A variational bayesian clustering approach to acoustic emission interpretation including soft labels

International audienceWe investigate Gaussian Mixture Models (GMM) with uncertain parameters to evaluate whether this model can help in interpreting acoustic emission data used in non-destructive testing. This model, called VBGMM (variational Bayesian GMM) allows the end-user to automatically determine the number of clusters which makes it relevant for this type of application where clusters are related to damages. In this work, we modify the training procedure to include prior knowledge about clusters. Experiments are made on a recently published bench-mark, ORION-AE, that aims at estimating the tightening levels in a bolted structure under vibrations. Preliminary results of the VBGMM with soft priors (VBGMM-SOFT) show good improvement over the standard VBGMM