Study on data augmentation for fatigue crack growth prediction model

Prediction techniques based on numerical analysis results of components containing cracks can contribute to the safety assessment of machinery and structures. The authors have been conducting research on surrogate modeling of fatigue crack propagation analysis using automatic mesh generation and the s-version finite element method (s-FEM). In previous studies, noise-based data augmentation was employed to improve prediction accuracy; however, data imbalance remained a challenge. To address this, we proposed a relative frequency equalized data augmentation method, which is expected to mitigate the imbalance problem. This study investigates the applicability of this method for improving the prediction accuracy of learning models for both single and multiple cracks, focusing on the format of input and output data. At 10,000 training iterations, the single-crack model showed results close to the ground truth, and the multiple-crack model showed a tendency to approach it as well. The results demonstrate that unifying the input and output formats enables the use of a common model architecture, allowing effective application of the relative frequency equalized data augmentation method.