Stiffness-Embedded GNN Model for Structural Component Importance Analysis

This study proposes a framework for evaluating structural component importance (SCI) in framed structures by combining graph neural networks (GNNs) with an energy-based method (EBM). EBM provides an accurate and comprehensive training data set through systematic analysis of SCI distribution characteristics and their influencing factors for a 3-span 4-story framed structure under multiple cases. It identifies the critical roles of component E, load paths, and structural topology in determining SCI. Based on this foundation, a stiffness-embedded GNN (S-GNN) model is developed. By embedding component stiffness into the adjacency matrix and integrating a localized sampling strategy with dual aggregation methods, S-GNN significantly enhances sensitivity to localized stiffness degradation. The proposed method employs graph representation techniques, where nodes represent structural components with ten extracted features, including type, location, stiffness, and load-related attributes, while edges represent the connections between components. Comparative experiments demonstrate that S-GNN achieves the best performance in SCI prediction for both intact and damaged structures, and its generalization capability is validated on unseen structures. The information aggregation of S-GNN effectively captures changes in component stiffness and their influence on system behavior and energy distribution, outperforming traditional models such as graph convolutional network (GCN) and graph sample and aggregation (GraphSAGE). The proposed model comprehensively integrates key factors influencing SCI, including component strength and stiffness, structural system topology, and load characteristics. This enables high-performance model training with limited data sets and achieves a computational efficiency over 200 times faster than traditional numerical methods.