Physics-Constrained Machine Learning-Driven Flow-Induced Vibration of Large-Diameter Siphon Steel Pipes

Addressing the challenges in monitoring flow-induced vibrations (FIV) of large-diameter siphon steel pipes and the low accuracy of traditional models, this study integrates physics constraints with machine learning to construct a high-fidelity intelligent simulation model. By employing Bayesian inference to quantify parameter uncertainty, the proposed method achieves quantitative assessment of structural safety status, significantly reducing the risks of false alarms and missed detections. Specifically, a Physics-Constrained Temporal Neural Network (PCTNN) is proposed, which embeds structural dynamics priors into a deep learning architecture and is trained using measured displacement data from multiple operating conditions. The SHAP (SHapley Additive exPlanations) method is utilized to analyze feature contributions, thereby revealing the decision-making mechanisms of the model. Furthermore, a Variational Bayesian inversion framework is constructed to perform posterior sampling on four key physical parameters: elastic modulus (E), density (ρ), damping coefficient(c), and pipe diameter(d), systematically evaluating their identifiability and uncertainty. The result shows that the proposed PCTNN-Bayesian joint framework enables high-precision, physics-consistent FIV simulation and reliable parameter identification for siphon pipes. It is recommended that, in practical applications, only highly identifiable parameters (E, ρ, d) be inverted, while the damping coefficient (c) should be fixed at its design value. This research provides a technical pathway for the intelligent monitoring and safety assessment of large-scale hydraulic metal structures, offering a balance of accuracy, robustness, and physical interpretability.