A lightweight improvement of YOLOv5 for insulator fault detection

Abstract

Abstract Along with the development of artificial intelligence, mobile terminal equipment patrol inspection has become the mainstream of power grid line patrol inspection. Insulator defect detection is an important part of power patrol inspection. To increase the detection speed under the condition of guaranteeing high precision of insulator detecting, an improved lightweight YOLOv5 algorithm is presented to achieve insulator defect detection. This algorithm uses the lightweight Ghost convolution to improve the general convolution and the Ghost Bottleneck module to improve the head module in YOLOv5. Based on the original Ghost lightweight, the algorithm improves the channel data suitable for insulator detection and decreases the number of convolutions. In the same research data and experimental environment setting, the effect is better than the unmodified Ghost optimization. Experiment results indicate that the mean precision of detecting insulator is 81%, the number of algorithm models and parameters is reduced, the speed of detection is increased under the premise of ensuring accuracy, and the improved algorithm model is more lightweight and easy to deploy and use in embedded mobile terminals.