A machine learning-based method for wide-band random fatigue damage calculation

Objectives To address the limited accuracy of conventional frequency-domain fatigue analysis methods under wide-band random loading, a machine learning-based method for calculating wide-band random fatigue damage in the frequency domain is proposed.

Methods First, a dataset linking a wide range of spectral moment parameters and spectral width parameters to the corresponding fatigue damage obtained by rainflow counting is constructed from 11 types of parameterized stress spectra. A neural network-based machine learning model is then developed for model training, hidden layer architecture optimization and performance evaluation. Second, numerical simulations are performed using an independently generated set of new stress spectra to systematically assess the prediction accuracy and generalization capability of the model. Third, SHAP (SHapley Additive exPlanations) analysis is conducted to interpret the internal learning mechanism of the model. Finally, the applicability of the model under extreme spectral width conditions and in real-world engineering scenarios is evaluated using cases with extreme spectral widths and actual engineering simulation data.

Results The results for all validation cases demonstrate that the proposed method generally outperforms conventional frequency-domain methods in terms of prediction accuracy. In most cases, the maximum relative error is within 5%, while the mean absolute percentage error remains below 1%. The average computational time per prediction is approximately 4.5 ms, demonstrating high computational efficiency while maintaining reliable fatigue damage predictions. SHAP-based interpretability analysis reveals that the fourth-order spectral moment serves as the dominant feature influencing the model predictions.

Conclusions The proposed ANN model effectively captures the complex nonlinear mapping between spectral moments and fatigue damage, overcoming the limitations of conventional frequency-domain methods, which rely on analytical formulations and therefore have difficulty adequately characterizing complex spectral features. Consequently, the proposed method enables efficient and accurate prediction of wide-band random fatigue damage.