Remaining life prediction and parameter identification in fatigue crack propagation of welded joints in OSDs using a physical-information neural network approach

The prediction of remaining fatigue crack growth life (FCGL) is crucial for the long-term performance maintenance of bridges. This study proposes a novel physics-informed neural network to identify crack growth parameters and predict remaining FCGL. Initially, a reliable physical numerical model was constructed to generate random crack propagation samples for evaluating the performance of machine learning models. Subsequently, the neural network established the relationship between crack depth and fatigue cycles. Automatic partial differential inverse identification techniques were used to determine the key parameters of crack propagation. Additionally, a composite loss function was designed to dynamically update crack propagation parameters, ensuring the physical consistency of FCGL predictions. The results show that the proposed method significantly improves extrapolation and generalization performance. FCGL predictions align with authoritative numerical methods, with all results within a 1.5-fold error threshold. Compared to traditional data-driven methods, the proposed method demonstrates greater robustness and accuracy. It can effectively replace existing numerical simulation methods for rapid FCGL prediction. • Physical principles are embedded into neural networks to guide the gradient descent process. • A self-updating PINN is developed to predict the remaining life of welded joints in OSDs. • The model updates crack parameters automatically, ensuring consistent and accurate physical predictions. • The model outperforms data-driven methods and yields predictions consistent with numerical simulations.