Controllable data augmentation and application of multiaxial fatigue experiments by fatigue conditional generative adversative network model

Multiaxial fatigue damage poses a significant threat to the reliability and safety of engineering components, while the scarcity of samples hinders the prediction performance of data-driven algorithms. To address this issue, a controllable fatigue conditional generative adversative network (FatCGAN) is proposed and validated with a series of 316L stainless steel data. With strain-controlled loading as conditional information and utilization of Wasserstein loss, the FatCGAN achieves efficient augmentation of stress response and fatigue life. The augmented samples show consistent distribution trends with experimental ones in qualitative and quantitative assessments. With the help of augmented samples, various data-driven models such as SVM, RF, XGBoost, ANN, CNN, and GRU have significantly improved prediction performance with reductions in RMSE by up to 46.2 %. The repeatability of the prediction results has also seen substantial improvement with the standard deviation reduced by up to 51.7 %. This approach offers a new solution for enhancing the training effects in small-sample scenarios.