Restoration of missing structural health monitoring data using spatiotemporal graph attention networks

For structural health monitoring systems with many low-cost sensors, missing data caused by sensor faults, power supply interruptions and data transmission errors are almost inevitable, significantly affecting structural diagnosis and evaluation. Considering the inherent spatial and temporal correlations in the sensor network, this study proposes a spatiotemporal graph attention network for restoration of missing data. The proposed model was stacked with a graph convolutional layer and several spatiotemporal blocks composed of spatial and temporal layers. The monitoring data of normal sensors were first mapped to all sensors through the graph convolutional layer, and attention mechanisms were used in the spatiotemporal blocks to model the spatial dependencies of sensors and the temporal dependencies of time steps, respectively. The extracted spatiotemporal features were assembled through a fully connected layer to reconstruct the missing signals. In this study, both homogeneous and heterogeneous monitoring items were used to calculate the spatial attention coefficients. The data restoration accuracy with and without the multi-source data fusion was discussed. Application on a long-span cable-stayed bridge to restore missing cable forces demonstrates that spatiotemporal attention modelling can achieve satisfactory restoring accuracy without any prior analysis.