Substation equipment target recognition and path planning for power inspection robots

Abstract

Reliable autonomous inspection in high-voltage substations requires a robot to recognize small and safety-critical equipment targets while planning routes that remain outside energized danger regions. This paper proposes an optical and light detection and ranging (LiDAR) perception and planning method for substation inspection robots. The method combines visible images, infrared images, LiDAR bird's-eye-view maps, and inertial odometry in a calibrated edge pipeline. A lightweight recognition network extracts multi-scale visual and thermal features, introduces a small-object attention branch for meters, bushings, and warning plates, and sends detection uncertainty to a semantic risk grid. The planner integrates global graph search with local velocity optimization, enabling the robot to visit assigned inspection targets while maintaining clearance from live equipment and newly detected obstacles. A reproducible substation-like benchmark with paired images, LiDAR scans, and route tasks is built to evaluate recognition and navigation. The proposed model achieves 62.8% mean average precision averaged over intersection-over-union (IoU) thresholds from 0.50 to 0.95 (mAP50:95), 89.4% small-object recall, and 43.7 frames per second (FPS) on an edge graphics processor. For route planning, it reduces mean replanning time to 0.18 s, increases mean clearance to 2.14 m, and reaches a 96.7% mission success rate in dynamic tests. The results indicate that coupling optical target recognition with uncertainty-aware path planning can improve the safety, repeatability, and automation level of substation inspection robots.