Joint Optimization of the Number of Clusters and Their Parameters in Acoustic Emission Clustering

Acoustic Emission (AE) is a non-destructive technique suitable for monitoring structural damages in complex systems. Interpretation of AE data is generally made by an unsupervised learning approach called clustering. A clustering method divides a dataset into different groups called clusters, which are expected to have a physical interpretation in terms of damages. The set of groups, called partition, is then evaluated through an external criterion. Since the number of clusters is a priori unknown, the standard approach for AE clustering consists in running a clustering method for different number of clusters and then to select the partition with the best value of the external criterion. The external criterion can vary across publications and aims at evidencing “natural” clusters. According to the criterion, the selected partition and its interpretation can be subjected to variability. In this publication, we explore a clustering method in which the parameters of clusters are optimised jointly with the number of clusters through an objective function and an optimisation procedure designed for that purpose. To our knowledge, it is the first method of this class in AE literature. Experimental results concern real data from tightening test.