A graph convolutional network method for rapid global stiffness prediction under multiple damage scenarios

Accurately predicting the overall structural stiffness under damage conditions is a crucial aspect of structural health monitoring, as it is essential for ensuring the safety and reliability of engineering structures, given that damage significantly impairs load-bearing capacity. Currently, research on structural stiffness degradation and load-bearing capacity assessment under damage scenarios remains limited, with few studies on structural performance evaluation. This article proposes a method based on graph convolutional networks (GCNs) to predict the overall structural stiffness under damage conditions. This study focuses on the four most common and critical damage conditions that cause significant stiffness reduction in bridge structures: pile cap settlement, bearing detachment, steel corrosion, and cracking. The stiffness prediction GCN method designs an adjacency matrix to label structural features under scenarios such as pile cap settlement, bearing detachment, steel corrosion, and cracking, incorporating prior information including initial stiffness and monitoring data. A finite element (FE) model is employed to simulate the overall structural stiffness degradation for these damage types, training the graph neural network (GNN) to achieve accurate predictions of global structural stiffness. For each damage scenario, the GNN learns the degradation patterns of the overall structural stiffness from the onset of damage to severe levels. Its accuracy is validated using a test set that covers both initial and severe damage stages. The method’s effectiveness is further demonstrated through a real bridge case study, accurately predicting stiffness degradation due to pile cap settlement, bearing detachment, steel corrosion, and cracking. Results show that the proposed method enables precise predictions of structural stiffness decline for these four damage types, with a maximum error of 1.86%. Compared to deep neural network methods, it achieves higher accuracy and significantly reduces computation time relative to FE analysis methods.