Research On Detection Method Of Wheelset Tread Defects Under Vision System

The train wheelset is a key component to ensure the smooth running of the trai n at high speed,and timely detection of defects in the wheelset tread can reduce train operation risks.Due to the shortcomings of manual inspection,such as low efficiency,high labor intensity,and subjective factors,intelligent inspection has become t he development direction of wheelset defect detection.Among them,machine vision technology is the current mainstream,which reflects such as high detection efficiency and dynamic detection.That is,the train does not need to stop and other advantages.I n addition to the hardware design of the wheelset tread machine vision inspection equipment,the most important thing is the research on the defect recognition algorithm.This article focuses on the deep learning technology,supplemented by the image processing algorithm,to study the tread defect recognition method.The main research results of this article are as follows:(1)Propose a dual target detection network recognition method.The task of wheelset tread defect detection is divided into two steps: wheelset tread area detection and tread defect detection.By using two target detection network models to identify tread defects in series,the experiment verifies the effectiveness of this method in improving the accuracy of defect recognition.(2)An adaptive direction correction method for wheelset tread area is designed.Aiming at the problem of deflection in different directions in the tread area caused by the collection reasons,the method proposed in this paper can correct the tread area of the wheel set in different deflection directions to the horizontal direction.And on this basis,the SSD network is trained,and the rapid identification of the wheel-to-tread area is used.The average accuracy is 99.8%,and the calculation time for a single image is only 31.8ms.(3)Propose an improvement method of YOLOv3 model.By an alyzing the size distribution of tread defects,combined with the target size range predicted by the anchor box assigned to each output structure by YOLOv3,it is concluded that the output structure used to predict large targets in YOLOv3 cannot play a rol e in the task of tread defect recognition.Therefore,the output structure is deleted to achieve the purpose of improving the calculation speed of the model.Experimental results show that compared with YOLOv3,the optimized T-YOLOv3 has an average accuracy loss of only 0.6%,but the average calculation time for each tread image is reduced by 7.1%.The experiment verifies the effectiveness of this method.(4)Propose a data enhancement method.Aiming at the problem that it is difficult to obtain the tread defect data set,this paper proposes a method of using a generative adversarial network to generate tread images containing defects in order to achieve the purpose of expanding the data set.The experimental results show that after adding the generated tread image to the original tread defect training set,the average recognition accuracy of the trained T-YOLOv3 model in the test set is increased from 89.9% to92.1%,which significantly improves the model training effect.