In this paper, an intelligent recognition and spatio-temporal evolution model combining physical constraints and deep learning is proposed. Firstly, a data conflict resolution mechanism based on the equilibrium equation of elasticity is constructed to unify multi-source heterogeneous monitoring data into a physically credible deformation field; Secondly, a multi-modal comparative learning network is designed, and the graph convolution encoder is pre-trained with unlabeled data to solve the problem of deformation pattern recognition in small sample scenes. Finally, a Spatio-Temporal Graph Attention Network (ST-GAT) is proposed, which adaptively captures the spatial propagation path of deformation through graph attention mechanism, and combines Gated Recurrent Units (GRU) to model temporal dependencies, achieving high-precision spatiotemporal prediction. The experiment is based on the three-year multi-source monitoring data of a subway tunnel in a city. The results show that the physical constraint fusion can reduce the displacement error of the conflict area by about 20%, the recognition accuracy of comparative learning can reach more than 85% with 1% labeled data, and the three-step prediction Root Mean Square Error (RMSE) of ST-GAT is 1.76 mm, which are significantly better than the traditional methods. The model effectively improves the intelligent level and interpretability of deformation monitoring, and provides reliable technical support for tunnel safety operation and maintenance.
Yang et al. (Sun,) studied this question.