Multi-step cable force prediction based on hypergraph spatiotemporal deep learning

Cables are crucial components of cable-stayed bridges, whose mechanical behavior are commonly used for bridge condition assessment. Accurate cable force prediction of bridges is one of the important functions of structural health monitoring (SHM) systems related to inherent spatiotemporal correlation mining. This study proposes a novel spatiotemporal deep learning model for accurate multi-step structural responses prediction based on spatiotemporal hypergraph convolutional network (STHGCN). The input spatiotemporal data is mapped to high-dimensional space, whose spatiotemporal features are fully mined by STHGCN block consisting of hypergraph convolutional layer and one-dimensional convolutional layers, and finally the prediction results are generated by the dimensionality reduction of the fully connected layer. The extraction of spatiotemporal features is done by STHGCN blocks without the involvement of other feature extraction modules (e.g., attention mechanism) to display the capability of hypergraph-based model in feature extraction. Based on different hypergraph generation methods, two STHGCN models: static method-based STHGCN(S) and dynamic method-based STHGCN(D) are established. Compared with three classical models and STHGCN(S) in a case study cable-stayed bridge, the prediction accuracy in 12-step parallel-based task of STHGCN(D) improves by 39.83 %, 39.68 %, 24.52 % and 2.68 %, respectively, which demonstrate the strong ability of hypergraph-based models in multi-step prediction tasks.