A novel graph neural network framework with self-evolutionary mechanism: Application to train-bridge coupled systems

Abstract

Deep learning (DL) methods have been widely applied for structural response prediction. However, classical DL methods rely heavily on training data with no consideration to the information at the structural level. They generally show poor generalization performance for unknown structural forms. To address this issue, a graph representation is proposed in this study to abstractly represent the actual structure as a graph structure, which is subsequently processed using the graph isomorphic network (GIN). Due to the unique self-evolutionary mechanism of the graph structure, the GIN model is able to disentangle from the training data, leading to excellent generalization performance on the task of response analysis with unknown structural forms. Taking train-bridge coupled (TBC) systems as examples, for different working conditions, the test results show that the prediction accuracy and generalization performance of the GIN model reach an extremely high level. Moreover, a GIN-based iterative system is proposed in this study. It exhibits significantly better generalization performance than classical DL methods for unknown structural forms, indicating its high potential for practical applications in various engineering fields. The content and findings of this study contribute to the future development of a new generation of DL methods with advanced performance.