TPS-GNN: Predictive model for bridge health monitoring data based on temporal periodicity and global spatial distribution characteristics

Abstract

In recent years, the investigation of spatiotemporal dynamics in bridge health monitoring data has gained substantial prominence, with spatiotemporal models becoming a focal point of research in this domain. Traditional multilayer graph neural networks (GNNs) struggle to effectively distinguish the influence of sensor data from various cross-sectional positions on target nodes due to the oversmoothing issue when capturing spatial dependencies. Similarly, recurrent neural networks exhibit limitations in managing long-term dependencies and capturing periodic patterns, reducing their efficacy in applications like bridge monitoring, where an in-depth analysis of long-term trends is essential. To address these challenges, we introduce the concept of temporal periodicity and global spatial (TPS) and propose a predictive model named TPS-GNN. This model is specifically designed to leverage the temporal periodicity of data and the global spatial distribution characteristics. For spatial features, TPS-GNN directly employs representations from different hop neighborhoods, allowing it to accurately capture the spatial dependencies of sensor data across varying hop distances. This approach effectively retains node feature information and capitalizes on both local and global information, enhancing the model’s capacity to represent complex spatial relationships. Regarding temporal features, we introduce a position-encoding scheme that integrates daily, weekly, and seasonal periodic information into the flow data. The current and historical models process continuous and periodic time series data separately, thus better capturing temporal dependencies. Experimental results on two real-world bridge health monitoring datasets demonstrate that TPS-GNN significantly outperforms existing state-of-the-art predictive models, achieving up to a 78% improvement in prediction performance. This highlights the model’s ability to accurately predict bridge health monitoring data, enhancing the precision of data forecasts and providing reliable support for bridge maintenance and management.