Lightweight, interpretable acoustic emission diagnostics for multi-mode damage in bonded CFRP joints using consensus-guided machine learning

Adhesively bonded carbon fibre-reinforced polymer (CFRP) joints exhibit complex, multimode damage involving concurrent matrix cracking, debonding, delamination and fibre breakage, which challenges acoustic-emission (AE)-based diagnosis under industrial hit-triggered acquisition. This study is framed as a focused methodological investigation on a single-lap bonded CFRP joint system, rather than as a claim of universal damage-frequency thresholds. The proposed workflow uses an offline consensus procedure to derive structure-specific operational passbands from standard AE descriptors, then uses these fixed passbands to build compact band-energy and entropy features for interpretable, online-oriented classification. AE waveforms from three nominally identical tensile tests are first represented by 18 conventional descriptors. Multiple clustering families and random seeds are combined through a co-association matrix to obtain a robust four-cluster macro-partition. High-confidence samples within each macro-cluster are then used to define four operational passbands from central peak-frequency quantiles. These passbands are frozen and used to construct an extended 26-dimensional representation that augments the baseline descriptors with band-energy shares and entropy measures. A second consensus step in the extended feature space provides 15 operational AE reference states that describe single-dominant and spectrally mixed acquisition windows. These states are used to train an XGBoost surrogate evaluated on a fully held-out experiment. The extended representation improves test accuracy from 0.68 to 0.91 and increases macro-F1 and macro-AUROC to about 0.88 and 0.90, respectively. SHAP analysis indicates that the classifier relies primarily on peak frequency, passband energy and entropy descriptors. The results demonstrate a trade-off between offline robustness and online simplicity: the numerical passbands are system-dependent and require commissioning for new structures, whereas the deployed decision path remains lightweight after calibration.