Research On Wheel Tread Defect Detection Based On Deep Learning

The train wheels are one of the important parts to ensure the safe running of the train.The accurate and rapid identification of the wheelset tread defects is conducive to timely maintenance,so that the occurrence of dangerous accidents is reduced,so that the interests of the people are guaranteed.With the rapid improvement of computer performance,it is a key factor to ensure the timely maintenance of wheelset and an important means to ensure the safe operation of trains to accurately and quickly identify the defects of wheelset tread by computer technology.On the basis of highly developed computer technology,modern deep learning technology has opened up a new world for image recognition.And use the corresponding data to analyze the model.The essence of deep learning technology is to use a large amount of data to learn,automatically mining out beneficial information,and finally meet the actual needs.The deep learning technology is to apply the deep learning model to the wheelset tread defect recognition work from a large amount of data,which greatly reduces the workload of manual feature design and makes the intelligent recognition of wheelset tread defects possible.In this paper,a classification method of wheel tread defect recognition based on convolutional neural network is proposed.By improving and optimizing the classic RESNET50 model,the common wheel tread defect images,such as cracks,peels,scrapes and pits,are detected and studied.Firstly,a large number of pictures of wheel tread defects were obtained,and the deep learning framework Py Torch was selected as the deep learning framework,and the identification method of wheel tread defects was studied using Py Torch deep learning framework.Based on the classical convolutional neural networks VGG16,VGG19,RESNET18,RESNET34,RESNET50 and RESNET101,the wheel tread detection was studied respectively.In the experimental process,transfer learning was adopted to improve the training efficiency,and a comparative experiment was conducted on the six models.The results show that the RESNET 50 model has better performance in wheel tread recognition.Then,the classic RESNET 50 model is improved and optimized.Add an attention mechanism after the first convolution layer of the Resnet50 model.At the same time,the loss function combined with Center Loss and Softmax Loss is quoted in the aspect of loss function,so that the depth features learned can be more distinguished,while the optimization algorithm uses the ADAM algorithm.Then,an automatic recognition model of wheel tread defects based on deep learning was obtained by training.In this research subject,a map library of wheel tread defects has been established,which contains nine kinds of defect pictures and one normal picture.The gallery contains 35,000 images,with an average of 3,500 for each category.Each image resolution is processed as 224×224.The improved model has great superiority in accuracy,and the recognition rate is as high as 91.20%.Experimental results show that the optimized recognition algorithm has good performance adopted in this paper is better than that of the traditional machine vision recognition method and the recognition method based on the unimproved Res Net50 network model.In the recognition of wheelset tread defects,the method adopted in this paper achieves better recognition accuracy and real-time performance.