Analysis Of Fatigue Crack Acoustic Emission Signal Experimental Data Based On Time-Frequency Image And CNN

The bogie of railway vehicles is one of the most important components in the structure of rail train,and the axle is the key component in the bogie and prone to failure under the condition of bearing long-term load.Besides,the feature extraction of fault acoustic emission signal is very important in various fault detection schemes.In order to realize intelligent fault diagnosis of axle faults,there is an intelligent diagnosis method of axle faults based on time-frequency image and Convolutional Neural Network(CNN)proposed.Time-frequency image,obtained by time-frequency analysis of signal,contains abundant time-frequency joint domain information and can reflect the instantaneous frequency and amplitude at each moment,which can be used to extract fault signal features.In addition,CNN,as one of the deep learning algorithms,has strong performance in image recognition,which can be applied to the classification of time-frequency images of axle fault signals.The time-frequency image is obtained by carrying out time-frequency analysis of the acoustic emission signal of train axle,and fed to CNN as sample data.Then,CNN trains time-frequency images of different fault types,and conducts final tests to complete classification,so as to realize intelligent fault diagnosis of train axles.Different time-frequency analysis methods will generate different time-frequency images,which will affect the final classification effect.Thus,there are three time-frequency analysis methods(short-time Fourier transform,S transform,and continuous wavelet transform)used in the time-frequency analysis of signal.Besides,different wavelet basis functions affect the continuous wavelet transform.For this reason,three kinds of wavelet basis functions(Morlet,db4,and cmor3-3)are used for time-frequency analysis and comparison to obtain the best time-frequency analysis method.In order to achieve the best time cost and recognition accuracy,three CNN network models(AlexNet,GoogLeNet,ResNet50)are used to compare the time of the classification recognition effect and network training.The results show that the scheme of continuous wavelet transform with cmor3-3 as wavelet basis function and AlexNet has the best recognition accuracy and time cost.Finally,the effects of two CNN structural parameters,batch size and Dropout ratio,on the recognition accuracy are studied to further improve the network performance.