Deep Reinforcement Learning for Predictive Maintenance in Dynamic Environments: From Continuous Prognostics to Robust Scheduling

Intelligent maintenance technology that can improve the automation and reliability of manufacturing systems has attracted much attention in Industrial Internet of Things. However, in real-world industrial scenarios, complex operating conditions (OC) and environmental changes lead to continuous changes in the distribution of operating data, which poses a significant challenge to data-driven predictive maintenance (PdM). This paper proposes an adaptive end-to-end PdM framework for continuous remaining useful life (RUL) prognostics and maintenance scheduling, where OC information is encoded as degradation features and maintenance states to enhance robustness to dynamic environments. First, a conformal Bayesian network is designed to predict the probabilistic RUL, improving the coverage of uncertainty interval. Afterwards, an uncertainty-aware orthogonal gradient projection modification is presented for continuous prognostics; parameter plasticity is defined using Bayesian posterior variance to impair the projection strength of elastic synapses. Eventually, a deep reinforcement learning-based maintenance environment is constructed using the states of RUL distribution and OC information, whose agent optimizes maintenance strategy via a distributional soft actor-critic algorithm. Continuous prognostic and maintenance experiments are conducted on aero-engine and bearing datasets under dynamic OC. The results demonstrate that the proposed framework can progressively enhance the accuracy of prognostics and the robustness of maintenance scheduling, presenting a promising PdM solution in dynamic environments.