A novel neural network model considering cyclic loading condition for low-cycle fatigue life prediction

Abstract

Existing data driven low-cycle fatigue (LCF) life prediction models exhibit limited attention to the heterogeneity of input features and relatively insufficient consideration of cyclic loading condition. This study introduces a novel approach based on a long short-term memory parallel hierarchical neural network (LSTM-PHNN). Firstly, the input features are classified into three categories based on their physical significance and distribution characteristics: elemental parameters, microstructural property parameters, and loading parameters. Next, waveform features over a complete cycle are reconstructed for different loading parameters to characterize the impact of time-series cyclic loading condition on fatigue life. Finally, three predictive models were developed: a fully connected neural network (FCNN), a parallel hierarchical neural network (PHNN), and the proposed LSTM-PHNN model. Comparative analysis was conducted using the small sample LCF dataset of 316L stainless steel. The results show that the proposed LSTM-PHNN model outperforms both the FCNN and PHNN models,with almost all predictive data falling within a scatter band of 1.5 times and the prediction accuracy on the test set reaching 0.954. The above demonstrate that the LSTM-PHNN model provides a highly accurate and robust method for predicting LCF life with a small amount of experimental data.