Research On Defect Detection Method Of Transmission Line Insulator Based On Deep Learning

As the electricity consumption of the whole society in our country continues to grow,the number and coverage of transmission lines are also increasing year by year.As an important part of transmission lines,insulators have multiple functions such as mechanical fixing,supporting lines and realizing electrical insulation.If the faulty insulator is not detected and replaced in time,it will cause a brief collapse of the power system and certain economic losses.Since most of the transmission lines are located far away from the city such as mountains and large ridges,The inspection method used by inspectors for field surveys is not only low in safety,but also high in cost.At present,some provinces and cities have gradually used drones to take inspection images.However,due to the unfixed shooting angle of the drone,the complicated background and the few samples of defective insulators,the detection of insulator defects is facing great challenges.Based on the scientific and technological project of Zhejiang Electric Power Company,this article takes the inspection images of insulators taken by drones as the research object.Based on deep learning,the SDMask R-CNN network is proposed to complete the positioning and segmentation of the insulators in the inspection image,and the segmented insulator pictures and corresponding masks are sent to the subsequent classification network for defect voting judgment,In this way,the self-explosion defect detection of insulators can be realized.The main research contents and results of this paper are as follows:This paper proposes the SDMask R-CNN network to locate the insulators in the inspection pictures.Compared with the MaskR-CNN network,there are several improvements as follows:Performs K-means one-dimensional clustering based on the prior knowledge of the insulator aspect ratio,and adjusts the corresponding RPN ANCHOR Ratio;Apply the SE module to ResNet101 to give different weights to different channels;Change the 3×3 convolution operation in the feature extraction module of Mask R-CNN to a hole convolution with a hole number of 2,and P6 in the feature pyramid is obtained by convolution with a step size of 2 and a number of holes of 2.,which increases the receptive field of the feature extraction network,and further improves the detection ability of Mask R-CNN for insulators.In this paper,a classification network is designed according to the defect classification requirements of insulators.VGG16 is used as the base of the convolutional layer,and the SPP network is introduced to solve the problem of different sizes of input insulator images.Besides,in order to improve the classification accuracy and prevent over-fitting,a series of methods are adopted:transfer learning,data enhancement,weighted loss function,Dropout layer,L2 regularization.In this paper,the positioning network and the classification network are cascaded to form the final insulator defect detection model.The insulators detected by the positioning network and the corresponding mask are sent to the classification network to vote whether they are defective insulators.The final test mAp can reach 82%,which is an increase of 16%compared to the direct detection with YOLOv3.