Machine learning-based fatigue life prediction of double-sided U-rib welded joints considering multiple factors

Abstract

This study systematically developed and evaluated five machine learning models, Support Vector Regression (SVR), Gaussian Process Regression (GPR), Neural Network Regression (NNR), Least Squares Boosting (LSBoost), and Random Feature Kernel Ridge Regression (RF-KRR), to address the challenges of crack depth and remaining fatigue life predictions for single- and double-sided U-rib welded joints in orthotropic steel decks, considering multiple influencing factors. A high-fidelity finite element model was established to analyze fatigue crack growth through integrated simulation using ABAQUS and FRANC3D, incorporating input features such as nominal stress range, deck thickness, welding type, residual stress, and crack length, with crack depth and remaining life as output targets. The results show that the evolution laws of the crack aspect ratio differ significantly for single- and double-sided U-rib welded joints. Correlation analysis reveals a strong positive relationship between crack length and crack depth. In contrast, residual stress, crack length, and crack depth exhibit significant negative correlations with remaining life, whereas double-sided welds and increased deck thickness contribute positively to enhanced fatigue resistance. In terms of model performance, the GPR model exhibits the highest accuracy and generalization in crack depth prediction, while the NNR model outperforms others in fatigue life prediction. Accordingly, the GPR model is recommended for crack depth prediction, and the NNR model is identified as the most suitable for fatigue life prediction.