This study develops a green, high-performance, cement-based grout by replacing manufactured sand with iron tailings sand (ITS) at ratios of 0–50% to address resource depletion. Fluidity, mechanical strength, and expansion rates were experimentally evaluated to determine engineering feasibility. The results indicate that while ITS inclusion reduces fluidity due to particle morphology, it significantly enhances compressive strength through a physical filling effect. Specifically, the 30% replacement group achieved a peak 28-day compressive strength of 100.4 MPa. Comprehensive analysis identifies 40% as the optimal replacement rate, where the grout strictly satisfies relevant industry specifications regarding fluidity, early strength, and volume stability. This research demonstrates the practical significance of utilizing industrial solid waste to produce high-performance sleeve grout for prefabricated construction.

Bridge deflection monitoring is essential for structural health monitoring, yet deflection signals are complex multiscale responses jointly influenced by environmental and traffic loads. To improve prediction accuracy while reducing computational cost, this study proposes a lightweight two-stage framework integrating Gaussian filtering decomposition with physical feature enhancement. A Gaussian low-pass filter first separates the original deflection signal into low-frequency trend and high-frequency detail components. Effective temperature features considering thermal lag and dynamic load-flow features based on influence-line weights are then constructed to enhance the physical interpretability of the inputs. Two parallel Long Short-Term Memory (LSTM) networks independently predict the trend and detail components, which are subsequently combined to obtain the final deflection response. In the presented comparison, the Trend–Detail model achieves lower RMSE and MAE than the Naive and Baseline models, indicating improved deflection prediction accuracy for the investigated bridge. With parallel execution of its independent trend and detail branches, the Trend–Detail model achieves a 1.91-fold inference speed-up relative to the Baseline model under the same hardware and execution settings.