Physics-guided data fusion for settlement prediction of high-fill railway embankments

Abstract

Settlement prediction for high-fill railway embankments is challenged by monitoring noise and stage-dependent deformation across construction and operation. This paper presents a physics-guided framework integrating multivariate construction information with settlement monitoring data. Monitoring sequences are denoised by a hybrid autoencoder with a signal-to-noise-ratio-based cleaning mechanism. A temporal graph convolutional neural network embeds a modified layer-wise summation formulation to guide variable interactions and settlement prediction. On the analysed dataset, the model achieved a coefficient of determination of 0.99 with the selected parameters. Compared with Informer, MTGNN, and TQNet, it achieved the lowest test-set mean squared error, averaged over ten random seeds, under both working conditions. Ablation experiments supported the effectiveness of the hybrid architecture. Component analyses indicated that the physics-guided module estimated settlement trends, while the data-driven module supplied local residual corrections. It provides a feasible framework for handling complex settlement-pattern transitions that are difficult to capture using conventional approaches.