Beyond static schedules: Dynamic maintenance optimization with double deep reinforcement learning

This study presents an adaptive framework for dynamic preventive maintenance optimization based on the Double Deep Q-Network (DDQN) algorithm. The objective is to learn cost-optimal preventive maintenance policies under stochastic and partially observable failure behavior, relying solely on observed failure and maintenance events rather than condition-monitoring data or known degradation models. Equipment hazard function is modeled using non-homogeneous Poisson processes, including power-law and bathtub models, while maintenance actions follow variable restoration levels defined through the proportional age-reduction model. Training is performed on simulated failure trajectories using a standard workstation in under two hours, and the trained agent performs inference nearly instantaneously. Results demonstrate that the DDQN-based adaptive policy consistently outperforms analytical periodic and static benchmarks, as well as a dynamic genetic algorithm and a standard reinforcement learning implementation, by achieving lower average maintenance costs and reduced variability across a wide range of corrective-to-preventive cost ratios. The method also remains robust under perturbed and uncertain hazard conditions, maintaining stable performance without retraining. These findings highlight the potential of the proposed DDQN approach as a computationally efficient and generalizable tool for reliability-centered maintenance optimization, capable of adapting to stochastic cost structures and cumulative corrective effects while operating effectively in data-limited industrial environments.