Non-destructive detection of interfacial defects in SCCS using percussion-based machine learning with SHAP interpretability

Additionally, discerning various defects can provide crucial insights into overall structural health and aid in determining appropriate maintenance strategies. Inspired by the classification capability of machine learning-based data processing, this study introduces a novel approach to identifying abnormal interface conditions, leveraging the percussion method in conjunction with a support vector machine (SVC). Initially, the distinct acoustic signals are acquired using a high-frequency microphone and prepared as the dataset. Subsequently, Mel Frequency Cepstrum Coefficients (MFCC) combined with Chroma Vector (CV) are extracted as fused features in the model. The validation experiment results demonstrate the excellent applicability of this method in classifying multiple types of damage, achieving a remarkable prediction accuracy of up to 96.85%. Furthermore, the integration of SHapley Additive exPlanation (SHAP) enables quantifying the contribution of features to the model prediction process, thereby further enhancing model interpretability. Importantly, systematic comparison works verify the superlative identification ability through the fused features and SVC. The proposed method provides a high-precise and non-destructive tool for addressing the multi-class recognition of interfacial damage in SCCS.