Hybrid-YOLO for classification of insulators defects in transmission lines based on UAV

Abstract

Transmission power lines are essential to supply electrical energy to consumption centers. Keeping a reliable transmission system requires the early identification of faults. Image-based inspection of transmission lines makes fault identification faster and more accessible since it can be carried out using unmanned aerial vehicles (UAVs) in hard-to-reach places. In this paper, it is proposed to use a hybrid version of the You Only Look Once (YOLO) using ResNet-18 classifier, for power system inspection based on real images of failed components recorded by UAVs. This work assumed a dataset including 1,593 power grid inspection pictures for a supervised training. Based on YOLOv5x, with an mAP of 0.99262, the proposed method was superior to YOLOv5n, YOLOv5s, YOLOv5m, and YOLOv5l for object detection. For the multiclassification task, with an F1_score result of 0.96216, the proposed Hybrid-YOLO was superior to distinct architectures as the VGG-11, VGG-13, VGG-16, VGG-19, ResNet-34, ResNet-50, ResNet-152, DenseNet-121, DenseNet-161, DenseNet-169, DenseNet-201, YOLOv5, YOLOv6, and YOLOv7 versions.