Near-real-time Identification of Seismic Damage by Graph Neural Network based on Structural Modes

Abstract

This paper proposes a near-real-time damage identification method based on the graph neural network (GNN) using the structural response data recorded during an earthquake event. The proposed method features a structural-mode-based weighted adjacency matrix to enable the GNN model to learn the spatial correlation and structural characteristics. The GNN model has an autoencoder architecture, one of the self-supervised deep neural networks that can detect anomalies in the input data by extracting important latent variables. The proposed method consists of ﾑencoderﾒ, ﾑgraph structure decoderﾒ, and ﾑfeature decoderﾒ that respectively learn latent variables of input data considering the spatial correlation, structural characteristics of the graph by reconstructing the adjacency matrix, and vibrational characteristics by reconstructing the response data. The GNN model is trained using the simulated structural responses and the structural-mode-based adjacency matrix of the target structure in a healthy state. The seismic damage of each member is then identified by the structural damage index calculated based on the reconstruction errors. As a numerical investigation, the proposed method is applied to two- and three-dimensional steel frame structures. Structural analyses are performed using ground motions from the PEER-NGA strong motion database to create the train, validation, and test datasets. The proposed method is verified by near-real-time simulations using the test dataset. The results demonstrate that the proposed GNN-based method can identify seismic damage accurately in near-real-time. The proposed method under various load conditions is expected to help reduce the time required for the post-disaster decision-making process by providing near-real-time damage identification.