Ultra-high cycle fatigue life prediction of titanium alloy with small sample size based on the PSO-BP model

Accurate prediction of the ultra-high cycle fatigue (UHCF) life of titanium alloys is essential for the design of safe and reliable aero-engine critical components. Few machine learning models have been utilized in ultra-high cycle fatigue prediction because the methods have limitations in dealing with data sparsity and overfitting issues. The current work aims to overcome the sparsity of the UHCF experimental data and to propose a simple and non-redundant ML prediction model. The size of the dataset is extended by the Gaussian Mixture Model (GMM), and an improved ML method is presented to analyze the synergistic effects of elastic modulus, tensile strength, yield strength, specimen size, and stress amplitude on titanium alloy. Compared with traditional machine learning methods, the model predicts fatigue life with significantly improved accuracy and stability. With sufficiently large amounts of data, the model achieves higher accuracy (R² = 89.9 %) and smaller prediction fluctuations (S_e = 0.0494), which is 4.93 % higher and 4.63 % lower, respectively, compared with the BP neural network that is much better in prediction. This study has important application value for the prediction of ultra-high cycle fatigue life of titanium alloy materials in aeronautical engines.