Digital twins for static and dynamic responses of a continuous rigid frame bridge based on multi-task learning and optimization algorithm

With the increasing demand for proactive maintenance of civil infrastructure, digital twin technology demonstrates significant value in the domain of bridge health monitoring. This paper presents a multi-task deep learning-optimization framework for constructing a high-fidelity digital twin of a continuous rigid frame bridge, enabling accurate static and dynamic responses simulation and performance prediction. The bridge is the physical entity, with in-situ tests conducted across its operational state, and a finite element model is constructed as the virtual entity. A deep learning model is developed to capture the nonlinear relationship between material parameters and bridge responses. Based on the parameter identification results from the physical entity as optimization targets, a multi-objective optimization algorithm is performed to update the virtual entity. Experimental validation demonstrates a significant reduction in the mean prediction error from 31.2 % to 11.7 % for static responses and from 22.39 % to 7.15 % for dynamic responses, which confirms the operational feasibility and predictive accuracy of digital twin. The digital twin framework is applied to the smart control of bearing reaction forces through height adjustment, providing a digital platform for lifecycle bridge management. This study introduces three key advances: (1) simultaneous use of static and dynamic responses within a unified MTNN; (2) an MSO strategy that decouples heterogeneous loss gradients for balanced multi-task learning; and (3) a closed-loop validation mechanism using intelligent bearings to ensure consistency between the updated FEM and structural behavior.