Research On Insulator Defect Detection Method Based On Deep Learning

Against the backdrop of the development of electricity during the 14 th Five Year Plan period,China adheres to the energy transmission pattern of "transmitting electricity from the west to the east and supplying electricity from the north to the south".The national power grid system is rapidly developing,and the range of lines that need to be installed nationwide is becoming wider and wider.Transmission lines are often damaged by harsh environments,natural disasters,and other factors,so ensuring the safe and efficient completion of circuit inspection tasks has become the top priority for protecting the stable operation of the power grid system.This article mainly focuses on insulator images collected by drones for defect detection of insulators.In response to the national strategic goal of "intelligent inspection",this article studies insulator defect detection methods based on deep learning background,proposes a hybrid attention mechanism encoding and decoding network to segment insulators in images,and proposes an improved algorithm based on YOLOv5 s to detect insulator defects in images.Firstly,to address the issues of varying insulator positions and complex backgrounds in aerial images,this paper proposes a hybrid attention based encoding and decoding network for insulator segmentation.The network mainly consists of three parts,namely encoder,mixed attention mechanism,and decoder.The activation convolution operation in the encoder adopts deep separable convolution calculation.During the skip connection process from encoder to decoder,a hybrid attention mechanism is integrated to highlight effective features and suppress redundant features.And comparative experiments were conducted with three networks,FCN,Unet,and Mask RCNN,respectively.Through experiments,it was found that the Dice coefficient of the hybrid attention encoding and decoding network designed in this paper is 0.90,which is 0.08,0.05,and 0.04 higher than the three models of FCN,Unet,and Mask RCNN,respectively.This verifies that the hybrid attention encoding and decoding network designed in this paper achieves better segmentation effect on insulator images,providing an effective foundation for subsequent defect detection of insulators.Secondly,on the basis of insulator segmentation,this paper improves the YOLOv5 s object detection algorithm to detect insulator defects with small targets.A bidirectional chain feature fusion module is designed,which adds four bidirectional fusion chain connections.By improving the network feature fusion part,the target information on the feature map is fully utilized,At the same time,the Semantic information of shallow features is further enhanced,and the small target feature information used for detection by feature map is fully obtained.This article conducted two separate experiments to verify the detection effect.Firstly,a comparative experiment was conducted with SSD,YOLOv3,and Faster RCNN models.The experiment showed that the improved YOLOv5 s algorithm achieved a defect detection m AP value of 91.2%,which was 8.4%,4.6%,and 2.6% higher than SSD,YOLOv3,and Faster RCNN models,respectively.It was verified that the improved YOLOv5 s network has stronger defect detection ability for insulators.Secondly,this article conducted comparative experiments to verify the impact of extracting insulators on detection accuracy.It was found that compared to defect detection without extracting insulators,the m AP value of extracting insulators for defect detection reached 92.6%,which increased by 5.9%.This verified the feasibility of the method of segmenting before detecting in this article.In summary,the hybrid attention encoding and decoding network proposed in this article and the improved YOLOv5 s insulator defect recognition method have certain theoretical significance and practical application value.