Research On Classification Technology Of Track Surface Diseases Based On Deep Learning

Under the environment of rapid social and economic development in China,high-speed railway transportation has become a very important mode of transportation.With the passage of time and the increasing number of trains,a variety of diseases have appeared on the surface of the rail.The appearance of rail surface disease will reduce the quality and corrosion resistance of rail,which will pose a threat to the safety of train running.In order to improve the classification accuracy of rail surface disease detection,reduce the workload of manual inspection or completely replace manual inspection,this paper proposes the rail surface disease classification technology and research based on deep learning,and introduces deep learning into the engineering application of rail surface disease classification.The main research contents of this paper are as follows:(1)The research status of rail surface disease and deep learning at home and abroad was investigated.Res Net-50,Incetion-V3 and Xception were selected to analyze and compare the target detection algorithms at the present stage.According to the advantages and disadvantages of the three networks,Res Net-50 network was selected to be used in rail surface disease classification.Secondly,the data set needed in this paper is obtained through different ways,including the production of training set and test set,and the problem of insufficient data is solved by means of data enhancement,and the corresponding data preprocessing is done.(2)In the actual detection,images collected under different lighting conditions will affect the performance of the model,leading to a low detection rate of disease classification.In order to solve this problem,an optimization of Res Net-50 network Conv5 structure based on Transformer is proposed to achieve dynamic attention and better generalization of the model,reduce the number of parameters and calculation amount,and improve the classification and detection accuracy of the model without changing translation and rotation.(3)According to the characteristics of the data set in this paper,there is the nature of implicit correlation between diseases and diseases and between diseases and backgrounds.In order to solve the problem that the convolutional layer cannot extract the long distance correlation from the feature space,A deep residual network model A-Res Net-50 based on the attention mechanism is proposed in this paper.In this model,an attention mechanism is added after the backbone network of Res Net-50,before the full connection layer and after each module of the overall structure of Res Net-50,which can better solve the problem of feature capture and improve the classification detection rate.(4)According to the characteristics of the data set in this paper,there are a large number of negative samples in the data,which interfere with loss in the process of classifier training.In this paper,a deep residual network model L-Res Net-50 based on loss function optimization is proposed.Modular length was introduced to pay different degrees of attention to positive and negative samples,simple and difficult samples in data set.The three loss functions of gradient coordination mechanism,Focal loss and cross entropy were applied to the Res Net-50 network,and the data set in this paper was used to train and test the network respectively.Under the premise of guaranteeing the model loss,the classification detection rate was significantly improved.