Surface crack propagation prediction under low cycle fatigue test based on machine learning with small data size

Abstract

Fatigue crack initiation and propagation in turbine discs can lead to catastrophic failures in aeroengines, compromising structural integrity and highlighting the critical need for effective crack detection and propagation prediction. This study proposes an adaptive parameter adjusting and online forecasting DLinear model, a data-driven model aimed at predicting surface fatigue crack propagation at turbine disc bolt holes utilizing limited experimental datasets. The model’s performance is evaluated through prediction error analysis and compared with existing machine learning and long-term time-series forecasting methods. Results demonstrate high predictive accuracy, with mean absolute percentage errors of 1.75%, 1.64%, 1.71%, 1.04%, 1.40%, and 1.41% for six bolt holes. The proposed model significantly outperforms ARIMA, RNN, and Informer, achieving average MAPE reductions of 5.36%, 17.02%, and 30.51%, respectively. This model enables reliable crack propagation prediction with small datasets, offering a practical solution for real-time monitoring and forecasting during component-level engineering tests.