Fatigue life prediction for orthotropic steel bridge decks welds using a Gaussian variational bayes network and small sample experimental data

Ensuring the structural integrity of orthotropic steel bridge decks (OSBDs) is paramount, particularly when predicting fatigue life of welds, which are susceptible to failure under repeated loading. Traditional S-N curve models for fatigue life prediction often struggle with limited generalization ability and low accuracy due to the influence of sensitive parameters. To address these challenges, this paper proposes a Gaussian Variational Bayes Network (GVBN) probabilistic prediction model specifically designed for small sample datasets. Leveraging Bayesian inference, GVBN effectively utilizes a normalized training dataset constructed from 27 experimental studies on rib-to-deck welds (RTDWs) in OSBDs. The GVBN model’s performance was evaluated against Back-Propagation Neural Network (BPNN), Gaussian Process Regression (GPR), and Bayesian Neural Network (BNN) based on metrics such as R2, MSE, and RMSE for both training and prediction datasets. A global sensitivity analysis using the Shapley Additive exPlanations (SHAP) method identified average stress ratio as the most influential parameter affecting weld fatigue life. Further analysis demonstrated that increasing the training data size significantly improves model accuracy. The proposed GVBN model exhibits superior fitting accuracy and generalization ability compared to traditional deep learning models, offering valuable insights for predicting fatigue life in OSBDs, particularly when data is scarce.