Feature imputation and label prediction of dam deformation behavior based on Graph Representation Learning

The real-time and accurate prediction of dam deformation behavior is crucial for ensuring the safe operation of dams. Most existing models for predicting dam deformation behavior are built under the assumption of no missing data. Nonetheless, such an optimal scenario is seldom realized in practice, given the influence of subjective and objective considerations like human observational inaccuracies and malfunctions in monitoring instruments. This study introduces a framework designed for feature imputation and label prediction of dam deformation behavior at the same time in the presence of missing monitoring data. The novel approach entails formulating the issue utilizing a graph framework, where observations and features are denoted as separate nodes within a bipartite graph linked by edges that signify observed feature values. This formulation allows for treating feature imputation as a task of predicting missing edges, and label prediction as a task of predicting node values. These tasks are independently solved using Graph Neural Networks (GNN) without interference between them. The test results on two deformation datasets of an arch dam show that, the approach proposed in this study yields 50% lower mean absolute error for imputation tasks and 15% lower for label prediction tasks, compared with other advanced methods.