Research On Defect Detection Method Of Key Components Of High-speed Railway Catenary Based On Deep Learning

The security of the power supply is of utmost importance for the high-speed rail traction system,given the rapid advancement of China’s high-speed rail,the safe operation of the train is directly impacted by the catenary,which is a crucial component.The quality and working condition of the catenary,which serves as the primary power source for the traction locomotive,is crucial in guaranteeing the safe and efficient functioning of the entire traction power supply system.The catenary system has the characteristics of many parts,a wide range of faults,and a long time to repair faults.Therefore,Improving the accuracy and speed of component defect detection in the catenary system is of utmost importance.At present,the key components that cause defects in the catenary mainly include insulators and hanging strings,which represent small parts and large parts respectively.In view of the poor detection effect of these large and small parts,this paper takes suspension strings and insulators as the research objects,and uses deep learning methods to conduct defect detection research on them.The main contents are as follows:In this thesis,the high-definition catenary images taken by the 4C system inspection vehicle are used as the research data basis.When constructing the network training data set,in view of the problem that there are few defect samples of suspension strings and insulator parts in the 4C images,the loop generation confrontation network is used for the defect samples.Carry out data expansion experiments,and manually select the data pictures after the experiment expansion to ensure that the network has sufficient positive and negative samples for training.Targeting the issues of inadequate detection accuracy and challenging identification in traditional methods utilized for detecting faults in catenary suspension strings,an improved Cascade R-CNN hanging string defect detection network is proposed,Introducing a recursive pyramid network structure enhances the multi-scale capability of the feature map,resulting in better and more significant image features that enhance the network’s ability to detect defects;in addition,to solve the problem that the position of the defect generated by the hanging string is not fixed,a hole-deformable convolution structure,that is,the SAC-DCNv2 structure,is added to make the network weak in the defect.More receptive field information can be obtained in feature extraction,so that the sampling position can be better moved to the area of interest of the network.According to the experimental results,the detection accuracy of the network has been enhanced,leading to an accurate identification of various defect states in the suspension string.In view of the small proportion of catenary insulators in 4C images and the difficulty of identifying contamination defects,this paper proposes a two-stage insulator contamination defect detection algorithm of "detect first,then identify".Insulator locations are recognized using the YOLOv5 target detection network in the first stage.Considering that the insulator on the inclined arm is in a tilted state,in order to reduce the interference of some useless background information contained in the horizontal positioning frame,a rotatable detection anchor frame mechanism is added to enable the network to detect the inclined insulator target,and finally the positioned Insulators are cut.In the second stage of the pollution recognition network,an enhanced U-Net semantic segmentation network is applied for identifying pollution defects in insulators,taking into account that the cropped image exclusively comprises of insulators.In order to enable the network to accurately segment the pollution defects of insulators,it is improved by adding a multi-scale context module and an adaptive perception module,so that the network can obtain richer context information,pay more attention to the contamination area of interest,and enhance the feature mapping ability of the network,thereby improving Pollution defect detection capability of insulators.Experimental results show that the two-stage algorithm can effectively detect contamination defects of insulators.