CSD-YOLOv8的输电线路故障目标检测  

Transmission line fault target detection of CSD-YOLOv8

作　　者：马旭[1] 王锐 邓军 常驰 郝帅[1] 李添麒 刘峥岐 李国亮 赵晴 MA Xu;WANG Rui;DENG Jun;CHANG Chi;HAO Shuai;LI Tianqi;LIU Zhengqi;LI Guoliang;ZHAO Qing(College of Electrical and Control Engineering,Xi’an University of Science and Technology,Xi’an 710054,China;College of Safety Science and Engineering,Xi’an University of Science and Technology,Xi’an 710054,China;Shaanxi Power Project Management Co,Ltd.,Xi’an 710075,China)

机构地区：[1]西安科技大学电气与控制工程学院,陕西西安710054 [2]西安科技大学安全科学与工程学院,陕西西安710054 [3]陕西电力项目管理有限公司,陕西西安710075

出　　处：《西安科技大学学报》2025年第2期383-392,共10页Journal of Xi’an University of Science and Technology

基　　金：国家自然科学基金项目(51804250);中国博士后科学基金项目(2020M683522);陕西省自然科学基础研究计划项目(2024JC-YBMS-490)。

摘　　要：针对无人机巡检输电线路过程中待检测目标受复杂背景干扰、故障目标部分遮挡以及目标多尺度造成传统算法难以准确检测的问题,提出一种基于CSD-YOLOv8的输电线路故障目标检测方法。首先,以YOLOv8网络作为基础框架,并在其主干网络中引入空间金字塔池化将不同尺度特征进行融合;然后,在检测网络头部中引入深度可分离卷积,并将其与交叉卷积连接模块结合,实现对部分遮挡目标的准确检测;此外,设计基于通道注意力机制的特征融合模块对不同层级特征进行加权融合,提高复杂背景下故障目标特征信息提取能力;最后,利用某电力巡检部门近5年的巡检数据对所提出算法进行验证。结果表明:相比于4种经典对比算法,所提方法在对12种故障类型检测效果的综合指标最好,平均检测精度为94.7%,召回率为93.0%。与此同时,所提算法具有较好的实时性,对于分辨率为1280×720的图像检测速度为45帧/s,为输电线路的智能巡检奠定了坚实的理论基础。To address the issues of complex background interference,partial occlusion of faulty targets,and difficulty in accurate detection caused by multi-scale targets in the process of unmanned aerial vehicle inspection of transmission lines,a transmission line fault target detection method based on CSD-YOLOv8 was proposed.First,the YOLOv8 network was used as the basic framework,and spatial pyramid pooling was introduced into the backbone network to fuse features of different scales;then,depth-separable convolution is introduced into the head of the detection network and combined with the cross-convolutional linkage module to achieve accurate detection of partially occluded targets;In addition,a feature fusion module based on the channel attention mechanism was designed to weight the features of different levels to improve the capability of fault target feature information extraction under complex background;finally,the inspection data of an electric power inspection department in the past 5 years were utilized to validate the proposed algorithm.The results show that:Compared with the four classical comparison algorithms,the proposed method has the best comprehensive index in detecting 12 kinds of fault types,the average detection accuracy is 94.7%,and the recall rate is 93.0%.At the same time,the proposed algorithm has good real-time performance,and the detection speed is 45 frames/s for the image resolution of 1280×720,which lays a solid theoretical foundation for the intelligent inspection of transmission lines.

关 键 词：YOLOv8 多尺度检测 通道注意力机制 特征融合 深度可分离模块 

分 类 号：TM755[电气工程—电力系统及自动化]