Structural Deformation Prediction and Uncertainty Quantification via Physics-Informed Data-Driven Learning

In structural health monitoring, purely data-driven methods for deformation prediction are often susceptible to time-varying boundary conditions under complex operating scenarios, leading to insufficient physical interpretability and limited generalization across different conditions. To address these challenges, this study proposes a Physics-Informed Dual-branch Long Short-Term Memory framework (PINN-DualSHM). The framework employs dual-branch LSTMs to separately extract temporal features of structural mechanical responses and environmental thermal effects. Dynamic decoupling and fusion of these heterogeneous features are achieved through an adaptive cross-attention mechanism. Furthermore, physical priors, including the thermodynamic superposition principle and structural settlement monotonicity, are embedded into the loss function as regularization terms, complemented by a dual uncertainty quantification system based on heteroscedastic regression and MC Dropout. Experimental results based on long-term measured data from an industrial base project in Shenzhen demonstrate that PINN-DualSHM significantly outperforms baseline models such as LSTM, CNN-LSTM, and GAT-LSTM. Specifically, the Root Mean Square Error (RMSE) is reduced by 65.25%, and the coefficient of determination (R2) reaches 0.925. Physical consistency analysis confirms that the introduction of physical constraints effectively suppresses anomalous predictive fluctuations that violate mechanical laws. Uncertainty decomposition reveals that aleatoric uncertainty is dominant (93.7%), objectively indicating that the current system’s accuracy bottleneck lies in sensor noise rather than model capability. By enhancing prediction accuracy while providing credible quantitative assessments and physical interpretability, the proposed method provides a scientific basis for the operation, maintenance optimization, and upgrading decisions of SHM systems.