High cycle fatigue S-N curve prediction of steels based on transfer learning guided long short term memory network

The stress-life (S-N) curve is a fundamental aspect in fatigue analysis. However, fatigue testing using S-N curve is very costly and time-consuming. To solve this, a novel method to predict S-N curve is proposed combining the long short-term memory network (LSTM) and transfer learning. A transfer LSTM framework (TR-LSTM) was developed, wherein the reversed torsion S-N curves prediction of low alloy steels was transferred from rotating bending S-N curves. The prediction results for twelve steel grades prove the rationality of the framework. The generality of the framework with respect to different data amount and model parameters was further investigated. Additionally, the model was also successfully extended for the curve prediction of very high cycle fatigue. This proposed prediction framework can significantly reduce the cost of fatigue property evaluation and realize the conversion among fatigue curves with different test costs.