Deep Reinforcement Learning for Optimal Planning of Production Line Maintenance With Deterioration  

Manufacturing companies often contend with deteriorating production systems that can significantly hinder overall performance. Therefore, optimizing maintenance planning strategies plays a crucial role in enhancing production efficiency. In this study, we address a novel maintenance decision-making problem in serial production lines, where machine degradation leads to product quality deterioration and throughput loss at the final stage. We propose a new framework that jointly considers product quality trends, buffer dynamics, and machine production states to optimize a preventive maintenance policy for the last machine in the line. To achieve this, we utilize an average-reward deep reinforcement learning (DRL) approach within a discrete event simulation environment and compare their effectiveness to conventional dispatching methods. Our results, based on a digital twin enriched with real production data, show that the proposed DRL approach consistently outperforms traditional dispatching rules and practitioner-designed heuristics. This study is the first to apply DRL to maintenance decision-making in serial production systems that explicitly model product quality degradation, machine production states and buffer interactions, addressing a key gap in current literature.