Machine-learning-based crack mode classification in plain concrete considering acoustic emission wave propagation effects

This study focuses on crack mode classification in concrete using acoustic emission (AE) considering the impacts of wave propagation and AE signal acquisition conditions. It was realized by performing symmetric and asymmetric four-point bending (FPB) tests on notched beams under crack mouth opening displacement (CMOD) control. Additionally, pull-out tests on grouted anchors in concrete samples were performed. The crack modes have been identified by cluster analysis on features extracted from AE waveforms. The impact of AE wave propagation and signal acquisition conditions such as wave propagation distance, wave propagation angle, damage progression, and AE sensor characteristics were investigated through a supplementary experimental campaign which indicated that AE wave propagation distance and AE sensor response can significantly affect waveform characteristics. It also helped to identify the features that were genuine indicators of crack modes (source mechanisms). The wave propagation distance was modeled using Support Vector Regression (SVR), optimized by the Bayesian Optimization Algorithm (BOA), based on pencil-lead break test results to modify the effect of wave propagation distance. An unsupervised machine-learning algorithm, Gaussian mixture model (GMM), was used for crack classification. Cluster analysis was performed separately on the AE hits from each sensor as AE sensors exhibited varying responses. Finally, the AE sources were classified into shear, tensile, and mixed modes based on a voting system. A comparison of the results was drawn between cluster analysis using modified and raw features. The applied approach showed potential for practical application for an efficient, qualitative crack mode classification.