A model–free reinforcement learning framework for statistical process monitoring with variable sampling rate

This study presents an artificial intelligence (AI)-driven reinforcement learning (RL) framework for statistical process monitoring with variable sample size (VSS) and variable sampling interval (VSI) as a data-driven alternative to classical adaptive control charts. Within this framework, the monitoring problem is formulated as a Markov Decision Process (MDP), enabling sampling decisions to be learned directly from the agent's interaction with the process environment rather than through application-specific analytical control-chart design. The proposed framework optimizes the sequential trade-off between detection speed and sampling effort under process uncertainty. Neural network-based models (NNQ-VSS and NNQ-VSI) consistently outperform Q-table variants (Q-VSS and Q-VSI) and achieve competitive and, in many cases, improved monitoring performance relative to adaptive control charts in numerical evaluations using standard Statistical Process Control (SPC) metrics. In particular, the NNQ models demonstrate improved shift-detection performance while maintaining higher sampling efficiency. The framework incorporates an automatic tuning mechanism for reward-related cost parameters, enabling faster and more stable learning of sampling policies while satisfying process monitoring constraints, namely the false-alarm and the sampling-rate constraints. Its practical applicability is further demonstrated through an industrial case study based on the Tennessee Eastman Process benchmark, where the proposed framework exhibits clear advantages in a realistic environment characterized by non-normality, autocorrelation, and complex process dynamics. Owing to its inherently model-free formulation, the proposed framework provides a flexible foundation for monitoring non-normal or autocorrelated processes, while the data-driven design facilitates extension to multivariate monitoring and simultaneous tracking of process mean and variance without requiring new analytical control-chart design formulations.