基于Retinex-Net和改进YOLOv5网络的超高压变电站巡检图像故障识别  

Fault identification of ultra-high voltage substation inspection images based on Retinex-Net and improved YOLOv5 network

作　　者：杨利民 李松 钟明才 潘盈君 曹吉 YANG Limin;LI Song;ZHONG Mingcai;PAN Yingjun;CAO Ji(State Grid Xinjiang Electric Power Co.,Ltd.,UHV Branch,Urumqi 830001,China;Huayan Intelligent Technology(Group)Co.,Ltd.,Chengdu 610041,China)

机构地区：[1]国网新疆电力有限公司超高压分公司,乌鲁木齐830001 [2]华雁智能科技(集团)股份有限公司,成都610041

出　　处：《国外电子测量技术》2025年第1期126-133,共8页Foreign Electronic Measurement Technology

基　　金：国网新疆电力有限公司科技项目(5230CD240009)。

摘　　要：为进一步提高对超高压变电站设备图像故障识别的精度和速度,提出一种基于VGG(Visual Geometry Group)网络+Retinex-Net和改进YOLOv5网络的图像故障识别方法。首先利用归一化层的改进VGG网络判别巡检图像中的低照度图像和正常照度图像;然后采用Retinex-Net网络对低照度图像进行增强;最后采用损失函数改进、卷积改进和引入卷积注意力机制模块(Convolutional Block Attention Module,CBAM)注意力机制对YOLOv5网络进行改进,对增强后的低照度图像和正常图像进行故障识别。结果表明,所提方法可对超高压变电站设备的图像故障进行识别,且不受低照度图像的影响,具有较高的故障识别性能,其中识别精确率、召回率和平均精度均达90%以上,检测速度达30帧/s以上,网络参数量小于15M,浮点运算次数小于20。由此得出,所提方法可提高超高压变电站设备图像故障识别的精度和速度。In order to improve the accuracy and speed of equipment fault recognition in ultra-high voltage substations,a fault recognition method was proposed based on intelligent inspection images of ultra-high voltage substations.The method first uses an improved VGG(Visual Geometry Group)network with a normalization layer to distinguish between low illumination images and normal illumination images in the intelligent inspection of ultra-high voltage substations.Then,Retinex Net network was used to enhance the low illumination images.Finally,loss function improvement,convolution improvement,and improved YOLOv5 network with CBAM(Convolutional Block Attention Module)attention mechanism were used to classify the enhanced low illumination images and normal images,achieving equipment fault recognition in ultra-high voltage substations.The simulation results show that the proposed method can identify equipment faults in ultra-high voltage substations using intelligent inspection images,and is not affected by low illumination of intelligent inspection images.It has excellent fault recognition performance,with recognition accuracy,recall rate,and average accuracy reaching over 90%,detection speed reaching over 30 frames/s,parameter quantity less than 15 M,and floating-point operation frequency less than 20.From this,it can be concluded that the proposed method can improve the accuracy and speed of image fault recognition for ultra-high voltage substation equipment.

关 键 词：超高压变电站 智能巡检图像 故障识别 低照度图像增强 YOLOv5网络 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]