Joint Optimization for Predictive Maintenance and Spare Parts Management via Multiagent Reinforcement Learning

Prognostics and health management (PHM) has emerged as a critical discipline in modern industrial systems, integrating condition monitoring, degradation prediction, and maintenance optimization to prevent catastrophic failures while minimizing operational costs. However, current PHM methodologies predominantly treat remaining useful life (RUL) prediction and maintenance planning as sequential rather than coupled processes. To address this limitation, this article proposes an integrated framework that bridges prognostic prediction with operational decision-making through predictive-driven dynamic coordination. The technical contribution is twofold: 1) a probabilistic residual convolutional network architecture incorporating Gaussian distribution modeling is developed to quantify prediction uncertainty through RUL probability distributions and 2) a multiagent reinforcement learning (MARL) paradigm is formulated to dynamically optimize maintenance scheduling and spare parts inventory control based on real-time prognostic updates and system status. Through the dual-agent mechanism of MARL, maintenance decisions and spare parts ordering are dynamically adjusted based on real-time RUL distributions, balancing cost efficiency and reliability. This empirical validation quantitatively establishes the operational advantages of synergizing uncertainty-aware prognostics with adaptive resource allocation in PHM systems.