Multifidelity modelling of damaged aerospace structures: A transfer learning based approach

A critical part of structural health monitoring is accurate detection of damages in the structure. Simulation based optimization for damage detection and identification requires numerous iterations with expensive simulated models which is impractical for real-time assessment. This thesis proposes a multi-fidelity Reduced Order Modeling (ROM) method based on transfer learning, to develop an emulator for fast online data generation. The online data are obtained using a machine learning algorithm trained with an offline database which is the result of a transfer learning method. In this method, finite element simulations with different fidelities are combined using Reduced Order models and manifold alignment to determine a common space where the accuracy of high fidelity simulations is fused with the spatial resolution of the low fidelity simulations. Then, it is proposed a comparison between the machine learning algorithms of gaussian regression, cokriging, regression trees and Self Organizing Maps (SOMs) to determine the most suited for cut damage detection in composite plates for aerospace application.