Global Damage Identification Method Based on Multi-Source Signals and ST-GNN for Frame Structures

A framework of a global damage identification method based on multi-source signals and spatiotemporal graph neural network (ST-GNN) is proposed for frame structures. A new architecture of the ST-GNN consisting of a temporal feature encoder, a spatial feature encoder and a classifier is constructed. In the ST-GNN, the structure is represented by a graph defined by graph nodes and graph edges with time sequential signals and a modal weighted adjacent matrix, and a multi-head attention mechanism is used to fuse the multi-source signals of the graph nodes and edges. The modal weighted adjacent matrix with mode shape information is employed to embed deep structural physical behaviors into the ST-GNN, leading the message-passing process of GNN to be more physically meaningful. A new loss function is designed to solve the imbalanced label problem during training of the damage identification model, and evaluation indicators are used to assess the performance of the model. The global damage identification model of a frame structure is created by means of the presented method, and its performances are compared with the models based on the CNN, LSTM, and GRU machine learning algorithms. It is shown that the global damage identification method based on the multi-source signals and the newly developed ST-GNN shows superior capabilities in global damage identification, stability, anti-noise performance, and generalization ability.