointly monitor the process mean and variability of residuals & autocorrelated actual observations were investigated in Sheu and Lu 45, 46 , respectively; assuming an ARMA(1,1) model for the time series data.
- In each investigated instance, the GWMA schemes were shown to have a better OOC ARL performance than the corresponding EWMA schemes.
Mean and variance using a single charting statistic
- The transformed mean and variance values of the Max statistic are obtained using an inverse standard normal distribution and a chi-square distribution, respectively.
- Max scheme can be found in Nguyen et al 49 and Hsu et al 50 using the thickness of the coating layer of tablets and the quality of displays from Thin Film Transistor -Liquid Crystal Display (TFT-LCD) datasets, respectively.
- In many situations, the results showed that the GWMA Max scheme has lower false alarm rates (or similarly, higher IC ARLs) for more levels of skewness when compared to the EWMA Max scheme.
- Max scheme; note though, the diagnostic analysis was not considered.
Coefficient of variation
- For processes where the mean and variability parameters vary in a fixed proportional way when the process is IC, it is more reasonable to monitor the coefficient of variation (CV).
- Using the ARL, SDRL and ratios of the ARLs, it is shown that, in many situations, the GWMA CV scheme has a better OOC performance as compared to the EWMA and DEWMA CV schemes.
- Hong et al 60 incorporated a FIR feature to the GWMA CV scheme and they showed that it has a significantly improved OOC performance compared to the DEWMA and GWMA schemes without a FIR feature.
3.2 Nonparametric schemes
- Currently, the only contributions that consider the monitoring of a nonparametric statistic using a GWMA scheme are all dedicated to the monitoring of the process location.
- Taking into account the shapes of the Normal, Student"s -and Gamma distribution, Chen et al 62 used the ARL and ASS metrics to show that the GWMA sign scheme with repetitive sampling outperforms the corresponding EWMA scheme based on SRS and repetitive sampling methods as well as the GWMA scheme based on SRS in detecting small shifts.
- As an alternative to the two-sample location shift parametric t-test, the Exceedance (EX) and Wilcoxon rank-sum (WRS) tests are usually recommended when the underlying process distribution is non-normal.
- Next, the GWMA WRS scheme was proposed by Mabude et al 66 , with the Normal, Student"s t-, Gamma, Log-logistic and Weibull distributions used to show the IC robustness and to study the OOC performance.
- Also, the effect of the Phase I reference sample size on the Phase II test samples" OOC performance is provided.
4.2 Nonparametric schemes
- The research works that exist for DGWMA-type nonparametric schemes are on monitoring the location parameter only.
- Lu 73 proposed a DGWMA sign scheme and showed that it is more sensitive than the corresponding EWMA, GWMA and DEWMA schemes in many situations.
- Phanyaem 74 briefly discussed the DGWMA sign scheme with repetitive sampling and showed that it is more sensitive than the corresponding DEWMA sign scheme in many situations.
- The same distributions as those in Chakraborty et al 65 are considered and the DGWMA EX scheme is observed to have more favourable results than the GWMA and EWMA EX schemes.
5. Shewhart-GWMA schemes
- Currently, no research work exists for nonparametric Shewhart-GWMA schemes.
- For parametric ones, some based on monitoring location as well as location and variability exist.
6. GWMA-CUSUM schemes and its reverse version
- Currently, no research work exist for nonparametric statistics using GWMA-CUSUM schemes.
- For parametric ones, some based on monitoring location and monitoring variability exist.
6.1 Location
- Lu 16 proposed the mixed GWMA-CUSUM ̅ scheme and its reverse version (CUSUM-GWMA ̅ scheme) to monitor individual observations.
- It is shown (using ARL only) that, in many situations, this scheme provides better OOC detection ability than the basic GWMA, EWMA and CUSUM schemes, as well as the mixed EWMA-CUSUM and CUSUM-EWMA ̅ schemes.
- Moreover, the CUSUM-GWMA scheme is shown to yield a slightly better OOC ARL performance over the corresponding GWMA-CUSUM scheme in many situations.
- At the same time, but independently, Ali and Haq 17 also proposed the GWMA-CUSUM ̅ scheme to monitor the process mean; however, of subgroup observations.
- Note though, Ali and Haq 17 did not consider the reverse CUSUM-GWMA ̅ scheme.
6.2 Variability
- Ali and Haq 78 proposed the mixed GWMA-CUSUM scheme (with and without FIR) to monitor upwards shifts for subgroup observations and showed that it yields better OOC detection ability than the CUSUM and mixed CUSUM-EWMA schemes.
- Note that Ali and Haq 78 did not consider the mixed CUSUM-GWMA scheme.
- Next, Huang et al 79 proposed the mixed GWMA-CUSUM scheme and the reverse version to monitor upwards shifts using individual observations.
- Both of the mixed schemes have better small shifts detection ability than many competitors, including those discussed in Ali and Haq 42 and, Sheu and Lu 41 .
- More importantly, it is observed that the CUSUM-GWMA scheme has a slight advantage over the corresponding GWMA-CUSUM scheme in many situations.
7. Concluding remarks
- The GWMA scheme is an extended version of the EWMA scheme with an additional adjustment parameter.
- Note that, with the aid of computer programs, the complexity in implementation can be significantly simplified.
- Effort needs to be paid to demonstrating how to monitor real-life datasets using the different GWMA-related schemes discussed herein.
3. As observed in
- Only a few attribute GWMA monitoring schemes have been proposed.
- There is no study on the economic and economic-statistical designs of the GWMA monitoring schemes that has been done in the literature so far.
- Therefore, researchers are encouraged to investigate these topics under both i.i.d. and correlated observations as well as for Case U. 12.
