基于Mask R-CNN与改进BP神经网络联合算法的变压器套管红外热故障诊断  

Thermal Fault Diagnosis of the Bushing Infrared Images Based on Mask R-CNN and Improved BP Neural Network Joint Algorithm

作　　者：李雪寒 刘沁怡 杨晓彤 胡海敏 王哲铭 周文强 卢武 LI Xuehan;LIU Qinyi;YANG Xiaotong;HU Haimin;WANG Zheming;ZHOU Wenqiang;LU Wu(Shanghai University of Electric Power,Shanghai 200090,China;State Grid Shanghai Electric Power Company Pudong Power Supply Company,Shanghai 200122,China;State Grid Shanghai Electric Power Supply Company,Shanghai 200080,China)

机构地区：[1]上海电力大学,上海200090 [2]国网上海市电力公司浦东供电公司,上海200122 [3]国网上海市电力公司市区供电公司,上海200080

出　　处：《上海电力大学学报》2023年第6期591-598,共8页Journal of Shanghai University of Electric Power

基　　金：国家自然科学基金(51707113);上海市教委及教育发展基金会“晨光计划”人才培养计划(21CGA63)。

摘　　要：为解决传统图像类算法在变压器套管状态诊断时存在的效率低、准确度不高以及复杂背景下变电设备目标识别困难等问题,提出了将Mask R-CNN与改进BP神经网络相结合的套管红外图像状态诊断方法。首先,利用Mask R-CNN解决套管红外图像背景复杂时分割困难的问题;其次,基于灰度特征的特征量提取方案,实现对红外伪彩图特征量的提取;最后,引入粒子群优化BP神经网络(PSO BP)算法对变压器套管特征进行分类识别。实验结果表明,该方法对红外图像中套管的运行状态具有较好的检测效果,对套管中介质损耗故障、接头故障和漏油故障的故障诊断准确率分别可达100.0%、88.9%和96.3%,平均准确率达到93.518%,优于传统BP算法和支撑向量机(SVM)算法。To solve the problems of low efficiency,low accuracy and difficult target recognition of substation equipment under complex backgrounds in the traditional image algorithm in the state diagnosis of transformer bushings,this paper proposes a combination of diagnosis method of tube infrared image state with Mask R-CNN and improved BP neural network.First,Mask R-CNN is used to solve the problem of difficult segmentation when the background of the casing infrared image is complex;secondly,the feature extraction scheme based on grey features realizes the extraction of infrared pseudo-color image features;finally,the particle swarm optimization BP(PSO BP)in neural network algorithm is introduced to classify and identify the transformer bushing features.The results show that this method has a good detection effect on the running state of the casing in the infrared image,and the fault diagnosis accuracy of the dielectric loss fault,joint fault and oil leakage fault in the casing can reach 100%,88.9%and 96.3%respectively,the average accuracy reaches 93.518%,which is superior to the traditional BP algorithm and SVM algorithm.

关 键 词：变压器绝缘套管 红外图像 Mask R-CNN 改进BP神经网络 状态诊断 

分 类 号：TM83[电气工程—高电压与绝缘技术] TP183[自动化与计算机技术—控制理论与控制工程]