Research on Adaptive Inspection Algorithm for Distribution Network Based on Multi-Source Fusion and Image Deep Learning Recognition

This paper addresses the challenges of low efficiency and poor safety in traditional manual inspections for distribution networks by proposing an adaptive inspection algorithm based on image deep learning recognition and multi-source data fusion. By integrating the advantages of visible light images (high resolution, rich features) and infrared images (strong anti-interference capability), combined with the You Only Look Once version 8(YOLOv8) deep learning framework, high-precision identification of distribution network equipment is achieved. Building on this, adaptive inspection algorithms for single-pole and continuous-pole are designed. For single-pole inspection mode, drones autonomously adjust spatial positions and gimbal angles at seven key inspection points through visual high-precision navigation, with inspection time optimized and reduced as sample data accumulates. For continuous-pole inspection mode, by optimizing waypoint transitions and flight paths, the average inspection time per tower is further reduced to 245 seconds, achieving over 50% efficiency improvement compared to traditional laser point cloud methods. Experimental results demonstrate that this method significantly outperforms manual inspections and laser point cloud technology in safety, timeliness, and cost-effectiveness. It autonomously generates continuous inspection routes, reduces manual intervention by over 90%, and provides an efficient, low-cost solution for intelligent operation and maintenance of distribution networks.