Aiming at the problem of serious interference and low detection accuracy of insulator pictures collected in UAV patrol inspection, the optimization is carried out based on YOLOv5s algorithm, and the simulation research of insulator fault detection is carried out based on the improved YOLOv5s algorithm. The original algorithm is improved by adding CBAM attention module to the neck network, using k-means clustering to recalculate the size of a priori frame, and using MetaAconC as the activation function. The experimental results are analyzed based on python. The experimental results show that the advantage of this scheme is that the average accuracy of the algorithm mAP reaches 96.7%, which is 3.3% higher than the original YOLOv5s model; In addition, the weight file size of the algorithm training in this scheme is only 15.1M, which is only 0.1M larger than the original YOLOv5s, and still maintains the lightweight feature, which has a good prospect in the deployment of intelligent patrol work.