Research On Catenary Equipment Fault Algorithm Based On Deep Learning

In recent years,with the rapid development of China’s electrified railway industry,the reliability and safety requirements of railway power supply are becoming higher and higher.Insulators are the main components of catenary equipment used for electrical insulation and mechanical fixation.Under the influence of complex natural factors and the surrounding environment,flashover fault of insulators happens from time to time.If the insulator flashover fault can be located quickly after occurrence,it is beneficial for the staff to timely repair the fault and save a lot of labor and time costs.Meanwhile,once the insulators are damaged,the insecurity of the power grid will be increased and the power supply of the power system will be affected.Therefore,monitoring the status of insulators and timely detecting of broken insulators are crucial to the normal operation of the power supply system.The main work of this paper includes the following aspects:(1)Based on the Unet+++ algorithm,the A-Unet +++ algorithm is proposed.In order to improve the feature extraction capability of the Unet+++ algorithm,atrous convolution is introduced in the jump connection stage of the encoder and decoder of Unet+++,so as to retain the spatial features of the image without losing the image information,obtain a larger receptive field,and further optimize the feature fusion problem of the network at different scales.(2)In the insulator flashover fault monitoring project,the whole experiment is divided into two steps: light extraction and location comparison,aiming at the characteristic that bright light will appear when flashover occurs.In the bright light extraction stage,aiming at the problem of insufficient number of existing bright light extraction data set samples,random clipping,random flipping and other methods were used to expand the image,and a total of 1800 images were generated.Using PS to generate bright light’s label images;then comparing with Lite Seg,Unet,Unet++,Unet+++ and A-Unet+++ to do the contrast experiment of light extraction.Compared with Unet+++ network,the test results show that the IOU value of the improved A-Unet+++ + algorithm is 10.18% higher and the speed is0.13 s every frame lower than that of Unet+++ network.The feasibility and practicability of A-Unet+++ in flashover fault identification are verified.In location comparison stage,using vision insulator field data for the data sets,then marked by Label Img,through the image of insulator with light position location,set up the threshold of 80%,if more than80% of the position of the insulator is bright,it is considered that the light belong to flashover,and the insulator has flashover fault,then outputing the position of the insulator.Finally,the image segmentation algorithm based on deep learning realizes the identification and location of insulator flashover fault monitoring.(3)In the insulator damage detection project,the close-up insulator images are used as the data set.In view of the problem that the data scale is too small,then using random rotation and translation transformation methods to realize data expansion,enhancing the diversity and completeness of the data set,and improving the model generalization ability.After that,labeling all images and generating label images at the same time.At last,three classical target detection algorithms,Yolov5 s,SSD and Faster-RCNN,were used for damage detection experiments.By comparing the test results,it can be seen that Yolov5 s network has better performance in the application of insulator state detection.