Research On The Defect Diagnosis Method Of Wheelset Tread Of High-speed Train Based On Convolutional Neural Network

As an important support and traveling component of high-speed trains,wheelsets will cause minor failures in the pedal surface due to the rolling contact of the wheel rails,and if they are allowed to deteriorate,they will easily cause serious failures such as wheel-to-pedal fractures that affect the safety of train operation.Traditional methods use manual detection,ultrasonic flaw detection,magnetic particle flaw detection,such methods usually require falling wheel processing,which is not conducive to the intelligent operation and maintenance of high-speed railways.In recent years,defect detection technology combined with computer vision and deep learning has become an effective solution for industrial defect detection.This is due to the powerful feature extraction and characterization capabilities of convolutional neural networks,making them more adept at processing complex industrial image data.Therefore,in this thesis,for the effective extraction of the characteristics of the wheelset of high-speed trains under a limited sample,a study of the fault diagnosis method of the wheelset of high-speed trains based on convolutional neural network is carried out,and the specific research content is as follows:(1)It is difficult to obtain the data sample of the wheel-on-wheel tread surface of the high-speed train required for model construction.Firstly,the wheelto-tread surface defect image acquisition system is used to collect the image;secondly,the wheel-to-tread defect category image is classified to make a wheelto-tread defect classification dataset;finally,the collected image is sorted out and labeled,and the wheel-to-tread defect detection data set is established.Provide data support for the subsequent design of wheel-to-tread defect classification model and wheel-to-tread defect detection method.(2)Aiming at the difficult problem of classifying the defects of the wheelpair tread of high-speed trains under small sample conditions,a neural network model based on parameter-less attention and spine is proposed.Firstly,the pretrained convolutional feature extraction network is used to extract the feature m AP of each category;secondly,the more expressive categorical features are extracted for the tread image under the limited training samples by the non-parameter attention module(Sim AM);finally,the local and overall semantics are obtained by the spine fully connected layer segmentation to make correlation decisions,and the strong distinguishing representation of each class is obtained,and the fault discrimination is performed.Experimental results show that this method is superior to some mainstream identification methods.(3)Aiming at the problem that the local semantics and position information extraction of traditional deep learning detection algorithms are insufficient,a Yolov4 high-speed train wheel-to-pedal defect detection method based on multiscale feature fusion is proposed.Specifically,a multi-scale feature fusion module is proposed;this module embeds Yolov4 to help the network integrate local semantics and position details,providing high-quality feature learning for the network.At the same time,the improved loss function can speed up the convergence of the network.Experimental results show that the method proposed in this section can effectively improve the detection accuracy of the detection model.In this thesis,the tread surface of the wheel pair of high-speed train is studied,the experimental data set required for the study is collected,the tread defect classification method based on convolutional neural network is proposed to solve the problem of small sample wheel tread defect classification,and the tread defect detection algorithm based on convolutional neural network is proposed,Solve the problem of low accuracy of wheel-to-wheel tread detection.Complete the classification and detection of wheel alignment tread defects in high-speed trains.