Research on bearing fault diagnosis and predictive maintenance algorithm of track moving equipment

This paper proposes an intelligent diagnostic and predictive maintenance algorithm that integrates multi-sensor features and deep reinforcement learning to address the issues of insufficient fault diagnosis accuracy and large prediction errors of remaining useful life (RUL) for bearings in rail mobile equipment under complex working conditions. First, a multimodal feature fusion framework with attention mechanism is designed for adaptive key time frequency domain features extraction of vibration, sound and temperature signals, achieving accurate recognition of cross condition fault patterns with an accuracy rate of 98.7%. Secondly, a joint model of LSTM-DQN (Long Short Term Memory Network-Deep Q Network) is designed to capture the degradation trend of bearings using LSTM. Combined with DQN dynamic optimization maintenance strategy, the RUL prediction error is significantly reduced (RMSE=520 km, MAPE=12.1%). The experiment is based on the full life cycle data of subway axle box bearings, and the results show that the proposed method reduces maintenance costs by 40% compared to traditional strategies while the unplanned shutdown event rate is less than 1%, providing a feasible solution for intelligent operation and maintenance of track equipment.