Dynamic Structural Early Warning for Bridge Based on Deep Learning: Methodology and Engineering Application

In bridge health monitoring, structural responses are strongly coupled with temperature effects and vehicle load effects, making it difficult for conventional fixed thresholds and single data-driven approaches to simultaneously achieve environmental adaptability and quantitative reliability assessment. To address this issue, this study proposes a deep-learning-based dynamic early-warning method for bridge structures, using health-monitoring data from an in-service long-span cable-stayed bridge as the research background. First, a two-month mid-span deflection time series is processed using variational mode decomposition optimized by the Porcupine Optimization Algorithm to separate temperature-induced effects. Subsequently, a hybrid prediction model integrating Informer and SEnet is constructed. Temperature and temperature-induced deflection components are used as input features, and a sliding-window strategy is adopted to achieve high-accuracy prediction of the temperature-induced deflection trend, which serves as the time-varying baseline of the dynamic threshold. On this basis, vehicle load effects are modeled by combining Pareto extreme value theory with finite element analysis and superimposed to establish a two-level dynamic early-warning threshold system that satisfies code requirements. Furthermore, a stochastic finite element Monte Carlo method is introduced to probabilistically model uncertainties associated with material parameters, load effects, and model prediction errors. The threshold failure probability at each time instant is taken as the evaluation metric, enabling quantitative characterization of threshold reliability. The results indicate that under combined multiple working conditions, the proposed method reduces the maximum failure probability of the first-level warning by 32.68% and that of the second-level warning by 93.48%, with more stable and consistent probabilistic responses. In engineering applications, simulation experiments based on stochastic traffic loading show that the warning accuracy is improved by up to 19.27%, while the error rate is reduced by up to 16.16%. The study demonstrates that the proposed method possesses a clear physical and statistical foundation as well as good engineering feasibility and provides a viable pathway for transforming bridge early-warning systems from experience-based schemes toward data-driven and risk-oriented frameworks.