Multi-agent Deep Q-Learning for Maintenance Scheduling of Engineering System with Large-Scale State Space

In recent years, deep reinforcement learning has rapidly advanced and found widespread application in dynamic decision-making for industrial applications. In the field of industrial maintenance, devising repair plans for components coupled in complex subsystems is a challenging problem. This paper addresses this issue by constructing a multi-agent reinforcement learning framework suitable for multi-agent policy formulation environments, based on Deep Q-learning and utilizing a normalized representation derived from the state matrices of multi-body systems. To assess the effectiveness of our proposed algorithm, we created a mining transport case study comprising 5 subsystems and 14 subcomponents to test our algorithm. We compared it against the commonly used Actor-Critic reinforcement learning framework for multi-agent scenarios. Our results demonstrate superior convergence speed and validation performance in the testing environment compared to the contrastive algorithm. When compared to traditional multi-agent deep reinforcement learning approaches, the proposed method exhibits better performance in exploring global optimality.