Generative Adversarial Networks (GANs)-based Timeseries Data Augmentation to Overcome Data Scarcity in the Context of Material Testing

Abstract

Data scarcity is one of the major challenges in applying machine learning (ML), as acquiring training data on the material behaviour is often difficult due to high costs and time constraints. To address this challenge, this study reports the development of a robust data augmentation method based on Generative Adversarial Networks (GANs). Firstly, the hyperparameters of the GANs model are optimised using Bayesian Optimisation (BO) by minimising the Fréchet inception distance (FID). Secondly, the quality of the generated data and its similarity to experimental datasets are evaluated using A-basis material allowables to determine the acceptability of the synthetic data. The accepted data is subsequently used to train a Gaussian Process Regression (GPR) model for predicting the strain-rate effect. Results show that the GPR model trained on a dataset of 1,800 data points achieved a lower negative log predictive density (NLPD) score than the model trained only on the experimental dataset. Furthermore, a hybrid approach is introduced to integrate the GPR model with a Cowper-Symonds model, thereby further improving prediction accuracy. Ultimately, this framework establishes a robust, statistically validated method for synthetic data generation and for overcoming data limitations in the accurate modelling of the strain-rate-dependent properties of carbon/epoxy composites.