Data-driven control chart optimization via reinforcement learning under multivariate dependence and noise

{Reinforcement Learning is applied to optimize multivariate control charts, with empirical evaluation using industrial data from a Siemens gas turbine case study within a realistic simulation framework derived from actual sensor logs.}Multi-component systems consist of interdependent subsystems operating under shared loads, each generating multiple process and vibration signals. Consequently, intra- and inter-component correlations, temporal autocorrelation, and environmental variability can propagate local degradation into system-level quality and availability losses. This study introduces an integrated monitoring and decision-making framework and validates it through a real-world case study of a Siemens gas turbine operating in a combined-cycle power plant. In the monitoring layer, multivariate signals are stabilized using time-series filtering to remove short-range autocorrelation, and a multivariate control statistic is computed on approximately uncorrelated residuals. Control limits are empirically calibrated to achieve target in-control performance, ensuring fair comparison across subsystems and accurate sensitivity to correlation and noise effects. In the decision layer, a safe reinforcement learning agent uses normalized monitoring evidence, including deviation-to-limit ratios, trend indicators, and operational features such as temperature, vibration, and fuel flow, to determine when to continue, inspect, or perform preventive maintenance in order to minimize total cost while maintaining system availability. A deterministic safety rule supplements the agent to ensure conservative actions under extreme deviations. The results of the case study demonstrate that the proposed framework maintains consistent in-control behavior, improves the cost–availability balance compared with fixed rules, reduces detection delay, and shifts maintenance spending from costly corrective actions toward timely preventive interventions, achieving competitive and interpretable performance under realistic operating conditions, with statistically significant improvement over Shewhart and age–based rules but comparable performance to the condition–belief baseline.