Enhanced process monitoring using machine learning-based control charts for poisson-distributed data

Abstract

The ability to detect shifts (e.g., outliers) in process monitoring is crucial for maintaining high-quality standards and operational efficiency in industrial environments. Control Charts (CCs) provide an organized framework for recognizing and managing anomalies, generally caused by assignable factors (e.g., identifiable issues) rather than inherent process variation. Traditional CCs, such as classical Shewhart, CUSUM, and EWMA, are commonly used to monitor Poisson observations in modern industries. The classical exponentially weighted moving average (EWMA) and cumulative sum (CUSUM) CCs are individually effective at detecting small-to-moderate shifts while Shewhart CCs identify moderate-to-large shifts in the process location and/or dispersion parameters. However, the classical CCs face limitations due to their sensitivity being constrained to specific ranges of shifts. To enhance the detection abilities of classical CCs in detecting all kinds of shifts in the process location parameter, this study proposes the integration of Machine Learning (ML) techniques into CCs to optimize the shift’s detection in process location parameter across a wider range. This study generates a dataset using the statistics of classical CCs based on Poisson-distributed data, which includes both in-control (stable process) and out-of-control (unstable process) processes. This dataset is used to train ML models, which are pre-processed through normalization and feature engineering through a heuristic approach before training. The performance of ML models is evaluated using standard regression metrics, specifically mean squared error and the coefficient of determination ($R^2$-score), to ensure effective generalization across varying process conditions. After training, these models are implemented within the proposed ML-based CC ($MLCC$) schemes. Their process monitoring capabilities are then rigorously compared with traditional and existing CCs, utilizing key performance indicators such as average run length and standard deviation of run length. These metrics are computed through a Python-based algorithm developed using the Monte Carlo simulation method. For practical purposes, implementing the proposed $MLCC$ schemes with real-life data in the food processing industry, specifically the packaging of frozen orange juice concentrate. This practical example highlights the superiority of proposed $MLCC$ schemes in the early detection of shift(s) in process location parameter(s) against classical CCs in real-life scenarios.