基于改进的YOLOv4绝缘子掉片故障检测方法  

Insulator Dropout Fault Detection Method Based on Improved YOLOv4

作　　者：党宏社[1] 薛萌 郭琴 DANG Hongshe;XUE Meng;GUO Qin(School of Electrical and Control Engineering,Shaanxi University of Science and Technology,Xi'an,710021,China;Xi'an Xirui Control Technology Co.,Ltd.,Xi'an,710021,China)

机构地区：[1]陕西科技大学电气与控制工程学院,西安710021 [2]西安西瑞控制技术股份有限公司,西安710021

出　　处：《电瓷避雷器》2022年第1期211-218,共8页Insulators and Surge Arresters

基　　金：陕西省自然科学基金资助项目(编号:2020JM-509)。

摘　　要：针对现有输电线路无人机巡检使用的目标检测算法速度较慢且模型文件较大的问题,提出了一种改进的YOLOv4目标检测算法。使用轻量型的MobileNetv2网络作为模型的主干特征提取网络,并将模型后续的标准卷积运算变为深度可分离卷积,减少了运算参数。使用K-means聚类得到了9种绝缘子及掉片故障锚框尺寸的先验知识。使用h-swish函数作为模型颈部网络的激活函数,减少了特征反复提取过程的信息损失。通过与主流算法进行实验对比,改进后的算法检测速度可达到10.51 FPS,相较SSD算法快了18.7%,模型文件大小为46.4 MB,仅为原算法的1/5,检测的平均精度mAP达到94.08%,在满足精度的情况下提升了检测速度,减少了模型体积,为实现无人机巡检的边采集边检测提供了可能。Aiming at the problem that the existing target detection algorithm used in UAV inspection on the transmission lines runs too slow and the model file is large,an improved YOLOv4 target detection algorithm is proposed.Lightweight MobileNetv2 network is used as the backbone network of the model,and depth separable convolution is adopted instead of the subsequent original convolution to reduce model parameters.K-means is applied to obtain the prior knowledge of 9 anchor frame sizes suitable for insulators and drop fault,and h-swish is adopted as the activation function of the model neck network,which reduces information loss in the repeated feature extraction process.By comparison with mainstream algorithms,the detection speed of improved algorithm can reach 10.51 FPS,which is 18.7%faster than SSD algorithm,and the model file size is 46.4 MB,only 1/5 of original algorithm,and then mean average precision reaches 94.08%,while meeting the accuracy,the detection speed is improved,and the model size is reduced,which makes it possible to realize the detection and the image acquisition of the UAV inspection.

关 键 词：绝缘子故障 目标检测 YOLOv4 MobileNetv2 深度学习 

分 类 号：TM75[电气工程—电力系统及自动化] TP391.41[自动化与计算机技术—计算机应用技术]