Physics-Enhanced Unsupervised Clustering for FRP-Concrete Debonding Using FEM–BEM Simulation-Guided Percussion Sound Reconstruction

Abstract

Percussion sound analysis is a promising non-destructive evaluation technique for detecting debonding between fiber-reinforced polymer (FRP) and concrete. However, conventional supervised learning approaches are constrained by their reliance on labeled training datasets and their limited capacity to reconcile the inherent variability and non-stationary characteristics of percussion sound signals, ultimately compromising model generalizability. To address these challenges, this study proposes an unsupervised clustering–based detection method guided by finite element–boundary element (FEM–BEM) simulations. First, FEM–BEM models simulate percussion sound signals under varying debonding conditions and percussion patterns, revealing common latent acoustic features strongly correlated with debonding across diverse percussion patterns. Guided by physics-derived insights, original percussion signals are reconstructed using wavelet packet decomposition, emphasizing debonding-related features while suppressing noise. Subsequently, unsupervised clustering is applied to the reconstructed signals to group percussion sounds without using labeled training data. The resulting clusters are then assigned semantic labels using a small set of reference signals from well-bonded regions. Experimental validation demonstrates that the method achieves an accuracy of 87.80%, outperforming supervised methods by 15.44%. The proposed method reduces the reliance on large amounts of labeled experimental data by using simulations to identify features sensitive to debonding, thereby guiding the detection algorithm to place greater emphasis on relevant frequency bands or features.