Data Augmentation of High-Rate Dynamic Testing via a Physics-Informed GAN Approach

Abstract

High-rate impact tests are essential in the prediction of component/system behavior subject to unplanned mechanical impacts which may lead to several damages. However, significant challenges exist when identifying damages in such events, considering its complexity, these diagnostic challenges inspire the employment of data driven approaches as feasible solution for such problems. However, most deep machine learning techniques require big amount of data to support an effective training to reach an accurate result, while the data collected from each test is extremely limited yet performing multiple tests to collect data is oftentimes unrealistically expensive. Therefore, data augmentation is very important to enhance the learning quality. Generative Adversarial Network (GAN) is a deep learning algorithm able to generate synthetic data under a recorded testing environment. However, a GAN uses random input as seeds to generate adversarial models, and with sufficient training, it may produce synthetic data with good quality. This paper proposes a hybrid approach which employs the output from an oversimplified FE model as the seed to drive a GAN generator, such that a drastic amount of computation is saved, and the GAN training will converge faster and more accurately than using just the random noise seeds. Variational Autoencoder (VAE) is combined with the approach to reduce the data dimension and the extracted features are classified via a Support Vector Machine (SVM). Results show that using the proposed physics-informed approach will improve the accuracy of the damage classifier and reduce the classification uncertainty, compared to using the original small dataset without augmentation.