Abstract: To address the limitations of traditional methods in monitoring multi-frequency GNSS deformation of floating bridges regarding fluid-structure interaction and nonlinear problem-solving,the insufficient physical credibility of purely data-driven deep learning approaches,the impacts of marine multipath effects and ionospheric delays on multi-frequency GNSS observations,and the inadequate fusion of multi-sensor data,this study proposes a physics-informed neural network (PINN) framework integrating multi-frequency GNSS observations with structural dynamics equation constraints.Partial differential equations describing fluid-structure interactions,structural vibrations,and conservation laws are embedded into the neural network's loss function,constructing an end-to-end model featuring spatiotemporal adaptive multi-scale weighting and marine error correction.Validation using long-term monitoring data from the Bergsøysund Floating Bridge demonstrates:Compared to Kalman filtering,the root mean square error (RMSE)during normal monitoring is reduced by 35%~47%;Compared to purely deep learning methods,training data requirements decrease by 1~2 orders of magnitude,and generalization capability improves by over 60%;After multi-frequency fusion,the L5 band exhibits a 40% enhancement in multi-path error resistance,achieving an overall positioning accuracy of 0.8 cm.This framework provides a novel solution for intelligent deformation monitoring and prediction of structures in complex environments.

Key words: multi-frequency GNSS, deformation prediction, fluid-structure interaction, physics-informed neural network, structural health monitoring