Reinforcement Learning-Based Predictive Maintenance for Degrading Multi-Axis Synchronization Systems

Performance degradation of actuators in complex multi-axis synchronization systems seriously threatens system reliability and operational safety. Traditional time-based maintenance cannot handle nonlinear degradation of multiple components, often resulting in synchronization failures. This paper presents a new predictive maintenance (PdM) framework combining statistical state estimation and reinforcement learning.First, Kalman filter, EM algorithm, and RTS smoother are used to accurately estimate actuator health states from noisy measurements. Then, a Mahalanobis distance indicator weighted by the Entropy Weight Method is applied to fuse multi-dimensional data into a unified Health Degree (HD). Finally, a Q-learning agent is designed to optimize maintenance decisions, balancing synchronization accuracy and maintenance cost.Simulation results based on industrial stage machinery show that the proposed method effectively extends the remaining useful life (RUL) and ensures high synchronization reliability under various degradation conditions.