Aiming at the problems of traditional centralized cloud computing occupies large computing resources and high latency, this paper proposes a fault detection scheme for insulators self-explosion based on edge computing and DL (deep learning). In order to solve the high amount of computation brought by the deep neural network and meet the limited computing resources at the edge, a lightweight SSD (Single Shot MultiBox Detector) target recognition network is designed at the edge, which adopts MobileNets network to replace VGG16 network in the original model to reduce redundant computing. In the cloud, three detection algorithms (Faster-RCNN, Retinanet, YOLOv3) with obvious difference in detection performance are selected to obtain the coordinates and confidence of insulator self-explosion area, and then the self-explosion fault detection of overhead transmission line is realized by a novel multi-model fusion algorithm. The experimental results show that the proposed scheme can effectively reduce the amount of uploaded data, and the average recognition accuracy of cloud is 95.75%. In addition, it only increases the power consumption of edge devices by about 25.6W/h in the working state. Compared with the existing online monitoring technology of insulator self-explosion at home and abroad, the proposed scheme has the advantages of low transmission delay, low communication cost and high diagnostic accuracy, which provides a new idea for online monitoring research of power internet of things equipment.