Near-real-time damage identification under vehicle loads using dynamic graph neural network based on proper orthogonal decomposition

Abstract

Structural health monitoring (SHM) is essential for managing infrastructure by continuously monitoring performance. To address the complexity of SHM for large systems, this study introduces a dynamic graph neural network (DynGNN) approach for near-real-time damage identification. The approach represents the infrastructure as a graph and uses a dynamic adjacency matrix based on proper orthogonal decomposition (POD) to capture the dynamic characteristics and spatial correlations of responses. The DynGNN model has an autoencoder architecture with one encoder for latent feature extraction and two decoders for capturing spatiotemporal changes and is trained on intact structural responses. The proposed approach employs a damage index based on reconstruction errors for damage identification. In applications to a steel truss bridge under vehicle loads, the proposed approach is tested by real-time simulation using the field test data obtained from the damaged bridge. By incorporating a dynamic adjacency matrix, the results successfully demonstrate that the proposed approach effectively adapts to the time-varying nature of structural responses and can accurately identify damage in near-real time. The novel integration of a dynamic graph structure with POD allows the proposed method to capture both spatial and temporal variations, providing a significant improvement over traditional static methods.