Machine Learning approaches to damage detection in composite structures combining experimental and simulation domains

Abstract

Composite materials are widely used across major industries such as the automotive, aerospace and wind power, due to their excellent mechanical properties. A strong effort is thus put into developing innovative damage detection methodologies, for which Non-Destructive Testing (NDT) techniques can play a vital role as advanced measurement methods. One such technique is Laser Doppler Vibrometry which allows to accurately measure high-frequency vibration behavior with dense grids of points, without mass loading the structure. For data analysis, Machine Learning (ML) techniques have achieved high success on a number of structural applications, and can be leveraged to build automated and reliable damage classifiers. In this work, three methodologies have been developed by combining Laser Doppler Vibrometer (LDV) measurements with ML approaches, for the task of detecting damages on a carbon-fiber reinforced polymer (CFRP) plate. Each damage detection methodology requires pre- and post-processing steps, which were optimized with Bayesian Optimization. Principal Component Analysis (PCA) was also explored to reduce the dimensionality of the data, before classification. Moreover, making use of Finite Element Analysis (FEA), simulation data was generated with the ability of characterizing the high-frequency dynamic behavior of ply-based composite plates. The simulation data enriched the damage detection methodologies in a Transfer Learning (TL) framework. Results are presented for each damage detection methodology, alongside with a comparative overview of the advantages and disadvantages of each method.