DT-PPO: A Real-Time multisensor-driven predictive maintenance framework

Prognostics and health management (PHM) encompasses both predictive maintenance and health monitoring efforts. However, most studies either focus on prognostics or decision-making in isolation, with only a few attempts to integrate residual life prediction with maintenance strategies. Yet the resulting solutions are often cumbersome and impractical for real-world application. To address this gap, this paper proposes a dynamic transition-proximal policy optimization (DT-PPO) network for predictive maintenance of mechanical equipment. The framework processes multisensory data containing degradation information at different scales, and accurately transforms the initial equipment states into belief states, enabling real-time maintenance decisions. First, a state transition network (TransStateNet) is built as a deep self-encoder to selectively extract multisensory features at different scales and output belief states. Second, a DT-PPO architecture based on belief states is designed to develop maintenance strategies, including spare parts ordering and downtime planning. Finally, an s-g policy is incorporated into the DT-PPO to guide action selection, while state sliding windows process multi-sensor sequence data to reduce uncertainty in the action space, enhancing the robustness of the maintenance strategy. Experimental comparisons against benchmark strategies validate the superiority of the proposed framework.