An unsupervised pattern recognition approach for AE data originating from fatigue tests on polymer-composite materials

International audienceAcoustic Emission (AE) technique is gaining more and more interest for structural health monitoring (SHM) in polymer-composite materials. Recent literature has shown that using appropriate pattern recognition techniques (PRT), the identification of the natural clusters of acoustic emission data can be obtained. Despite these recent and valuable advances and to achieve health assessment of composite materials, the scientific community faces two major challenges: (i) develop real-time approaches and (ii) propose clustering approaches able to process in in-service-like situation, i.e. in case of high AE activity generated simultaneously from many damage sources in material, from damage progression and cumulated damage and from noise.This work investigates acoustic emission generated during tension fatigue tests carried out on a carbon fiber reinforced polymer (CFRP) composite specimen. Since fatigue data processing, especially noise reduction remains a big challenge in AE data analysis; a simple null-value filtering and a noise modeling have been proposed in the present work to tackle this problem. A Davies-Bouldin-index-based progressive feature selection has been implemented to reduce high dimensional fatigue dataset. A classifier offline-learned from quasi-static data is then used to classify the processed data to different AE sources. An adaptation has been studied to enable the classifier to generate new class, i.e. AE source, for unidentified AE events. With efficient proposed noise removal and automatic separation of AE events, the results of this work provide an insight into fatigue damage development in composites and then ability to health assessment which is necessary for residual life prediction