Defect Detection Of Cotter Pin On Railway Catenary Based On Deep Learning

As a key infrastructure for providing reliable and uninterrupted power supply to high-speed railway multiple units and electric locomotives,the stable working state of the railway catenary is a safety guarantee in railway transportation.In practical situations,the catenary is prone to faults due to the impact of external environment and its own factors,leading to problems in the traction power supply of trains and ultimately leading to a series of safety accidents.In the structure of the overhead contact system,the cotter pin is a fastening component in the cantilever structure of the overhead contact system.However,the harsh outdoor environment and vibration generated during long-term operation of the train may cause it to fall off,causing the cantilever structure to become loose,affecting the safe operation of the train.Therefore,it is necessary to timely detect the status of the cotter pin to ensure that the train can safely travel.The traditional method of detecting the defects of cotter pins is to conduct patrol inspection manually,which is often accompanied by low efficiency and prone to missed inspections.The defect detection method based on machine vision can effectively improve the shortcomings of manual detection methods,so the rise of machine vision provides a new idea for the defect detection work of cotter pins.Based on the deep learning technology in machine vision,this topic analyzes and studies the positioning,defect dataset expansion,and defect recognition of OCS cotter pins,mainly in the following aspects:(1)Generally,the captured cotter pin images generally have problems such as high resolution,and the gray value of the cotter pin position is relatively close to the background area.If the captured images are directly used to train the network,it will bring huge computational complexity to the network.In order to save computing resources and improve the detection efficiency of the network,it is necessary to locate and extract the cotter pin regions in the original image to avoid the network from ineffective processing of non cotter pin parts of the image.To solve the problem of locating the cotter pin region,this paper proposes a network model that combines efficient attention mechanism SKNet and multi-scale wide residual.This model achieves accurate positioning of the cotter pin region by extracting the spatial features of the original image.Finally,the two-source region center point method is used to synthesize the detection results to complete the output of the cotter pin position information,facilitating subsequent defect detection work.(2)Aiming at the problems of unstable training and slow convergence rate in deep convolutional generation confrontation network(DCGAN)for data enhancement of split pin defect samples,the algorithm and structure of DCGAN were optimized respectively.This paper proposes a depth residual generation adversary network(CDRGAN)based on Canberra distance.In algorithm,the JS divergence in the original network is replaced by Canberra distance,and gradient penalty optimization is introduced on Canberra distance,which improves the stability of the model during training;Structurally,a residual network is added to the generator to accelerate the convergence process of the model.In order to improve the quality of generated samples,a lightweight CNN network model sensitive to local feature differences in images is built to achieve secondary filtering of generated samples.The experiment shows that compared with the defect samples generated by DCGAN,the optimization rate of the improved model on multiple evaluation indicators is greater than6.0%;In addition,Alex Net trained using defect samples generated by CDRGAN improved its recognition accuracy by 4.0%,effectively improving the data enhancement effect of DCGAN on split sales defect samples.(3)In order to overcome the limitations of previous studies such as limited types of identifiable defects,inability to locate defects,large model parameters,and slow detection speed,an improved YOLOv5 network was used to detect and locate defects in cotter pin images.Training the network on the CDRGAN expansion dataset,using the Ghost Module to replace the commonly used convolution module of the YOLOv5 backbone extraction network,reduces the amount of network model parameters;To ensure good detection performance,a Squeeze and Elimination(SE)attention module is added to the backbone network to enhance the algorithm’s ability to detect targets;Introducing the structure of Bidirectional Feature Pyramid Network(Bi FPN)into the feature fusion network further enhances the network feature fusion capability.The results show that the model successfully identifies and locates common defects.Compared to the YOLOv5 algorithm,the model volume is reduced by 21%,and the single sheet detection speed is increased by 17.4%.