Damage Classification And Performance Prediction Of Fiber-reinforced Composite Materials Based On Acoustic Emission Multi-feature Statistical Analysis

Fiber-reinforced composite materials have become the first choice for a variety of engineering structural materials due to their excellent properties such as light weight,high strength,easy formation and strong designability.The harsh service environment and complex load conditions have brought huge challenges to the long-term stable service of composite structures.Therefore,it is necessary to develop non-destructive damage detection techniques and structural health monitoring techniques,among which acoustic emission(AE)technique shows great development potential.The damage diagnostics and prognostics of composite materials based on AE technique mostly rely on complex signal processing.Therefore,there are still many challenges in the current research on the AE-based damage identification.Feature selection methods for streaming AE data are still lacking,and the research on the AE-based damage prognostics is also relatively scarce.Based on this,research on AE-based damage classification and performance prediction of composite materials is carried out for the typical structural forms of composite materials,i.e.,laminated structures and adhesively-bonded structures.The main contents and conclusions of this thesis are as follows:(1)The corresponding relationships between AE response characteristics and damage mechanisms in adhesively-bonded composite specimens are established.The AE characteristics of different artificially-excited acoustic sources are experimentally investigated,and the accuracy of linear positioning in specimen-scale composites is verified.AE features are screened based on the principal component analysis,the Laplacian score and the correlation coefficient.The kmeans++ and CFSFDP clustering algorithms are introduced so as to realize damage mode identification of composite double-lap joints.The influence characteristics of the number of clusters,the metric of spatial similarity and the cutoff coefficient on the identification results of the CFSFDP algorithm are obtained.(2)An AE signal representation model is established based on the wavelet packet decomposition.The cosine similarity and the k-means++ clustering algorithm are employed to achieve damage mode identification of composite laminates based on the proportion of wavelet packet energy.The effects of hygrothermal aging on the AE response characteristics of damage mechanisms and the damage source localization are obtained.The failure process of the laminate is analyzed based on the cumulative wavelet packet energy of the signal in each frequency band.The influence of the wavelet basis function selection on damage mode identification and failure process analyis based on this signal representation model is explored.(3)Feature selection methods for dynamic data streams in adhesively-bonded composite specimens are developed.The Laplace score and the multi-cluster feature selection method are introduced and developed into feature selection mothods suitable for dynamic data streams.The real-time evaluation results of the conventional features and wavelet packet energy proportions of signals generated by composite single-lap specimens under tensile loads are analyzed.The damage characterization functions of the conventional features and wavelet packet decomposition features are obtained,including damage mode identification,singular signal detection and damage process characterization.(4)Models for predicting the remaining load-bearing capacity of laminated specimens are established based on dynamic data streams.AE feature evaluation is accomplished based on the RRelief F algorithm and the neighborhood component analysis.Based on the convolutional neural network,the remaining load-bearing capacity of the end-notched composite laminate under threepoint-bending load is predicted.The influence characteristics of the length of the time window in the model on the prediction results are explored.The Mel-frequency cepstral analysis is modified and the features sensitive to damage process are extracted.The cumulative moving average method is employed to establish the association relationship between the Mel-frequency cepstrum coefficients and the load or displacement of the laminate,which simplifies the prediction model of the remaining load-bearing capacity of laminates based on the convolutional neural network.