A Power Distribution Equipment Corrosion Cascade Detection Method and the Edge Deployment Based on Jetson Nano Platform

Corrosion detection of distribution equipment is of great significance to ensure the safe and stable operation of power grid systems. Since the corroded area typically exhibits irregular shapes, blurred edges, and significant scale differences, the traditional detection method based on a fixed convolution kernel has problems with missed and false detection. This paper aims to realize corrosion detection and its edge deployment based on Jetson Nano B01, and proposes a cascade corrosion detection method that integrates improved deformable convolution and Transformer classification modules. This method introduces the dilated DCNv3 (D-DCNv3) convolution module, into the YOLOv7 model to enhance the modeling ability of irregular targets. At the same time, Swin Transformer V2 is introduced as a subsequent fine classifier to reclassify the detection area and further improve overall detection accuracy. Experimental results show that the proposed method achieves an AP50 index of 69.0%, which is 3.1% higher than the baseline, and the classification accuracy reaches 99.5%. In addition, through INT8 quantization and model pruning, lightweight deployment is achieved on the Jetson Nano B01 platform. The inference speed reaches 18 FPS at a resolution of 320 × 320 pixel, and AP50 remains at 67.8%, with only a 1.2% loss in accuracy compared to the baseline model, meeting the distribution station inspection delay requirement of no more than 50 ms. During the trial operation of the proposed system in a 220 kV distribution station, the system has achieved a 20-fold increase in efficiency compared to manual inspection, and power consumption is controlled at 8.5 W, providing a low-cost solution for unmanned operation and maintenance.