Railway Track Monitoring Based on Multi-Sensor Spatio-Temporal Attention Fusion and Reinforcement Learning Intelligent Detection System

Abstract

To address the issues of unstable perception and decision-making delays in railway track inspection under complex scenarios such as rainy or foggy weather and switch areas, this paper proposes a multi-sensor spatiotemporal attention fusion and hierarchical decision-making system. The system integratesLight Detection and Ranging (LiDAR), visual cameras, and millimeter-wave radar, enhancing the robustness of multi-source data fusion through spatial and temporal attention mechanisms. At the decision-making level, a hierarchical architecture combining high-level planning and low-level control is adopted. The high-level layer performs global path planning and decision-making for critical areas based on reinforcement learning, while the low-level layer achieves precise trajectory tracking and safe action output through model predictive control. Experimental results show that under dense fog conditions (visibility < 100 m), the track gauge detection Mean Absolute Error (MAE) is 3.8 mm, the horizontal alignment MAE is 4.2 mm, the obstacle detection mean Average Precision (mAP) reaches 0.58, the turnout crossing success rate is 98%, the end-to-end decision latency is 85 ms, and the overall defect detection rate exceeds 95%. This work provides a reliable engineering solution for intelligent railway inspection, which can well support the development of autonomous maintenance robotics in the railway field.