Wayside Acoustic Diagnosis Of Train Bearing Based On STFT And Improved LENET5 Networks

Wheelset bearing is one of the key parts of train running part.Its healthy operation is an important guarantee of train safety.With the characteristics of non-contact measurement,low cost and early fault warning,the wayside acoustic detection technology has a good application prospect.In recent years,deep learning has developed rapidly,and has been well applied in all walks of life.Among them,convolutional neural network is an important part of deep learning,which has strong capability of feature extraction,feature fusion and feature recognition.In this paper,based on the characteristics of Doppler distortion of wayside sound signal of train bearing,combining the use of short-time Fourier transform and convolutional neural network,the method of rail side acoustic fault diagnosis of train at different speeds is studied.The main research contents are as follows:(1)A method for wayside acoustic diagnosis of train bearing based on STFT and improved Lenet5 network was studied,and the effectiveness and advantages of the proposed method were verified by experimental comparative analysis.First of all,this paper studied the method of making the image set of train bearing wayside sound signal based on STFT,and built the network structure based on the traditional Lenet5 network according to the signal characteristics.Secondly,the network parameters are designed,the pooling layer mode is selected,and the connection mode between S2 and C3 is selected.Then,the superiority of the proposed method is verified by comparing the experimental data and analyzing the proposed method with the time-domain sample image input method.Finally,compared with the traditional KNN algorithm,the results show that the proposed network has a higher recognition rate for bearing fault types.(2)A training sample expansion method based on resampling technology was studied to improve the generalization ability of the aforementioned Lenet5 network.In this method,timedomain interpolation resampling method is used to construct wayside signal samples at different speeds,which can be used as Lenet5 network training samples to improve its fault identification ability at different speeds,and the effectiveness of the proposed method is verified by experimental comparison.Firstly,introduce the principle of interpolation resampling method in time domain,and verify the validity of the method by simulation and experiment.Then,elaborate the training sample expansion method based on resampling technique in detail,and the experimental data is expanded in different combinations,which lays a good foundation for the subsequent comparative experiment.Finally,carry out the comparison test of training samples of different combinations,and the result verifies that the expanded training samples can effectively improve the identification ability of Lenet5 network to samples under different speeds and enhance the generalization ability of the network.