Physics-Enhanced Damage Classification of Sparse Datasets Using Transfer Learning

Abstract

High-rate, high-acceleration dynamic events produce especially limited and sparse data for two main reasons: high-acceleration loadings can destroy the test article, and the required laboratory equipment is typically expensive and complicated to operate. In many cases, these limitations prevent researchers from collecting additional data, driving the need for machine learning algorithms that utilize small datasets. Despite deep learning's preference for thousands or millions of training examples, the dataset considered in this work contains only six independent examples. Finite element analysis software simulates the dynamic response of an electronic structure, supplementing this small dataset with additional training examples. A hybrid deep learning model first learns the dynamic response of the simulated structure and is then adapted to predict the actual electronic structure's damage levels. This work shows that physics-enhanced transfer learning improves structural damage classification accuracy (𑃠= 0.0879).