Maintenance planning for a continuous monitoring system using deep reinforcement learning

This paper proposes a maintenance decision-making framework for multi-unit systems using Machine Learning (ML). Specifically, we propose to use Deep Reinforcement Learning (RL) for a dynamic maintenance model of a multi-unit parallel system that is subject to stochastic degradation and random failures. As each unit deteriorates independently in a three-state homogeneous Markov process, we consider each unit to be in one of three states: healthy, unhealthy, or a failed state. We model the interaction among system states based on the Birth/Birth-Death process. By combining individual component states, we define the overall system state. To minimize costs, we use the Markov Decision Process (MDP) framework to solve the optimal maintenance policy. We apply the Double Deep Q Networks (DDQN) algorithm to solve the problem, making the proposed RL solution more practical and effective in terms of time and cost savings than traditional MDP approaches. A numerical example is provided which demonstrates how the RL can be used to find the optimal maintenance policy for the system under study.