Spatio-temporal graph neural network for damage detection and global structural condition assessment

Data-driven structural health monitoring (SHM) systems are designed to assess the structural health conditions and to detect any potential damage, ensuring both structural safety and functionality. These systems rely on machine learning-based methods to process and analyze structural response data. The key step of these methods is damage sensitive feature extraction. Signal processing techniques, statistical modeling and neural networks were widely used in this step. However, these techniques presents limitations as information loss and computational complexity. Additionally, machine learning methods that use these techniques are supervised, which make them impractical because of damage and undamaged labels lack. To overcome these limitations, an unsupervised deep learning method based on spatio-temporel graph neural network is proposed in this paper. The method doesn’t require any preprocessing step to extract damage sensitive features. It integrates three key steps: First, Dynamic Time Warping (DTW) is used to construct a graph that captures the interactions among sensor measurements by assessing similarities between signals data. Second, a hybrid neural network architecture combining Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) layers is established to automatically capture the spatio-temporal dependencies from the historical sensor data of the undamaged structure. This combination enables accurate forecasting of expected sensor responses under normal conditions. Finally, the model prediction errors are analyzed to identify potential damage under unknown conditions; significant deviations between predicted and actual sensor data suggest damage presence. To quantify these deviations, the Kolmogorov–Smirnov test is employed, measuring the differences between the error distributions for undamaged and damaged scenarios. The proposed method is applied on the full-scale Tianjin bridge in order to demonstrate its efficiency in identifying the presence of damage and assessing the global structural condition of the bridge.