Identification Method Of Acoustic Emission Signal Based On Time Series Feature

As a nondestructive testing method,acoustic emission signals detection method has obvious characteristic frequency and high signal-to-noise ratio,and can be used to identify early fault signal.Furthermore,acoustic emission signal is a kind of stress wave signal with high frequency,which is generated by fault structure itself.Therefore,there is no necessary for outside signal sources and no request for geometric shape and operating state of the detected target.It is appropriate for online fault detection.The nondeterminacy,unpredictability and noise sensitivity of acoustic emission signal cause that the existing signal processing methods cannot effectively extract and identify the characteristics of signals.Time series analysis is a useful quantitative prediction method.The statistical model of target signal is able to reflect the law of development of the object and the regression parameter of the model contain the current state characteristics of the object and the average trend of dynamic variation.As an effective method of signal processing which can reflect the characteristics of the object,time series analysis is gradually applied in the field of fault detection in recent year.This paper takes the railway vehicle axle as research objects,and studies the time series analysis based feature extraction method and classification method of acoustic emission signals.There are two method of fault detection based on time series analysis proposed in this paper for acoustic emission signal.The first one is a frequency domain method which combines the energy feature with AR model spectral estimation.The second method analyzes the regression parameters directly by using deep learning model.The detail of the study in this paper is as follow:(1)By taking the fracture process of axle as the starting point,the principle of acoustic emission signal is analyzed.Then,the constitution of the collected signal is analyzed after data acquisition step of the acoustic emission signals during the process of axle fracture is introduced.And a simple classification of collected signal is presented based on different conditions of axle.(2)Because the acoustic emission signal generated by fracture is non-stationary and cannot be used for time series analysis directly,empirical mode decomposition is introduced in this paper to pretreat the collected signals.empirical mode decomposition can transforms the non-stationary signal to stationary one and separates different signals in different bands in the meanwhile.This kind of ability can reduce the complexity of feature extraction process and improve the characterization of features.(3)An AR model spectral estimation based frequency domain feature analysis method is proposed in this paper.In this study,AR model spectral estimation is used to replace the nonparametric method such as wavelet analysis and fast fourier transformation in the process of time-frequency transformation.AR model spectral estimation obtains power spectrum by self-adaptive analysis method rather than using primary function or time window.On the other hand,energy features such as power distribution and power spectrum entropy are used for feature extraction instead of widely used manual recognition method.This kind of advance can improve the descriptive ability of characteristic and the accuracy of identification.(4)A regression parameter analysis method by using deep learning model is proposed.The parameters of regression model contain a lot status information of the object,which cannot be extracted efficiently by traditional feature extraction methods.In this paper,Deep Belief Network is used to learn the characteristic of regression parameters.The characteristic information can be sum up and conclude automatically and the features obtained by Deep Belief Network can be more intact.Comparing with other traditional machine learning methods and spectrum based methods,this study has higher accuracy on signal classification but takes more time to learn features.