Augmentation method of fatigue data of welded structures based on physics-informed CTGAN

Variable amplitude loading is frequently applied to welded structures in practical engineering, and fatigue failure is a prevalent problem. In recent years, machine learning is a useful technique for predicting fatigue life. However, it is challenging to acquire a sufficient number of reliable training samples for fatigue tests under variable amplitude loading. The machine learning models' accuracy and generalization capabilities are impacted by this. This work introduces a novel data augmentation approach utilizing physics-informed Generative Adversarial Networks (GAN). Data augmentation is accomplished by incorporating the traditional damage model - Ye model  as constraints within the loss function of the Conditional Tabular GAN (CTGAN). The method combines physical laws of damage with CTGAN, which makes generated fatigue data conform to physical characteristics under two-step loading. Then the impact of generated data on model performance is evaluated on four machine learning models and compared to traditional damage models. The experimental results show that generated fatigue data helps machine learning models to get better prediction results compared with traditional models and unaugmented machine learning models, which significantly enhancing the precision of fatigue life predictions.