Research on edge intelligent recognition method oriented to transmission line insulator fault detection

Abstract

Aiming at the problem of high delay in the existing centralized cloud insulator fault detection methods, an insulator self-explosion fault detection scheme integrating mobile edge computing(MEC) and deep learning is proposed in this paper. In order to run the deep neural network on the edge devices with limited computing resources, the lightweight network MobilieNetv3 is replaced by the backbone network CSPDarknet53 of YOLOv4, which can effectively reduce the network parameters. The model detection speed and generalization ability are further improved by improving the activation function of MobilieNetv3 network and the loss function of YOLOv4. In addition, in order to meet the real-time requirements of insulator fault detection and maximize the utilization of UAV (Unmanned Aerial Vehicle) computing resources, a deep neural network partition algorithm based on binary particle swarm optimization is proposed. The algorithm realizes the division of the optimal partition points of deep neural network under the constraints of UAV energy consumption and system delay. The experimental results show that the edge intelligent recognition method can realize the terminal level recognition of insulator self explosion. The detection accuracy of the proposed object detection algorithm can reach 94.5% and the detection speed is 58.5 frames/s. At the same time, the deep neural network partition algorithm proposed in this paper has a significant effect on reducing UAV energy consumption, system cost and network delay.