A Physics-Informed Multi-Modal Deep Learning Framework for Stochastic Prediction and Reliability Assessment of Human-Induced Vibrations in Lightweight and Composite Floor Systems

Human-induced vibrations in lightweight and composite floor systems pose important serviceability concerns because these structures typically have lower stiffness, reduced mass, and higher sensitivity to dynamic loading. Conventional deterministic design methods often overlook the natural variability and short-term amplification effects caused by walking and other intermittent human activities. This study presents a physics-guided, multi-modal data-driven framework for predicting vibration responses and assessing structural reliability under human-induced loading conditions. Monitoring conducted over a 1000-second period shows bounded but dynamically fluctuating acceleration responses ranging between 9.5 m/s² and 10.2 m/s², concentrated around gravitational acceleration. Statistical evaluation indicates an approximately Gaussian distribution with minimal skewness, suggesting stable damping behavior and consistent stiffness properties. Frequency-domain analysis highlights dominant low-frequency components without noticeable resonance amplification within the typical walking frequency range of 1–4 Hz. Rolling mean trends reveal negligible long-term drift, confirming stable structural performance. The framework combines acceleration, strain, and environmental measurements to improve predictive capability and reliability evaluation. Correlation analysis shows that strain measurements strengthen response prediction, while temperature has a comparatively minor effect. Reliability assessment based on a probabilistic threshold model (μ + 2σ) identifies very few exceedance events, indicating a high reliability level and compliance with serviceability requirements. By integrating probabilistic modeling with physics-based structural constraints, the approach enables more accurate forecasting of vibration behavior and real-time reliability assessment. Overall, the results demonstrate improved prediction performance and provide a practical methodology for monitoring and evaluating serviceability in modern lightweight and composite floor systems.