Pattern Recognition Of Acoustic Emission Signals Based On Deep Learning For CFRP Composite Damage

Carbon fiber reinforced plastic(CFRP),recognized for its high strength,lightweight,and resistance to corrosion,has wide-ranging applications in fields such as aerospace,civil engineering,transportation,maritime,chemical equipment,and medical devices.However,during its service life,CFRP may be influenced by various physical conditions,resulting in diverse forms of damage.Some of these damages,particularly severe and irreversible ones,can lead to component failure,thereby posing significant safety risks and incurring substantial economic losses.Hence,conducting research on damage identification in carbon fiber composites is of paramount importance.This study undertakes a comprehensive examination of intricate recognition of unidirectional tensile acoustic emission signals within diverse environmental scenarios,encompassing mechanical,mechanical-thermal,and mechanical-thermal-hygro environments.The focal point of our investigation is Carbon Fiber Reinforced Plastic(CFRP)material,and we employ acoustic emission technology as the primary research tool,with deep learning technology as the analytical methodology for processing acoustic emission data.Within this research,we introduce an innovative,deep learning-based approach for the identification of acoustic emission signals indicative of damage in CFRP composites subjected to varying environmental conditions.Firstly,an experimental setup for unidirectional tensile acoustic emission of CFRP materials was established to simulate three typical damage modes: fiber breakage,matrix cracking,and delamination.Consequently,three distinct datasets capturing typical damage acoustic emission were acquired.For the subsequent supervised learning,two types of input data were considered:acoustic emission time series data and frequency domain sequence data following fast Fourier transform(FFT).Eight deep learning neural networks were constructed,trained,validated,and tested using both raw waveform time series data and FFT-derived frequency domain sequence data.These networks include FCN,Res Net,XRes Net,LSTM＿FCN,Inception Time,Xception Time,m WDN,and LSTM.The findings demonstrate that utilizing time series data as input for deep learning-based damage classification yields superior accuracy compared to using frequency domain sequence data.Moreover,this approach exhibits potential for effectively classifying unbalanced datasets.Secondly,to address the challenge of recognizing acoustic emission signals from CFRP composite damage in mechanical-thermal and mechanical-thermal-hygro environments,experimental setups for acoustic emission testing of CFRP composites were established under mechanical-thermal and mechanical-thermal-hygro environments.Acoustic emission datasets were subsequently generated,and an innovative hierarchical clustering labeling approach for acoustic emission feature parameters was introduced to tackle the issue of unavailable damage category labels.The labeling dataset was subjected to training,validation,and testing,including a comparative analysis between Res Net50 and VGG11 networks using 2D images obtained after continuous wavelet transform and Mayer transform as inputs.Additionally,Inception Time,Xception Time,and Res Net end-to-end modeling networks were used with time series data as inputs.The results reveal that the end-to-end deep learning model,utilizing time series data for classification,achieves higher accuracy when compared to deep learning models using 2D images as input.This approach not only enhances classification accuracy but also reduces the computational load associated with 2D feature extraction.Such computational efficiency is particularly valuable for damage recognition in composite materials subjected to varying layups,temperatures,and humidity levels in practical engineering scenarios.Finally,the trained supervised deep learning model is deployed to analyze the acoustic emission data of CFRP composite damage in mechanical-thermal and mechanical-thermalhygro environments.Initially,labeled datasets for acoustic emission in these mechanicalthermal and mechanical-thermal-hygro environments are constructed based on hierarchical clustering of feature covariates.Acoustic emission time-series data are utilized as inputs to effectively identify the type of CFRP composite damage,thereby further validating the model’s generalization ability.Secondly,unlabeled datasets of acoustic emission data from mechanicalthermal and mechanical-thermal-hygro environments are created.The knowledge and capability of the trained model are applied to the classification and identification task of this unlabeled data,enabling the identification of CFRP composite damage.These research results provide new ideas and methods for the deep learning-based recognition of acoustic emission signals associated with CFRP composite damage.