Reliable acoustic-emission-based damage clustering in composite adhesive joints with Integrated denoising and confidence metrics

The Acoustic Emission (AE) technique is a sensitive and high-resolution non-destructive method ideal for monitoring bonded composite joints. However, noise interference and unlabeled signals make damage identification challenging, especially when noise overlaps with useful frequencies and the dataset structure is unknown. This study proposes a conceptual and methodological framework combining noise separation and confidence evaluation to improve damage classification. AE events are collected from composite single-lap joints with varying adhesive lengths under tensile loading. Signals are first decomposed by the proposed noise separation methodology based on Variational Mode Decomposition (VMD) and Fast Independent Component Analysis (FastICA). Subsequently, the t-SNE-based Confidence Evaluation (tSCE) is developed as a reference for clustering analysis, which can provide the quantitative confidence metrics of individual AE signals and the proportion of dubious events prone to misclassification. Finally, the Gaussian Mixture Model (GMM) was applied for clustering damage patterns. The results demonstrate that integrating noise reduction with confidence evaluation notably increases the average clustering probability of the GMM from 95.8% to 98.1%. It also enhances the consistency of the damage evolution process under different adhesive lengths. Dubious points are more likely to correspond to mixed damage, and the AE confidence score can serve as a tool for early monitoring in SHM. This paper provides a reliable damage clustering framework and valuable insights for understanding failure modes in composites.