Rework caused by quality issues frequently occurs in modern manufacturing, disrupting production flow and creating instability in both scheduling and maintenance planning. Such stochastic disturbances make traditional decoupled strategies inadequate, as maintenance decisions often ignore scheduling priorities, while scheduling neglects machine health conditions. To address this challenge, this study develops a simulation optimization framework that smartly integrates predictive maintenance (PdM) and dynamic scheduling using a multi-agent reinforcement learning (MARL) architecture. Four specialized agents—Maintenance, Production, Order Selection, and Coordination—jointly make decisions under uncertainty, guided by Remaining Useful Life (RUL) and Health Index (HI) indicators. The coordination agent dynamically resolves conflicts between maintenance and scheduling actions, ensuring feasible and cost-efficient joint decisions. To further enhance performance, an Expected Improvement Trust-Region Response-Surface Method (EI-TRSM) is incorporated to optimize MARL hyperparameters and balance the trade-off between total cost and on-time delivery rate (OTD). Simulation experiments on multi-stage production lines with stochastic arrivals and rework orders demonstrate that the proposed PdM–Scheduling integration significantly reduces maintenance and tardiness costs, improves delivery reliability, and exhibits robust adaptability compared with heuristic and rule-based baselines.

Keywords

predictive maintenance, dynamic scheduling, reinforcement learning, simulation optimization, response surface method