Open Access
Subscription or Fee Access
Clustering Acoustic Emission Time-series Using Unsupervised-shapelets
Abstract
The sensitivity of AE sensors makes the AE technique very interesting for detecting damages at various scales as encountered in complex heterogeneous materials made of various constituents with different damage kinetics. However, there are unknowns behind the generation process of AE signals, as well as behind the modification of those signals along the propagation path until the sensors. Those make the interpretation of AE signals a difficult task in terms of pattern recognition. Even though some damage families are expected to occur for a given material under specific loading, there is a lack of knowledge to bridge the gap between AE signals collected on sensors with the related AE source. This accounts for the use of unsupervised learning when one is interested to discover relationships between AE data. When using clustering for unsupervised learning, the output is a set of clusters representing a set of labels assigned to each datum and forming a partition. In AE work, clusters are estimated using features, computed from AE signals. Feature extraction has become a must-to-do step due to the use of particular clustering methods and to the fact that AE streaming has not been particularly analysed for damage sources identification. The contribution of this paper is a new methodology for AE signal clustering which does not require feature extraction as usually done. It works directly on raw AE signals derives from several tools of the literature, namely unsupervised shapelets, anomaly detection and statistical modelling by Autoregressive Weakly Hidden Markov Models. When used in inference on unknown signals or streaming, those models allow to generate a set of novelty scores which are then processed by a chronology-based clustering algorithm to get a partition, accompanied by the uncertainty around clusters and a quantification of the robustness of the results obtained.
DOI
10.12783/shm2019/32263
10.12783/shm2019/32263