A physics-informed neural network for predicting vibration-induced human annoyance in over-track buildings

Existing vibration assessment standards for over-track buildings are mainly based on vibration-limit criteria and do not explicitly account for nonlinear perceptual thresholds or individual differences in human sensitivity. To address this limitation, this study proposes a physics-informed neural-network framework for predicting vibration-induced human annoyance. The artificial intelligence contribution lies in incorporating monotonicity constraints and boundary conditions into the loss function, enabling the model to learn a physically consistent nonlinear relationship between vibration levels and human annoyance from survey data. The engineering application focuses on over-track buildings above a metro depot. A vehicle-track coupled model and a three-dimensional finite element model of the soil-track-depot-building system are developed to calculate the maximum Z-weighted vibration levels at key response points in different functional zones. These vibration responses are then used as inputs for spatial annoyance prediction. The results show that the proposed model achieves high predictive accuracy and good generalization, with a mean squared error below 0.01 and a coefficient of determination above 0.95 across all evaluated scenarios. The predicted annoyance decreases with lateral distance from the track in commercial buildings and with floor height in residential buildings. Compared with the Chinese national environmental vibration standard GB 10070-88, the proposed framework identifies localized annoyance exceedances even under nominally compliant vibration conditions and provides a continuous spatial distribution of vibration-induced environmental impacts.