基于深度学习的架空输电导线缺陷检测方法研究  

Defect Detection of Overhead Transmission Wires Based on Deep Learning

作　　者：翟学明[1] 李晓 翟羽佳 ZHAI Xueming;LI Xiao;ZHAI Yujia(College of Control and Computer Engineering,North China Electric Power University,Baoding 071003,Hebei Province,China;State Grid Shijiazhuang Power Supply Company,Shijiazhuang 050000,Hebei Province,China)

机构地区：[1]华北电力大学控制与计算机工程学院,河北省保定市071003 [2]国网石家庄供电公司,河北省石家庄市050000

出　　处：《电网技术》2023年第3期1022-1030,共9页Power System Technology

摘　　要：无人机巡检图像中,架空输电导线断股、表面磨损等缺陷存在人工复检效率低、误检漏检率高的问题,为此提出了一种基于深度学习的架空输电导线缺陷智能检测方法。该方法以Unet为基础网络,结合迁移学习的思想,将VGG16(visual geometry group,16 weight layers)作为主干特征提取网络,并且将VGG16在ImageNet数据集上训练的权重作为预训练权重,以增强训练效果;然后将网络中的普通卷积用深度可分离卷积代替,有效地减少了网络的参数量;最后引入轻量级的高效通道注意力模块(efficient channel attention,ECA),实现不降维的局部跨信道交互策略,突出重要特征的同时克服了性能和复杂性之间的矛盾。在自建的输电导线缺陷数据集上,对方法进行了功能与性能测试,实验结果表明所提方法在导线断股检测上的准确率达到89.81%,在表面擦痕检测上的准确率达到90.86%,在表面刮损检测上的准确率达到93.58%,平均交并比(mean intersection over union,MIoU)值为86.12%,单张检测速度相对于Unet网络提升了8倍左右,提高了网络检测速度和检测精度。Aiming at the problems of low efficiency of the manual reinspection and high rate of the misdetection and undetection errors in checking the broken strands or the surface abrasion of the overhead transmission conductors in the UAV inspection images, an intelligent detection for the overhead transmission wire defects based on the deep learning is proposed. Taking the Unet as the base network and combined with the idea of migration learning, the VGG16 is used as the backbone feature extraction network. The weights of the VGG16 trained on the ImageNet dataset are firstly adopted as the pre-training weights to enhance the training effect;Then,the depthwise separable convolution(DS) is taken to replace the ordinary convolution, which effectively reduces the amount of the parameters in the network;Finally, a lightweight efficient channel attention(ECA) module is introduced to achieve a local cross-channel interaction strategy without doing dimensionality reduction, highlighting the important features while overcoming the contradiction between the performance and the complexity. The model is tested for functionality and performance on a self-built transmission conductor defect dataset. Experimental results show that the accuracy of this paper’s model reaches 89.81% for the conductor strand break detection, 90.86% for the surface scrape detection, 93.58% for the surface scratch detection, anf 86.12% the MIoU value. The detection speed of a single sheet is about 8 times higher than that of the Unet network, which effectively improves the model detection speed and detection accuracy.

关 键 词：无人机巡检图像 输电导线缺陷检测 迁移学习 深度可分离卷积 高效通道注意力 

分 类 号：TM721[电气工程—电力系统及自动化]