Fault Detection and Identification of Multi-Source Insulators Based on Improved YOLOv7

To improve the accuracy of fault detection for heterogeneous insulators with complex backgrounds, an improved YOLOv7 model-based insulator fault recognition method based on heterogeneous images is proposed in this paper. An image of a heterogeneous insulator was registered and fused to highlight the location of the insulators and fault information. Then, to reduce the computational complexity and improve portability, the original backbone feature extraction network of YOLOv7 was replaced by the MOBELINET network. To reduce the problems of missing and false detections of insulators in complex backgrounds, the original loss function of YOLOv7 was changed from CIOU to FOICAL-EIOU to further improve the regression effect of the model prediction box. Finally, a deformable convolutional DCNv2 was introduced in the YOLOv7 detection head to enhance its adaptability to insulator heating fault areas of different scales. The experimental results show that the mAP value of the improved model is 96.6%, which is 9.9% higher than that of the original YOLOv7 model, the number of parameters decreases by 30.5%, and the floating-point operator decreases by 49.2%, which are 12.2% and 12.4% higher than those of the YOLOV5 and YOLOV8 target detection models, respectively. The proposed improved model could effectively detect and identify heterogeneous insulators.