Analysis And Research Of Acoustic Emission Signal Of Fatigue Crack Of Axle Based On DBN And Support Vector Machine

With the rapid economic development,the running speed and carrying capacity of rail transit vehicles have also gradually increased.The running part is an important part of the rolling stock.The axle is the key part of the entire running part.Its operating conditions directly determine the running state of the bogie and affect the safety of passenger travel and cargo transportation.However,under long-term high-intensity operating conditions,the axles are prone to fatigue damage and cracks.When the cracks develop to a certain depth,cold cutting will occur.Therefore,it is of great significance to formulate plans for real-time monitoring and analysis of locomotive axles.In order to better analyze the acoustic emission data(Acoustic Emission,AE)of the axle,this paper proposes a model algorithm based on deep belief network(Deep Belief Nets,DBNs).This article first introduces the basic principles and algorithm flow of DBN,and then conducts a comprehensive analysis of the main parameters that affect network performance,and draws important conclusions about the combination of various variables.In the analysis and research of feature fusion and extraction of DBN,Principal Component Analysis(Principal Component Analysis,PCA)is used to reduce the dimensionality,so that the layer-by-layer extraction effect of the network can be visualized in three-dimensional space.In the process of classification and recognition of the collected and fused acoustic emission signal data,two hybrid algorithms EMD-DBN and DBN-SVM based on deep confidence network were constructed.The former uses Empirical Mode Decomposition(EMD)to pre-process the data and uses Tsallis energy entropy to reconstruct the data and then uses the DBN network for classification and recognition;the latter directly uses the DBN algorithm to perform feature fusion and extraction,And then use the processed features as support vector machine(Support Vector Machine,SVM)input data to identify and classify,and through horizontal comparison with a variety of traditional networks,it is found that the DBN-SVM network model has a higher performance in processing mixed signals The superiority and strong pattern recognition ability.Then,a DBN-SVR network model is constructed to perform time series prediction and life regression analysis on the identified crack AE signals,respectively.The model uses a support vector regression(Support Vactor Regression,SVR)layer to replace the softmax layer on top of the DBN structure,so that the model has higher preprocessing performance and stronger robustness,and then combines MSE and R^2 evaluation indicators It can be found that the better timing prediction effect of the DBN-SVR model is basically consistent with the original data set samples.In life prediction,the traditional root mean square(RMS)index is improved and optimized to obtain the relative root mean square value(Relative root mean square,RRMS)and the reconstructed root mean square value(ORRMS)According tothis,the main stage of crack propagation is obtained and regression prediction is carried out.The horizontal comparison with the traditional neural network shows that the DBN-SVR model has high prediction accuracy and good engineering practical value.