基于改进Cascade R-CNN的绝缘子故障检测方法研究  

Study of insulator fault detection algorithm based on improved Cascade R-CNN network

作　　者：居来提·阿不力孜 刘玉龙 曹留 黄杰 张勇[2] 朱彦卿[2] ABULIZI Julaiti;LIU Yulong;CAO Liu;HUANG Jie;ZHANG Yong;ZHU Yanqing(Urumqi Power Supply Company of State Grid Xinjiang Electric Power Co.,Ltd.,Urumqi 830000,China;College of Electrical and Information Engineering,Hunan University,Changsha 410082,China)

机构地区：[1]国网新疆电力有限公司乌鲁木齐供电公司,新疆乌鲁木齐830000 [2]湖南大学电气与信息工程学院,湖南长沙410082

出　　处：《电力科学与技术学报》2023年第3期140-148,共9页Journal of Electric Power Science And Technology

基　　金：国家电网有限公司科技项目(SGXJWL00SDJS2100845)。

摘　　要：针对航拍图中存在的绝缘子故障位置在图像中占比小、背景环境复杂导致的故障检测准确率低的问题,提出一种基于改进Cascade R-CNN模型的绝缘子故障检测方法。在原有Cascade R-CNN模型的基础上,在骨干网络中引入可变形卷积学习几何变换能力,在检测器中引入平衡损失函数平衡难易样本。在模型训练阶段,使用Copy-Paste与Mosica丰富故障绝缘子样本,平衡正负样本。使用该模型对航拍绝缘子图片进行故障检测实验,改进损失函数的模型与传统Cascade R-CNN模型相比平均召回率提升0.38%,引入可变卷积后的Cascade R-CNN模型与Faster R-CNN模型,相比平均召回率,从原来的89.78%变成93.49%,结果表明该模型能够有效克服样本遮挡以及样本不平衡的干扰。Aiming at the low accuracy problem of insulator fault detection caused by the fault position occupies a small proportion in the image and complex background environment in aerial images,an insulator fault detection method based on optimized Cascade R-CNN model is proposed in this paper.Based on the original Cascade R-CNN model,deformable convolution is inserted into the backbone network to learn geometric transformation capabilities,and balance loss function is introduced in the detector to balance difficult and easy samples.In the model training phase,the faulty insulator samples are enriched by using Copy-Paste and Mosica,and the positive and negative samples are balanced.The proposed model is tested for insulator fault detection.Compared with the traditional Cascade R-CNN model,the average recall of the optimized loss function model improves 0.38%.Comparing with the Faster R-CNN model,the average recall of the Cascade R-CNN model after introducing variable convolution improves from 89.78%to 93.49%.The results indicate that the proposed model can overcome the interference of samples shielding and sample imbalance effectively.

关 键 词：绝缘子 故障检测 改进Cascade R-CNN 可变形卷积 

分 类 号：TM855[电气工程—高电压与绝缘技术]