Experimental Data Analysis Of Acoustic Emission Signal Of Axle Fatigue Crack Based On Wavelet Transform And DBN

High speed and heavy load are the major trends in the development of rail transit vehicles in China.Therefore,how to realize the online detection of the health status of trains has also become a hot issue for the majority of railway workers and scientific research institutions.The axle is one of the main parts of the running part of the train.The working environment is extremely complex,and its reliability and service life will affect the operation safety of the entire vehicle.Therefore,it has very important research value for the detection and identification of axle failure and the prediction of remaining life.In this paper,the fatigue crack signal collected from the axle fatigue fatigue acoustic emission experiment is taken as the research object.First,in order to identify the crack signal from the noise and knock interference signals,a method of combining wavelet transform and deep belief network is proposed.In this paper,the three types of signals obtained by the acoustic emission experiment are first grouped.The selected base wavelet is Morlet wavelet suitable for mechanical fault signal analysis.The three types of grouped signals are continuously wavelet transformed to obtain the transformed wavelet coefficients for each group of signals,Extract six kinds of time domain features and two kinds of time-frequency domain features for wavelet coefficients to form a feature vector matrix and input it into the DBN recognition network.At the same time,this paper also inputs the same data into a traditional recognition network(MLP neural network).Compared with the DBN recognition network,the results show that the DBN recognition network has more obvious recognition performance advantages compared with the traditional neural network,the recognition accuracy is higher,more stable and the calculation cost is lower.Then in order to realize the prediction of the remaining life of the axle,this paper proposes a DBNnn network prediction method.In the prediction process,this paper mainly explores the effects of different input layers,hidden layer nodes,iteration times,training set,and test set samples on data reconstruction,network prediction error,and network calculation time.Next,in order to explore the role of the DBN network in the prediction method of the DBNnn network proposed in this paper,the same data is input into the nn prediction network with the DBN network removed,and compared with the prediction result of the DBNnn network.The data reconstruction error is smaller,the prediction result output by the network is more accurate,the prediction error is smaller,and the stability of the network is better.