基于二阶段目标增强网络的低照度复杂环境下绝缘子故障检测方法  

Insulator Faults Detection in Low Illuminance Complex Environment Based TOE-Net

作　　者：田子建[1] 吴佳奇 张文琪 陈伟[1,2] 杨伟 王帅[1,3] TIAN Zijian;WU Jiaqi;ZHANG Wenqi;CHEN Wei;YANG Wei;WANG Shuai(School of Mechanical Electronic&Information Engineering,China University of Mining and Technology(Beijing),Haidian District,Beijing 100083,China;School of Computer Science&Technology,China University of Mining and Technology,Xuzhou 221116,Jiangsu Province,China;Inner Mongolia Bureau of the National Mine Safety Administration,Hohhot 010010,Inner Mongolia Autonomous Region,China)

机构地区：[1]中国矿业大学(北京)机电与信息工程学院,北京市海淀区100083 [2]中国矿业大学计算机科学与技术学院,江苏省徐州市221116 [3]国家矿山安全监察局内蒙古局,内蒙古自治区呼和浩特010010

出　　处：《电网技术》2024年第3期1331-1340,共10页Power System Technology

基　　金：国家自然科学基金项目(52074305,52274160,51874300);国家自然科学基金委员会–山西省人民政府煤基低碳联合基金(U1510115)。

摘　　要：从低照度户外环境中航拍采集的绝缘子影像,存在照度低、背景复杂、绝缘子故障目标小等缺陷,严重影响低照度环境下绝缘子故障检测准确性。为解决上述问题,文章提出一种基于TOE-Net的低照度复杂环境下绝缘子故障检测方法,提出TOE-Net进行图像预处理方法,再使用YOLOv7-OL作为故障检测模块检测小目标绝缘子故障。在二阶段目标增强网络(two-stage object enhancement network,TOE-Net)中,设计零目标图像增强损失函数实现预增强网络(preparation enhancement network,PreEnNet)和深度增强网络(deep enhancement network,DeepEnNet)的无监督学习;使用信道级注意力模块跳跃式通道注意力机制(skip squeeze excitation networt,Skip_SENet)和跳跃式通道注意力机制(skip convolutional block attention module,Skip_CBAM)模块改进原始小目标特征增强单次多框检测算法(small object detection enhancement single shot multiBox detector,SDE-SSD),从而提升定位网络的小目标检测能力;设计弱监督机制使预增强网络根据小目标特征增强SSD的要求来提升图像增强能力,直到小目标特征增强SSD能够从增强图像中准确定位绝缘子串位置;使用深度增强网络深度增强绝缘子串区域,提升各类故障的特征显著性。故障检测模块中,将YOLOv7目标检测算法改进为面向小目标YOLOv7,在原模型中添加结合多尺度特征自适应融合网络的小目标检测通道,并将原始损失函数的CIOU改进为BIOU,从而提高模型的小目标检测性能。在低照度环境绝缘子故障检测实验中,该算法与5种目前常用目标检测算法相比具有较大优势,并且相较于低光目标检测算法IA-YOLO、GenISP with RetinaNet,m AP提升9.77%、10.35%,检测速度提升7.23%、10.16%,证明该算法适用于低照度复杂环境下小目标绝缘子故障检测任务;在正常光照绝缘子故障检测实验中该算法仍保持出色性能,证明该算法能够实现常规光照条件下The insulator images collected from the aerial photography in a low illuminance outdoor environment have such defects as low illuminance,complex background,small insulator faulty objects,etc.,which seriously affect the accuracy of the small object insulator faults detection in a low illuminance environment.In order to solve the above problems,the paper proposed an insulator faults detection in a low illuminance complex environment based the TOE-Net.The TOE-Net is built for image preprocessing,and then the YOLOv7-OL is used as the faults detection module to detect the small object insulator faults.In the TOE-Net,a non-object image enhancement loss is designed to achieve the unsupervised training of the PreEnNet and the DeepEnNet;The channel-attention module Skip_SENet and Skip_CBAM module are used to improve the original SSD to enhance the small target detection capability of the location network;The weakly supervised mechanism is designed to improve the image enhancement capability of the PreEnNet according to the requirements of the SDE-SSD until the SDE-SSD can accurately locate the insulator position from the image enhanced by the PreEnNet;Finally,the DeepEnNet is used to enhance the insulator area in depth to improve the feature significance of various faults.In the faults detection module,the YOLOv7 is improved to the YOLOv7-OL.By adding a small target detection channel combining with the multi-scale feature adaptive fusion network to the original model, the CIOU of the original loss function is improved to the BIOU, which improves the small object detection performance of the model. In the experiment of the insulator faults detection in a low illuminance environment, the proposed algorithm has greater advantages over the five frequently-used object detection algorithms. Compared with the low illuminance environment object detection algorithms of the IA-YOLO and the GenISP with RetinaNet, the mAP of this method is improved by 9.77% and 10.35%, and the detection speed is improved by 7.23% and 10.16%, which bo

关 键 词：绝缘子故障检测 低光复杂环境目标检测 小目标检测 二阶段目标增强网络 弱监督机制 零目标图像增强损失函数 小目标特征增强SSD YOLOv7小目标检测算法 

分 类 号：TM721[电气工程—电力系统及自动化]