Condition assessment of cable-stayed bridges based on heterogeneous data correlation mining using multichannel spatiotemporal graph convolutional network

Heterogeneous monitoring data describe the state of bridges. Extracting their spatiotemporal relations for condition assessment is a challenging issue in the field of structural health monitoring (SHM). This study conducted a condition assessment approach of cable-stayed bridges by analyzing the spatiotemporal characteristics of heterogeneous monitoring data (cable force, girder vertical deflection, and pylon slope) using a multichannel spatiotemporal graph convolutional network. The cable dynamometers, hydraulic pressure transmitters, and inclinometers were represented as heterogeneous vertices on a directed graph. The spatial and temporal features were extracted using learnable mapping matrices generalized from the adjacency matrix and one-dimensional convolutional neural networks, respectively. Using the designated penalty terms, the proposed model ensures that each sensor can hierarchically aggregate information from both homogeneous and heterogeneous vertices within each neighborhood order. By combining adjacency relations with regression residuals, the confusing problem of local data pattern anomalies and global data pattern anomalies can be addressed based on the local or global anomalies of the spatiotemporal model parameters. Finally, the effectiveness of the proposed method was validated in a long-span cable-stayed bridge for faulty sensor localization and structural change detection.