| As an important support and running part of high-speed train,train wheel has high-speed rolling contact with track in complex environment for a long time,and its tread damage has become one of the important factors affecting the safety and comfort of train operation.At present,the diagnosis method of wheel set tread damage by manual or auxiliary equipment has some problems,such as slow diagnosis speed,low efficiency and poor economy.Therefore,this thesis carried out the diagnosis research on high-speed train wheelset tread damage based on convolutional neural network.The main research contents are as follows:(1)Aiming at the problem of low efficiency of the traditional diagnosis method of artificial wheel set tread damage,a wheelset tread damage diagnosis method based on convolutional neural network is proposed.Firstly,the mechanism of wheel tread damage of high-speed train is analyzed to show the necessity of diagnosis and research on wheel tread damage.Then,the image data of wheel tread damage is collected and its characteristics are analyzed to describe the texture particularity of the damage area.Secondly,the framework and training process of ordinary convolutional neural network are introduced in detail,and the reason why it is suitable for wheel set tread damage diagnosis is explained.Finally,the diagnosis flow of this network is briefly described and the advantages and disadvantages of ordinary convolutional neural network are discussed.(2)Aiming at the problem that the accuracy of traditional convolutional neural network is not high in the diagnosis process of wheelset tread damage,a high-speed train wheelset tread damage diagnosis method based on pyramid split attention is proposed.Firstly,a transfer learning method was used to pre-train Image Net data set to obtain model weights,which were transferred to wheel tread damage features and fine-tuned.Secondly,the 3×3 convolution in RESNET-50 residual block was replaced by PSA module to obtain the new EPSA-Res NET,which integrated the multi-level features of spatial and channel attention and carried out feature re-calibration adaptively.Finally,the diagnosis results of wheel-set tread damage were obtained through the test set,and the comparative analysis were conducted on the same data set.Experimental results show that this method can diagnose wheel tread damage of high-speed train accurately.(3)Aiming at the problem that it is difficult to identify the damage category of high-speed train wheelset tread,a multi-category damage diagnosis method for high-speed train wheel set tread based on Inception-Res Net network is proposed.Firstly,the collected damage image samples are preprocessed,and multi-category damage data sets are made.Secondly,the principle of InceptionRes Net network is analyzed.This network can extract the damage characteristics of wheelset tread in parallel,and based on this network,a multi-category damage diagnosis model of high-speed train wheelset tread is constructed using tensorflow framework.Finally,the model was used to diagnose the damage of the test data set,and the diagnosis results of multiple types of damage were obtained,and the comparative experiments were conducted on the same data set.Experimental results show that this method can effectively diagnose the damage state information of wheelset tread by multiple categories. |
