基于深度学习的红外图像中劣化绝缘子片的分割方法

刘云鹏; 张喆; 裴少通; 武建华; 梁利辉; 马子儒

刘云鹏,张喆,裴少通,武建华,梁利辉,马子儒.基于深度学习的红外图像中劣化绝缘子片的分割方法[J].电测与仪表,2022,59(9):63-68.
Liu Yunpeng,Zhang Zhe,Pei Shaotong,Wu Jianhua,Liang Lihui,Ma Ziru.Faulty insulator segmentation method in infrared image based on deep learning[J].Electrical Measurement & Instrumentation,2022,59(9):63-68.

基于深度学习的红外图像中劣化绝缘子片的分割方法

Faulty insulator segmentation method in infrared image based on deep learning

DOI：10.19753/j.issn1001-1390.2002.09.009

中文关键词: 红外成像劣化绝缘子深度学习图像语义分割全卷积神经网络

英文关键词:Infrared imaging, Faulty insulators, Deep learning, Image semantic segmentation, Full convolutional neural network

基金项目:国家重点研发计划(2018YFF01011900)

作者中文名	作者英文名	单位
刘云鹏	Liu Yunpeng	华北电力大学电力工程系
张喆	Zhang Zhe	华北电力大学电力工程系
裴少通	Pei Shaotong	华北电力大学河北省输变电设备安全防御重点实验室
武建华	Wu Jianhua	国家电网河北省电力有限公司检修分公司
梁利辉	Liang Lihui	国家电网河北省电力有限公司检修分公司
马子儒	Ma Ziru	华北电力大学电气工程系

摘要点击次数: 1543

中文摘要:

红外成像作技术为绝缘子在线检测的重要手段已日趋成熟,而当前对于电气设备红外图像的分析仍依靠大多依赖人工经验,缺乏智能化。传统的红外图像中劣化绝缘子的分割方法需要复杂的图像预处理和人工提取目标特征,因而存在着泛化能力差、复杂背景下识别精度低等缺陷。基于以上问题,文中提出一种基于深度学习的红外图像中劣化绝缘子的分割方法。首先搭建实验平台,获得绝缘子串在复杂背景(有其他设备和发热源干扰)下的红外图像。为了保证红外图像的多样性,拍摄在多种污秽等级下进行,选取了多种阻值的劣化绝缘子片放置在了多种位置,并不断改变了拍摄的角度。然后构建多尺度的方法实现全卷积神经网络(FCN)的3种子模型架构：FCN-32s、FCN-16s和FCN-8s,使用随机梯度下降的训练方法对模型进行端到端的训练,最终实现了红外图像中的劣化绝缘子片的自主分割提取。通过对三种子模型训练测试对比后,得出以下结果：FCN-8s模型为劣化绝缘子片分割提取最佳模型,对于验证测试数据集像素分类准确率为89.23%。结果表明文中所描述的智能化红外劣化绝缘子片分割方法,为绝缘子和其他高压电气设备的红外在线监测诊断打下了坚实基础。

英文摘要:

As an important means of on-line detection of insulators, infrared imaging technology has matured. But the current analysis of infrared images still mostly depends on manual experience, which is lack of intelligence. The traditional methods of faulty insulator segmentation in infrared images requires complex image preprocessing and manual extraction of target features,which leads to the disadvantages of poor generalization capability and low recognition accuracy in complex contexts. Based on the above problems, this paper proposes a method of faulty insulator segmentation in infrared images based on deep learning. First, an experimental platform was set up to obtain infrared images of insulator strings under complex background (with other equipment and heat source interference). In order to ensure the diversity of infrared images, shooting was conducted in a variety of contamination levels, shooting angles, and backgrounds. At the same time, insulator pieces with various degrees of deterioration were selected and placed in various positions. Then build a multi-scale approach to achieve the three sub-models of Fully Convolutional Networks(FCN) architecture: FCN-32s, FCN-16s, and FCN-8s. The stochastic gradient descent method was applied to the training of this end-to-end model, and finally the autonomous division of faulty insulators in infrared images is achieved. After training and comparing the three sub-models, the following results are obtained: The FCN-8s model is the best model for faulty insulator segmentation, and the accuracy rate for verifying the test data set is 89.23%. The results show that the intelligent infrared faulty insulator segmentation method described in this paper has a good segmentation effect, and lays a solid foundation for intelligent infrared on-line monitoring and diagnosis of other high-voltage electrical equipment.

查看全文查看/发表评论下载PDF阅读器

关闭