Augmenting Fatigue Datasets for Improved Multiaxial Fatigue Strength Prediction with Neural Networks

Accurate fatigue prediction is essential for ensuring the reliability and durability of engineering systems. Suitable predictive performance was achieved by artificial neural networks trained on the FatLim dataset; however, further improvements are needed due to its small sample size. This study explored the impact of dataset augmentation on model performance by exemplarily expanding the FatLim dataset from 294 to 1732 cases and comparing results against the original dataset. The dataset was augmented by generating additional uniaxial stress scenarios and applying tensor transformations to simulate varied stress orientations. Neural network models were trained separately on the original and expanded datasets, and their predictive performance was evaluated. The results demonstrate that the model trained on the augmented dataset achieved better accuracy, with the mean prediction error decreasing from 0.95% to 0.31% when tested on the original dataset, confirming the effectiveness of dataset expansion in improving fatigue prediction. This research underscores the potential of data augmentation techniques to enhance machine learning models for fatigue analysis.