Crack pattern identification in cementitious materials based on acoustic emission and machine learning

Abstract

The cracking patterns in cementitious materials before failure are closely related to acoustic emission (AE) monitoring signals. Traditional rise angle-average frequency analysis methods rely on empirical judgment for boundary line determination, lacking effective automated recognition methods to distinguish between tensile, shear, and mixed crack types. This paper presents an unsupervised crack-type identification model based on ensembled clustering with support vector machine (SVM) and a supervised crack-type classification model using graph neural networks (GNN). The study achieves automatic annotation of four crack signal patterns by ensembling the results of k-means++ and GMMSEQ clustering and combining them with SVM. Subsequently, with labeled data obtained, the study encodes AE signals into graphs and uses a GNN for automatic crack-type classification. The effectiveness of these methods is validated on AE test data from cementitious materials. The findings show that the mutation values of AE event interval functions can accurately pinpoint the material fracture moment, and the AE spectrum’s mid-frequency band is associated with cement–river sand interface damage. Different compositions and loading methods in cementitious materials complicate the association between signals and damage. The ensembled clustering and SVM methods can automatically distinguish different types of cracks and provide quantifiable boundary lines. With horizontal visibility graph (HVG) encoding, the supervised GraphSAGE model accurately identifies various crack types, achieving an average accuracy of up to 97.5% on the test dataset. The proposed AE time–frequency domain characterization method offers new analytical tools for understanding the damage evolution process in cementitious materials, promoting the application of machine learning in the recognition of AE signal patterns.