Acoustic Emission Signals Classification For Structural Health Monitoring Based On Deep Learning

Acoustic emission technology has the advantages of being sensitive to damage,being less affected by geometric structure,and having a long detection distance,so it is suitable for online monitoring of civil engineering structures.Using acoustic emission to carry out structural health monitoring requires the quantitative analysis of acoustic emission signals to achieve the purpose of damage identification,location and evaluation.Acoustic emission signals contain dynamic microscopic damage information of materials and signals with different damage mechanisms usually have different instantaneous frequency components.Aiming at the demand of quantitative analysis of acoustic emission signals,this paper proposes a classification method of acoustic emission signals based on time-frequency analysis and deep learning.The wavelet transform is used to study the time-frequency energy distribution of acoustic emission signals,and then the deep learning model of convolutional neural network is established to automatically extract the time-frequency characteristics of signals,and the signals with different damage mechanisms are classified,to realize the structural damage identification and quantitative evaluation.The main research contents of this paper include:(1)The research ideas of AE signal classification based on time-frequency analysis and deep learning are proposed,and the relevant theoretical basis and research status are summarized.The basic theory of wavelet transform is introduced,including continuous wavelet transform,synchrosqueezed wavelet transform,empirical wavelet transform,and the advantages and disadvantages of each wavelet transform are compared.The basic principle,network structure,training method and optimization way of convolutional neural network are introduced.(2)Aiming at the acoustic emission monitoring problem of rail crack,the signal classification of supervised learning was studied.A signal classification method based on synchronous extrusion wavelet transform and multi-branch convolutional neural network was proposed to identify three typical signals of rail operation noise,wheel impact and crack growth.Because three kinds of acoustic emission signals are similar,synchronous extrusion wavelet transform can reveal the time-frequency energy distribution of the signals more clearly.Then a multi-branch convolutional neural network was established to extract the energy distribution features of different frequency bands to improve the accuracy of classification.Then,through transfer learning and Bayesian super parameter optimization,the crack identification under super noise is realized.Finally,the validity of the proposed method is verified by a dataset consisting of 5200 AE signals obtained from field rail crack experiments and laboratory fatigue experiments.(3)In view of the high strength bolt connection state problem of monitoring acoustic emission signal classification of unsupervised learning study,put forward a kind of experience based on wavelet transform and convolution neural network-signal classification method,gaussian mixture model is used for the static friction force of high strength bolt and relative slip,screw elastic deformation,plastic deformation,and fracture of three stages of state recognition.Because the three stages of acoustic emission signal have microcosmic mechanism crossover,the empirical wavelet transform can be used to adaptively divide the characteristic frequency band of the signal.Then a convolutional neural network model was established to extract time-frequency features and reduce dimensionality.Then,the characteristic vectors were input into the Gaussian mixture model cluster,and combined with the analysis of material failure mechanism,the typical acoustic emission signals of high-strength bolts in different stress stages were recognized.Finally,the validity of the proposed method is verified by the data set of 7282 AE signals obtained from the shear failure experiments of high-strength bolted connection members in the laboratory.