Damage identification in a tubular composite/metal joint through chronology-based robust clustering of acoustic emissions

The aim of this work is to determine the first failure mode of a tubular composite/metal joint, when tested in tension. PWAS are bonded to both the composite and metal parts of the joint for detection of acoustic emissions (AE) during the tensile test. The generation of AE hits is correlated with strain as measured by surface mounted strain gauges and digital image correlation (DIC) during the test. AE data analysis is completed using an unsupervised clustering approach applied to the AE streaming data in order to identify the damage mechanisms. The approach makes use of the Gustafson-Kessel clustering algorithm to identify the clusters with arbitrary shape in the feature space. In the proposed method, different subsets of features are considered, in place of a unique subset as in standard approaches. Each subset provides one clustering result (a partition). An unsupervised information fusion process is then performed to get a representative partition, taking multiple partitions into account. The approach makes use of bootstrapped ensembles to select the number of clusters which allows to both maximise robustness of the results according to the clustering parameterisation and evaluate the uncertainty on pattern recognition results. Subsets of features are optimised to emphasise the chronology about the onsets of AE sources and their evolution. Application of this pattern recognition chain on the AE streaming provides insights on the damage process involved in composite and metal parts of the joint.