Waveform-based clustering of fatigue cracking acoustic signatures

Identifying signatures of fatigue crack growth and detecting early warnings in service in Non-Destructive Testing applications remain major challenges for research and industry alike. To address these issues, we propose a waveform-based signal processing technique applied to Acoustic Emission (AE) data. This study explores how a fully unsupervised, data-driven approach can cluster acoustic signals based on their physical emission mechanisms associated with fatigue cracking. Using AE datasets recorded during fatigue crack growth tests on compact tension specimens made of aluminum or steel alloys, we demonstrate that hierarchical clustering of acoustic multiplets - signatures of a unique source - is consistent with the phenomenological classification of multiplets previously established distinguishing multiplets generated by local contacts between crack surfaces from those associated with fatigue crack growth at crack tip. Here we show that we can assign physical emission mechanisms to each multiplet solely based on acoustic information through waveform-based dissimilarity measures - unlike standard AE signal processing techniques. Furthermore, by taking advantage of the multiple levels of analysis offered by hierarchical clustering, we significantly improve multiplets detection by reducing dependency on user-parameters. This improvement opens the possibility of developing a more automated Non-Destructive Testing architecture for fatigue cracking signatures detection, thus reducing computational complexity. • Fatigue crack propagation and rubbing signatures are unsupervisedly classified. • Signal waveforms similarity measures capture information about physical sources. • Signals clustering based on hand-made descriptors are not sufficiently efficient. • Hierarchical clustering of multiplets centroids corrects signatures detection errors. • Fatigue signatures detection method robustness and automation is improved.