A framework for modeling and optimization of mechanical equipment considering maintenance cost and dynamic reliability via deep reinforcement learning

Effective maintenance of high-value mechanical equipment must balance cost efficiency with reliability, especially in safety-critical systems such as aerospace technologies, high-speed rail, and marine vessels. However, most existing studies prioritize cost reduction alone, which can inadvertently elevate failure risk in advanced systems. To tackle this challenge, this study proposes a novel deep reinforcement learning (DRL) framework that jointly optimizes dynamic maintenance costs and equipment reliability in real time. In the proposed approach, the equipment's degradation is modeled using a hybrid Gamma-Wiener process, capturing both progressive wear and random shocks for more accurate reliability estimation than traditional models. We further introduce an innovative reliability reconstruction method that reshapes the reliability metric during training, guiding the DRL agent to balance reliability and cost more effectively. Unlike conventional maintenance strategies with fixed thresholds or schedules, our DRL-based agent continuously learns and adapts maintenance decisions based on the equipment's state, eliminating the need for preset maintenance thresholds. Compared to prior reinforcement learning approaches that often optimize only cost or use simplistic degradation assumptions, this framework integrates a realistic reliability model and a multi-objective reward, resulting in more robust and safer decision-making. The framework is validated by the framework on CNC machine tool and aircraft engine case studies, where the learned policies significantly reduce maintenance costs while maintaining high reliability. The proposed method outperforms baseline strategies in cost savings and reliability trade-offs, demonstrating its superior effectiveness and adaptability for intelligent maintenance planning.