Research On State Detection Method Of High-speed Rail Catenary Small Fasteners Based On Deep Learning

In recent years,high-speed railways has developed rapidly at home and abroad.It has played a pivotal role in transportation.As the main artery of the high-speed railway,the catenary is one of the main components of the high-speed railway power supply system and the only power supply network in the high-speed railway system.The reliability of the working state of the catenary support device determines the safety and reliability of the traction power supply.The small sleeve fastener is an important connection fixture of the catenary support device,which is of great significance to ensure the safe and stable operation of the catenary.The research work in this paper is based on the catenary suspension state detection and monitoring device(4C)project in the 6C system,and takes the catenary arm images collected at night by the inspection vehicle as the research object.And proposes a threelevel cascade detection architecture from positioning to pre-processing of pictures to status detection,which is suitable for small fasteners of catenary.Firstly,to solve the problem of poor positioning of small components in catenary.a positioning data set including split pins and brace sleeve screw of the small target fasteners of the catenary support device was constructed.Based on the SSD algorithm,the input,number of layers,loss function,default boxes,etc.of the model are optimized according to the characteristics of the catenary arm picture and the fasteners,a algorithm named SSD512 is designed to realize the accurate positioning of small fasteners of the catenary support device.Then,in order to solve the problem of the imbalance in the number of categories of the small fasteners,the poor adaptability of the detection methods in complex railway scenes,and the lack of robustness,the pre-processing work was carried out on the pictures.In order to avoid the overfitting of the model caused by the class imbalance problem,DCGAN expands the defect samples,so that the number of samples reached a balance;And a data set for the training of semantic segmentation model was constructed.Semantic segmentation based on U-Net realizes the segmentation in the picture,filters out irrelevant image background,and makes full use of the normal samples in the case of insufficient defect samples,which can also strengthens the robustness,efficiency and accuracy of the model.Finally,based on VGG16,an image classifier based on a convolutional neural network and a color classifier are designed to realize the state detection of small fasteners.A three-level cascaded detection architecture is designed,and split pins and brace sleeve screw are taken as examples.The accuracy and robustness of the proposed method are verified through experiments.