Reinforcement learning-driven maintenance strategy: A novel solution for long-term aircraft maintenance decision optimization

A novel Reinforcement Learning (RL) driven maintenance strategy is proposed in this paper for solving the problem of aircraft long-term maintenance decision optimization. Specifically, it is targeted to process the information of aircraft future mission requirement, repair cost, spare components storage and aircraft Prognostics and Health Management (PHM) output, and provide real-time End-to-End sequential maintenance action decisions based on the coordination between short and long-term operation performance. The proposed RL-driven strategy is designed in the RL framework with Extreme Learning Machine based Q-learning algorithm, and an integrated aircraft maintenance simulation model is developed for training/testing RL-driven strategy. We test the proposed RL-driven strategy in several simulated dynamic aircraft maintenance scenarios together with 3 other commonly used maintenance strategies. The obtained results demonstrate that RL-driven strategy has prior performance in adjusting its decision principle for handling the variations of mission reward, repair/spare component storage cost and PHM ability in different maintenance scenarios. Some practical application suggestions and future perspectives of RL-driven strategy are discussed based on the obtained experiment results.