Adaptive real-time scheduling for production and maintenance: Integrating RUL prediction with multi-agent deep reinforcement learning

Production scheduling and machine maintenance are crucial for efficient production. Joint optimization of production and maintenance presents significant challenges due to the need to simultaneously consider time and resource constraints. Previous studies have attempted to address this issue by implementing fixed maintenance intervals and inserting production tasks accordingly. However, such methods often struggle to adapt to real-time changes in dynamic environments, leading to suboptimal performance. Therefore, this study focuses on the integration problem of dynamic production and predictive maintenance scheduling (DPPdMS). To solve this problem, an adaptive real-time scheduling method for production and maintenance (ARSPM) is presented. It combines a remaining useful life (RUL) prediction model with multi-agent deep reinforcement learning. Specifically, the prediction model is developed to accurately predict machine’s RUL. The predictive information derived from this model serves as the foundation for generating state features used by maintenance agent. Moreover, a multi-agent double deep Q network (MADDQN) is developed to train production and maintenance agents, with carefully designed state spaces, action spaces, and reward mechanisms. This enables production and maintenance agents to adaptively select dispatching rules and maintenance actions in real-time, respectively. Experiments demonstrate the efficiency of the proposed ARSPM through comprehensive numerical, statistical, stability, and case analysis.