Detection of Defects on Cut-Out Switches in High-Resolution Images Based on YOLOv5 Algorithm

The reliability of a cut-out switch (COS) directly affects the stable operation of electric power distribution systems. Detecting a defective COS plays a critical role in protecting the power distribution line transformer. Currently, the conditions of these devices are captured and monitored using high-resolution cameras, but the human visual interpretation is still required. This study presents a method of detecting four COS defects based on You Only Look Once (YOLO) v5. Its default feature network structure is modified to enhance high-resolution images' minimal feature extraction ability. We have improved the loss function to ensure challenging cases can draw more attention while training. The proposed approach based on YOLOv5 is reliable and accurate for defect detection in high-resolution and imbalanced data. The final COS defect-detection accuracy has been calculated as 81.1% mAP@.5, improving the baseline performance by 4.6%. The accuracy of crack and arc defects has been improved by 10.8% and 5.5%, respectively.