Unsupervised Consensus Clustering of Acoustic Emission Time-Series for Robust Damage Sequence Estimation in Composites

This paper suggests a new approach for unsupervised pattern recognition in acoustic emission (AE) time-series issued from composite materials. The originality holds in the development of a clustering ensemble method able to emphasize sudden growths of damages in composites under solicitations. The method combines multiple partitions issued from different parameterizations, initial conditions, and algorithms. A first stage automatically selects multifarious subsets of features based on the entropy of sequences of damages detected by clustering. A polygonal representation of the sequences is suggested to emphasize the kinetics of fracture events. The second stage allows estimating the optimal number of clusters necessary to represent the structure of the AE data stream. The data structure is estimated by consensus clustering with bootstrap ensembles, which allows estimating the uncertainty envelopes of each cluster and giving access to an interval of cumulated loading thresholds necessary to activate a particular damage. A qualitative evaluation phase is proposed on simulated data sets to statistically assess and underline both the robustness and accuracy of the proposed clustering fusion method, comparing $K$ -means, Gustafson-Kessel algorithm, and Hidden Markov models. An application is then presented for the detection of early signs of failure in high-performance carbon fiber-reinforced thermoset matrix composites dedicated to severe operating conditions. Despite the complexity of the configuration (ring-shaped specimens and high emissivity), it is demonstrated that the method emphasizes damage onsets and kinetics (fiber tow breakage, hoop splitting, and delamination) within the unevenly spaced AE time-series recorded during loading.