Multiaxial Fatigue Life Prediction of Metallic Specimens Using Deep Learning Algorithms

Accurately predicting fatigue life under multiaxial fatigue damage conditions is essential for ensuring the safety of critical components in service. However, due to the complexity of fatigue failure mechanisms, achieving accurate multiaxial fatigue life predictions remains challenging. Traditional multiaxial fatigue prediction models are often limited by specific material properties and loading conditions, making it difficult to maintain reliable life prediction results beyond these constraints. This paper presents a study on the impact of seven key feature quantities on multiaxial fatigue life, using Convolutional Neural Networks (CNN), Long Short-Term Memory Networks (LSTM), and Fully Connected Neural Networks (FCNN) within a deep learning framework. Fatigue test results from eight metal specimens were analyzed to identify these feature quantities, which were then extracted as critical time-series features. Using a CNN-LSTM network, these features were combined to form a feature matrix, which was subsequently input into an FCNN to predict metal fatigue life. A comparison of the fatigue life prediction results from the STFAN model with those from traditional prediction models—namely, the equivalent strain method, the maximum shear strain method, and the critical plane method—shows that the majority of predictions for the five metal materials and various loading conditions based on the STFAN model fall within an error band of 1.5 times. Additionally, all data points are within an error band of 2 times. These findings indicate that the STFAN model provides superior prediction accuracy compared to the traditional models, highlighting its broad applicability and high precision.