摘要
为了对高压输电线路的绝缘子串航拍图像进行精确快速的定位与状态识别,提出了一种基于改进YOLOv3的绝缘子串诊断方法。在自主建立航拍绝缘子串数据库的基础上,针对数据集样本存在的复杂度不均衡和类别不均衡现象,采用Focal Loss函数和均衡交叉熵函数改进YOLOv3算法的损失函数;然后,对原网络在COCO数据集上训练的卷积层过滤器进行可视化分析,选择冻结层并采用多阶段迁移学习策略来训练网络。在Python环境下训练并测试实例,结果表明:改进的损失函数可优化网络训练的损失值,提升算法精确度;多阶段迁移学习策略在提高算法精确度的同时,能有效应对数据集小而易过拟合问题;所提方法可端对端实现绝缘子串的定位与状态识别,且诊断精确度达到0.918。研究结果证明所提方法具有较高的精确性和实时性。
In order to realize accurate and rapid positioning and state recognition of an insulator aerial image of high-voltage transmission lines,an insulator string diagnosis method based on improved YOLOv3 algorithm is proposed.Based on the self-established aerial insulator substring database,the Focal Loss function and the Balanced Cross Entropy function are used to improve the loss function of the YOLOv3 algorithm to sovle the problems of complexity imbalance and class imbalance of the data set samples.The convolutional layer filters trained by the original network on the COCO data set are visually analyzed.The frozen layers are selected and the network is trained by the multi-stage transfer learning strategy.The results of training and testing in Python environment show that the improved loss function can optimize the loss value of network training and improve the accuracy of the algorithm;multi-stage transfer learning can effectively deal with the easy fitting problem of the small data set while improving the accuracy of the algorithm.The proposed method can realize the positioning and state recognition of the insulator string end-to-end,and the diagnostic accuracy reaches 0.918.The results fully verify that the accuracy and real-time of the algorithm are better.
作者
颜宏文
陈金鑫
YAN Hongwen;CHEN Jinxin(School of Computer&Communication Engineering,Changsha University of Science&Technology,Changsha 410114,China)
出处
《高电压技术》
EI
CAS
CSCD
北大核心
2020年第2期423-432,共10页
High Voltage Engineering
基金
国家自然科学基金(51407013).
关键词
绝缘子串
状态识别
YOLOv3
损失函数
迁移学习
高压输电线路
insulator string
state recognition
YOLOv3
loss function
transfer learning
high-voltage transmission lines
作者简介
通信作者:颜宏文,1968—,女,博士,教授,硕导,主要研究方向为智能电网中的数据挖掘,E-mail:yanhongwen7@126.com
