An Improved Physics‐Informed CTGAN Data Augmentation Method for Fatigue Life Prediction of Metal Materials Under Two‐Level Loading

Due to the complexity and stochasticity of fatigue processes under multilevel variable amplitude loading, traditional life prediction models often fail to meet the requirements of practical engineering applications. Data‐driven models based on intelligent computing techniques can effectively capture effects such as load sequence and load interaction, offering a promising approach to fatigue life prediction under multilevel loading. However, the destructive nature and long duration of fatigue experiments result in limited and scattered data, which limits the accuracy of data‐driven models for fatigue life prediction. In this work, an improved physics‐informed CTGAN method is presented for fatigue data augmentation, aiming to overcome the practical limitation of insufficient experimental samples in engineering applications. The LLR M–H model is incorporated as prior knowledge into the generator's loss function. Additionally, an S – N curve‐based screening strategy is employed to remove physically inconsistent samples. Experimental results indicate that the average RMSE of the four data‐driven models is reduced by 15.9%–32.4% following data augmentation, and the proposed method achieves a 25.2% improvement in prediction accuracy relative to the original physics‐informed CTGAN.