The Research Of Defect Detection For Overhead Contact Line Components Based On Deep Learning

The miles of electrified railway is growing continuously,and the passenger and freight transportation are also growing year by year.This makes higher requirement for the safety of traction power supply system.As there is no spare of contact network,so the maintenance of contact network is very important.The main work of operation and maintenance is to detect the contact network,find the defects in time and eliminate them.In recent years,railway has widely adopted catenary suspension state detection and monitoring device(4C device)to implement catenary defect detection.By collecting catenary image and then analyzing the image to form a report,but the image analysis in the 4C device still does not break away from the artificial realization,which limits the further play of the detection ability of 4C device.In this paper,take catenary image as the research object,explore the use of image processing algorithm to achieve machine vision,instead of manual image analysis work,so as to achieve the 4C device can truly collect images,image analysis and form reports in one.At present,the image processing algorithm with the greatest development potential is deep learning,which has the characteristics of automatic learning parameters and good generalization.Therefore,this paper studies the use of deep learning to realize the task of defect detection of contact network.The main tasks of this paper are as follows: firstly,the catenary image is preprocessed,and the image is enhanced by limiting contrast histogram equalization,so as to improve the accuracy of subsequent detection tasks;Then,two classical algorithms Faster R-CNN and YOLOv2 in deep learning were selected to realize the positioning of catenary parts.In order to improve the feature extraction capability of the model,Res Net-50 was used to replace the original feature extraction network.After the training and testing of catenary image of Dacheng line,it was found that both algorithms could achieve target positioning.And can meet the actual use;Then,the YOLOv2 Res Net-50 model was used to locate the defects of parts,and it was found that the location effect of the defects of insulators was poor.Next the research is conducted on insulator defect recognition.Due to insufficient defect samples,the generative adverssion network in unsupervised learning is adopted to extract features from the input image by using the depth encoder,and then the residual module is used to optimize the feature map,and then the depth decoder is used to restore the feature map to obtain the reconstructed image.By comparing and analyzing the reconstructed image and the input image,the judgment of the defect is realized.By testing the insulator samples located,the validity of the model for insulator defect detection is verified.Finally,the network is cascaded,and the overhead contact line image is used to test,which verifies the feasibility of the model detecting the defects of parts.In addition to model building,the most important thing in the experiment is model training,and the training results are directly related to the detection results of the model.In this paper,manual labeling is adopted to make data sets,and Adma optimizer is adopted in the gradient descent algorithm to monitor changes in the verification results of the verification set during the training process,so as to prevent overfitting of the model.