Design of attribute EWMA type control charts with reliable run length performance

Shu Wu, Philippe Castagliola, Athanasios C. Rakitzis, Petros E. MaravelakisCommun. Stat. Simul. Comput 2023 · DOI

Introduction

Statistical process monitoring (SPM) is a cornerstone of quality control, enabling practitioners to detect shifts in process parameters before they lead to significant waste. While Exponentially Weighted Moving Average (EWMA) control charts are highly effective for continuous data, their application to attribute data—discrete counts like defects or nonconformities—presents significant mathematical challenges. This paper addresses the instability of performance metric calculations for attribute EWMA charts by introducing a novel "continuousify" method that bridges the gap between discrete data and continuous numerical methods.

The Challenge of Discrete Data

In manufacturing and service industries, attribute data often follows Poisson or Binomial distributions. To increase sensitivity to small process shifts, engineers frequently employ EWMA schemes. However, designing an optimal EWMA chart requires accurate calculation of Run Length (RL) properties, particularly the Average Run Length (ARL).

Standard numerical methods, such as the Markov chain approach by Brook and Evans (1972), involve partitioning the control limits into $m$ subintervals. For continuous data, these methods converge smoothly as $m$ increases. For discrete data, however, the inherent "staircase" nature of the probability mass function causes the calculated ARL to fluctuate wildly and unpredictably based on the choice of $m$. This instability often forces researchers to rely on computationally intensive Monte Carlo simulations, which are too slow for the iterative process of finding optimal chart parameters.

The "Continuousify" Method

To resolve this, the authors propose the "continuousify" method. This technique smooths the discrete probability distribution by transforming the discrete random variable into a continuous one. Specifically, an observed discrete count $\omega$ is replaced by a continuous variable drawn from a Normal distribution $N(\omega, \sigma)$, where $\sigma$ is a small, user-defined smoothing parameter.

This transformation creates a continuous mixture distribution that retains the statistical characteristics of the original discrete data while providing the mathematical smoothness required for stable numerical integration. Once "continuousified," the standard Markov chain equations can be solved using matrix algebra, providing fast and reliable ARL and Standard Deviation of Run Length (SDRL) values for both Poisson EWMA (PEWMA) and Binomial EWMA (BEWMA) charts.

Practical Application and Datasets

The efficacy of the method was validated using two classic datasets from standard quality control literature (Montgomery, 2013):

  1. Printed Circuit Boards: A Poisson-distributed dataset tracking nonconformities across 44 samples of 100 boards.
  2. Orange Juice Concentrate Cans: A Binomial-distributed dataset tracking nonconforming items in samples of $n=50$ across 64 samples.

Key Insights

  • Stability through Smoothing: Adding a small amount of continuous noise (optimally $\sigma \in [0.1, 0.2]$) effectively eliminates the mathematical instability of the Markov chain method without distorting the underlying process statistics. If $\sigma$ is too small (e.g., 0.05), instability remains; if too large (e.g., 0.4), accuracy degrades.
  • Kernel Robustness: The choice of smoothing kernel (Normal, Uniform, or Parabolic) has a negligible impact on the final ARL results, suggesting the method is robust to different implementation choices.
  • Efficiency: By replacing brute-force simulation with stable matrix calculations, the method allows for the rapid generation of optimal design tables for various shift sizes, making it a practical tool for engineers.

Results and Discussion

The "continuousify" method successfully eliminated the erratic fluctuations seen in traditional discrete calculations. For a PEWMA chart where traditional methods failed to converge (yielding values ranging from -82.8 to 221.7), the new method stabilized the out-of-control ARL at 28.4 for $m \ge 100$. This result is highly accurate, closely matching the 28.3 value obtained from $10^6$ Monte Carlo simulations.

While the method slightly inflates ARL values compared to raw simulations, the difference is practically insignificant for quality control applications. The authors have leveraged this stability to provide comprehensive lookup tables for optimal smoothing parameters ($\lambda^*$) and control limits ($K^*$) for both PEWMA and BEWMA charts.

Conclusion

The "continuousify" method provides a robust bridge between the discrete reality of attribute data and the continuous requirements of efficient numerical analysis. By enabling fast, stable, and accurate performance evaluation, it empowers practitioners to design more reliable EWMA charts for monitoring count data. Future research may extend this framework to more complex distributions, such as Zero-Inflated Poisson or Negative Binomial models, and explore the impact of alternative smoothing kernels.

Summary updated

BibTeX
@article{wu2023design,
  author = {Shu Wu and Philippe Castagliola and Athanasios C. Rakitzis and Petros E. Maravelakis},
  title = {{Design of attribute EWMA type control charts with reliable run length performance}},
  journal = {Commun. Stat. Simul. Comput.},
  year = {2023},
  doi = {10.1080/03610918.2021.1955263},
}
Figure