A machine learning and multi-source authentic data-driven framework for accurate fatigue life prediction of welds in existing steel bridge decks

Abstract

To address the challenge of inaccurate fatigue life prediction for welds in existing orthotropic steel bridge decks (OSBDs), which stems from incomplete information and insufficient model generalizability, this paper proposes a precise prediction framework that integrates multi-source authentic data and machine learning. Firstly, a U-Net model is employed to achieve automated identification and measurement of the OSBDs’ geometric parameters, resolving the issue of distorted resistance information. Secondly, based on wavelet analysis, strain data from bridge health monitoring is denoised and reconstructed, establishing an authentic stress time-history database. To tackle the scarcity of fatigue test data, a Gaussian Variational Bayesian Network- Attention (GVBN-Attention) prediction model is developed. This model enhances the identification of key features through a self-attention mechanism and quantifies parameter uncertainty using variational Bayesian inference. Experimental results on 192 sample sets demonstrate that the proposed GVBN-Attention model significantly outperforms comparative models, including Gaussian Process Regression (GPR) and Bayesian Neural Networks (BNNs), on key metrics such as the coefficient of determination and root mean square error, exhibiting superior predictive accuracy and generalization capability. Furthermore, SHAP (SHapley Additive exPlanations)-based interpretability analysis reveals the influence mechanisms of input features like stress amplitude and deck plate thickness on fatigue life, validating the rationality of the model's decision-making. This study provides a new method for fatigue life assessment of steel deck welds under data-scarce conditions, characterized by high accuracy, interpretability, and uncertainty quantification.