- Only one publication on multivariate schemes is available in the literature.
- The weight structure of the observations for the EWMA and GWMA schemes are similar, i.e., geometrically decreasing; however, the weights structure or weight function kurtosis used by the GWMA scheme differs from that EWMA scheme due to the additional adjustment parameter.
5. The design parameter found in
- Step 4 is called the optimal design parameter.
- Record the optimal and its corresponding control limits.
- Calculate the charting statistic(s) and compare to the control limit(s) found in Step 5.
- Then record the number of subgroups plotted until an OOC signal occurs.
- This number represents one value of the distribution.
Did you find this useful? Give us your feedback
Citations
15 citations
14 citations
Cites background or methods from "Generally weighted moving average m..."
...Thus, the DEWMAW chart produces an OOC signal when Dt value plots beyond the control limits in (8)....
[...]
...Hence, the upper and lower control limits of the DEWMAW chart are computed as follows: UCLt∕LCLt = μD ± LDσDt , (8) where LD > 0 is the control limits coefficient, which has to be calculated such that the attained ARL0 is close or equal to the prespecified nominal ARL0....
[...]
10 citations
9 citations
Additional excerpts
...Additionally, recent developments of distribution-free EWMA-type schemes can be found in Raza et al.,18 Alevizakos et al.,19 Mabude et al.,20 Perdikis et al.21 The use of reliablemetrics that can efficientlymeasure the performance of a control schemeplays a vital role in its design phase....
[...]
9 citations
References
1,624 citations
851 citations
532 citations
480 citations
Related Papers (5)
Frequently Asked Questions (19)
Q2. Why is the use of a composite Shewhart-GWMA scheme not advised?
Due to the complexity in implementation and excessive false alarms along with a very minor OOC improvement, the use of composite Shewhart-GWMA ̅ schemes with runs-rules is not advised.
Q3. What distributions are used to show that the GWMA sign scheme has a better O?
Using the Normal, Student‟s -, Logistic and Uniform distributions, it is shown that the GWMA signed-rank scheme outperforms the GWMA sign scheme in many situations; however, when using the Laplace distribution, the GWMA sign scheme has a slightly better small shifts detection ability.
Q4. What are the main statistics used for GWMA monitoring?
For joint monitoring of the process mean and variability, there are many test statistics used (i.e.Max, Semi-circle, Sum of squares, separate charting statistics, etc.).
Q5. what is the relevance of multivariate schemes in real-life applications?
Given the relevanceof multivariate schemes in real-life applications, there is a lot of research works on GWMA schemes that need to be done based on parametric and nonparametric settings.
Q6. What is the common example of a random shock in a ZIP model?
In a ZIP model, some random shocks occur independently with probability and the number of nonconformities follows a Poisson distribution with parameter .
Q7. What are the main characteristics of the GWMA scheme?
Using the ARL and average sample size (ASS) metrics, it is shown that it outperforms the corresponding GWMA and hybrid EWMA schemes based on the SRS method in detecting small shifts.
Q8. What is the significance of the moving average?
the moving average tends to be a representation of the more recent process performance, as larger weights are allocated to the most recent observations.
Q9. What are the other options for adaptive EWMA schemes?
adaptive EWMA schemes also exist in SPM literature (i.e. variable sample size (VSS), variable sample interval (VSI), variable sampling size and interval (VSSI)).
Q10. What is the significance of the GWMA-CUSUM scheme?
More importantly, the GWMA-related monitoring schemes can be useful for quality practitioners in a variety of applications where the EWMA-related schemes are being currently used, as replacements.
Q11. How many publications have been published on GWMA?
Since 2003, the year of publication of the first article, there have been a total of 61 publications on the GWMA-related monitoring schemes and their enhancements.
Q12. Is it recommended to use it in real-life applications?
Since the implementation of the composite ShewhartGWMA scheme is relatively complex, it is neither not advised to use it in real-life applications.
Q13. What is the corresponding steady-state performance of the GWMA scheme?
The corresponding steady-state performance is discussed in Chiu and Lu 69 , where it is shown that it is preferred for downward shifts, while the GWMA scheme is more competitive for upward shifts.
Q14. What are the different types of defects in a ZIP?
For instance, defects are classified in terms of categories or classes, e.g. „Very serious‟, „Serious‟, „Moderately serious‟ and „Minor‟.
Q15. How can the GWMA scheme be extended to the case U scenario?
it is worth mentioning that Chakraborty et al 30 briefly discussed how the GWMA TBE scheme can be used to monitor downwards shifts in the variance for normally distributed data in Case K and they commented on how this can be extended to the Case U scenario.
Q16. What are the alternatives to the two-sample location shift parametric t-test?
As an alternative to the two-sample location shift parametric t-test, the Exceedance (EX) and Wilcoxon rank-sum (WRS) tests are usually recommended when the underlying process distribution is non-normal.
Q17. What are the known enhancements of the GWMA-related scheme?
So far, the existing known enhancements of the GWMA-related scheme are: the double GWMA scheme – denoted as DGWMA scheme; the composite Shewhart-GWMA scheme; the mixed GWMA-CUSUM scheme and its reverse version, the mixed CUSUM-GWMAscheme.
Q18. What is the way to study the GWMA Max scheme?
Using the ARL and SDRL, it is shown that the GWMA schemes have a better detection ability than the corresponding EWMA schemes, especially for small shifts; however, they have similar diagnostics abilities.
Q19. What is the need for GWMA monitoring?
R programs or any other commercial / open source statistical software for any general charting statistic need to be made readily available so that more research can be fast-tracked for GWMA-related monitoring schemes.