A generic framework for application of machine learning in acoustic emission-based damage identification

Advanced non-destructive monitoring scheme is necessary for modern-day lightweight composite structures used in aerospace industry, due to their susceptibility to barely visible damages from minor impact loads. Acoustic emission (AE) based monitoring of these structures has received significant attention in the past few years primarily due to their possibility of use in operating structures under service loads. However, localization and characterization of damages using AE is still an open area of research. The exploration of the space of signal features collected by a distributed sensor network and its reliable mapping to damage metrics (such as location, nature, intensity) is still far from conclusive. This problem becomes more critical for composite structures with complex features/geometry where the localized effects of discontinuity in geometric or mechanical properties do not make it appropriate to rely on simple signal features (such as time difference of arrival, peak amplitude, etc.) to identify damage. In this work, the AE signal features (which are spatially and temporally correlated) have been mapped to the damage properties empirically with a training dataset using metamodeling techniques. This is used in the online monitoring phase to infer the probabilistic description of the acoustic emission source within a hierarchical Bayesian inference framework. The methodology is tested on a carbon fibre composite panel with stiffeners that is subjected to impact and dynamic fatigue loading. The study presents a generalized machine learning-based automated AE damage detection methodology which both localizes and characterizes damage under varying operational loads