Rapid postearthquake modelling method for deformation monitoring models of high arch dams based on metalearning and graph attention

•Proposal of a SHM model for high arch dams suitable for small sample sizes post-earthquake.•Introduction of a meta-training framework for post-earthquake monitoring data of dams.•Application of the method to the world’s only 200 m-high arch dam subjected to strong near-field earthquakes. Southwest China is the world’s most densely populated area for high dams over 200 m and is also a region with high seismic activity. Earthquakes can significantly alter dam structures, resulting in substantial discrepancies between preearthquake and postearthquake deformation monitoring data. Deformation is a critical indicator of the structural response of dams to internal and external environmental factors. Establishing a dam deformation structural health monitoring (SHM) model promptly after an earthquake is crucial for postearthquake structural health analysis and preventing major accidents. In this paper, we propose a rapid modelling method for postearthquake deformation SHM of high arch dams that uses metalearning and graph attention techniques. First, we develop an SHM model tailored for postseismic small-sample data modelling, integrating a multihead attention mechanism with hydraulic-temporal graph feature fusion. On this basis, we introduce a metalearning framework to derive the initial model parameters from preearthquake data. The proposed model is applied to vertical radial deformation monitoring of the world’s only 200-metre-high arch dam subjected to strong near-field earthquakes. The effectiveness of our metalearning framework for postearthquake data is validated by comparing it with the transfer learning framework. Through a comparison with nine baseline models across six postearthquake modelling scenarios, we demonstrate that the proposed model achieves the highest accuracy and exhibits unique engineering applicability for rapid postearthquake modelling tasks. Ablation experiments further confirm the effectiveness of the proposed modules.