Inclined Support Sleeve Defect Detection Of High-speed Railway Catenary Parts Based On Deep Learning

With the rapid development of high-speed railway,the state has higher and higher requirements for the power supply safety and power quality of high-speed railway electric traction syste.There are many kinds of faults that may occur in traction power supply system,among which the probability of catenary system failure accounts for a large proportion.Therefore,the catenary parts defect detection is very important,because it is related to high-speed rail safety.As an important component of catenary,the defect detection of inclined support sleeve is of great significance.At present,the method of defect detection of inclined support sleeve is mainly machine learning method,and its accuracy and speed are not high enough.In order to improve the accuracy and speed of defect detection of inclined support sleeve,this paper adopts deep learning method.In this paper,the defect detection task of inclined support sleeve is divided into two steps.The first step is to complete the location recognition of inclined support sleeve,and the second step is to complete the defect recognition of inclined support sleeve.First of all,the data set of location recognition of inclined support sleeve is produced.In this paper,a large number of images of catenary parts are copied from Pingxiang North Station,and then the images of the initial data set are obtained by preprocessing the image data.Then labelimg software is used to annotate the initial data set images and get the initial data set.then the data set is extended by affine transformation to get the final data set.Secondly,an appropriate target detection model is improved to complete the location recognition and interception task of inclined support sleeve.Three target detection models with excellent performance are selected from the current mainstream target detection models,namely Faster R-CNN,Yolov3 and Efficientdet models.The three models are used to locate and identify the inclined support sleeve.After comparison,it is found that the comprehensive performance of Efficientdet model is the best.Then,the Efficientdet model is improved,and the improvement based on feature fusion network is proposed.Based on this,the improved new network model is used to locate and identify the inclined support sleeve,and the effect is very good.The improved Efficientdet model can well complete the location and identification task of the inclined support sleeve.Finally,a suitable classification neural network model is improved to complete the defect recognition task of inclined support sleeve.Three networks with excellent performance are selected from the current mainstream classification neural networks,namely,Vgg16,resnet50 and mobilenetv2 networks.The three classification networks are used to identify the defects of the inclined support sleeve.After comparison,it is found that the mobilenetv2 model has the best comprehensive performance.Then,the mobilenetv2 network is improved,and the backbone network and classifier network of the mobilenetv2 model are improved.Based on this,the improved new network model is used to identify the defects of the inclined support sleeve,and the effect is very good.The improved mobilenetv2 model can perfectly complete the defect identification task of the inclined support sleeve.