Research Of Detection And Defect Recognition Algorithm For Components And Parts Of High-speed Railway Catenary

Catenary is one of the important equipment to ensure the safe and reliable operation of high-speed railway,and its safety monitoring are very strict.Based on a detailed review of the domestic and foreign research status of components and parts defect recognition in catenary images,a two-stage defect recognition framework based on deep learning was studied,which realized automatic positioning and defect recognition of components and parts in catenary images.The main research results are as follows:(1)A kind of two stage generic defeat recognition framework was developed.An improved fully convolutional one-stage object detector(FCOS)was used to locate the components and parts on high-speed railway catenary,and the Res Net50 classification network with focal loss(Res Net50-FL)was applied to recognize defective parts.A high-speed railway catenary data sets was established,and it was verified that the framework had high commonality,strong robustness on this data sets.(2)Considering lack of high-speed catenary parts defect samples,generation model based on cycle-consistent generative adversarial network(Cycle GAN)was used to generate defect samples.The advantages and disadvantages of defect sample generation based on generation adversarial network and style transfer were analyzed.The experiments on the defeat data sets compared the generation effect of deep convolution generative adversarial network(DCGAN)and Cycle GAN,indicating the availability of defect sample generation based on Cycle GAN.(3)Aiming at the scene that defect parts are unknown or defect samples are difficult to collect,defect recognition algorithm based on generation adversarial network was adopted.This algorithm only learns the distribution of normal data through the network,and can distinguish the defect or normal state of the image by comparing the difference between the reconstructed image and the input image.The validity of the proposed algorithm was verified by a comparative experiment on the hanger data set.