基于YOLOv7的均压环缺陷及倾斜角度检测方法  

Detection Method for Defects and Tilt Angle of Grading Ring Based on YOLOv7

作　　者：黄泽泽 董晓杰 李少龄 付建文 康泰安 戚银城[1,3] HUANG Zeze;DONG Xiaojie;LI Shaoling;FU Jianwen;KANG Taian;QI Yincheng(Department of Electronic and Communication Engineering,North China Electric Power University,Baoding 071003,Hebei Province,China;Dongfang Electronics Co.,Ltd.,Yantai 264000,Shandong Province,China;Hebei Key Laboratory of Power Internet of Things Technology,North China Electric Power University,Baoding 071003,Hebei Province,China)

机构地区：[1]华北电力大学电子与通信工程系,河北省保定市071003 [2]东方电子股份有限公司,山东省烟台市264000 [3]华北电力大学河北省电力物联网技术重点实验室,河北省保定市071003

出　　处：《电力信息与通信技术》2025年第4期25-34,共10页Electric Power Information and Communication Technology

基　　金：河北省省级科技计划资助(SZX2020034)。

摘　　要：针对输电线路均压环缺陷及倾斜角度难以检测的问题,文章提出一种改进的YOLOv7均压环缺陷及旋转目标检测模型。将YOLOv7的主干网络设计为双分支,使CNN的局部特征与Transformer的全局表示相结合以增强目标特征的表达能力。采用渐进式特征融合结构,使得模型更有效地将浅层特征和深层特征进行融合。在检测头中引入了高斯建模表示的旋转框检测,提高了角度检测的精准性。最后基于改进后的旋转目标检测模型,设计了一种均压环倾斜角度检测算法,实现了任意视角下均压环倾斜角度的检测。实验结果表明,改进后的算法平均角度误差降低了0.97°,仅为5.32°,mAP提高了4.1%,达到91.5%,算法显著提高了均压环缺陷的检测效果,为均压环缺陷及倾斜角度的定量检测提供了一个有效的解决方案。Aiming at the problem that it is difficult to detect the defects and tilt angle of the grading ring of transmission lines,an improved YOLOv7 grading ring defect and rotated object detection model is proposed.The backbone network of YOLOv7 is designed as a two-branch network,which combines the local features of CNN and the global representation of Transformer to enhance the expression ability of object features.A progressive feature fusion structure is used to make the model more effective in fusing shallow and deep features.A Gaussian modeling representation of rotated box detection is introduced in the detection head to improve the accuracy of angle detection.Finally,based on the improved rotated object detection model,a grading ring tilt angle detection algorithm is designed to realize the detection of the grading ring tilt angle under any viewing angle.The experimental results show that the average angular error of the improved algorithm is reduced by 0.97°to only 5.32°,and the mAP increased by 4.1%to 91.5%.The algorithm significantly improves the mean average precision in the defect detection of grading ring,and provides an effective solution for the quantitative detection of grading ring defects and tilt angle.

关 键 词：均压环 深度学习 倾斜角度检测 缺陷检测 特征融合 

分 类 号：TM75[电气工程—电力系统及自动化] TP18[自动化与计算机技术—控制理论与控制工程]