Analysis Of Acoustic Emission Signal Of Axle Fatigue Crack Based On Two-Dimensional CNN

Axle is the key supporting part of the train and one of the most basic rotating parts of the traveling part.During the operation of rail vehicles,wheelset axles are faced with a very complex operating environment while bearing loads,which is easy to cause faults.Fault diagnosis and prediction of the remaining life of the axle can not only facilitate the formulation of reasonable maintenance strategies,but also effectively avoid the occurrence of accidents,so as to improve safety and reduce the loss caused by the fault.In order to achieve classification identification and lifetime prediction of axle faults,this paper proposes a two-dimensional convolutional neural network-based approach for axle fault diagnosis.Convolutional neural networks(CNN)are a method of deep learning that,compared to traditional methods,automatically extracts features from the data,eliminating as much as possible the influence of expert experience on the final result.Fault classification identification: First,this paper adopts three analytical methods,namely,time domain analysis,frequency domain analysis and time-frequency joint domain(timefrequency)analysis,to convert the one-dimensional acoustic emission signal into twodimensional image data,and obtain the corresponding time domain diagram,frequency domain diagram and time frequency diagram.The three images contain rich time domain,frequency domain,and time-frequency joint domain information,respectively.The automatic extraction of convolutional and pooled layers using the operations of Alex net(a type of CNN model)can characterize the different types of faults and use the classifier for the purpose of fault classification.Then,the best fault classification identification method for three image data combined with Alex net was analyzed by comparison of classification identification accuracy and training time.Improvements were made to the Alex net model after determining the best combined scheme in order to reduce training time with no change in classification identification accuracy,and the improved Alex net model was compared to the original model.After the experiment,the classification identification accuracy of the improved network remained basically unchanged at about 98%,while the training time was reduced by almost a third.Among them,the frequency domain transformation of one-dimensional signals is the Fast Fourier Transform(FFT),the time-frequency transformation is the Continuous Wavelet Transform(CWT),and the wavelet basis function of CWT is Cmor3-3.Axle residual life prediction: The time domain signal of the axle fatigue cracking acoustic emission signal is transformed by FFT into a frequency domain amplitude signal,which is then normalized to a 2D pixel image as an input to a self-built CNN model.The abstract features that can be annotated at the axle failure stage are automatically extracted by the operation of convolutional and pooling layers to characterize the percentage of axle life at different periods,construct health indicators(HI),evaluate the performance of the life prediction network using RMSE and MAE as evaluation indicators,and verify the reliability and generalization of the network.