A Cross-Domain Transfer Learning Method Incorporating Spatiotemporal Features for Structural Damage Identification

Abstract

In the field of Structural Damage Identification (SDI), the scarcity of actual damage data from engineering structures presents significant challenges for the application of deep learning models. In recent years, researchers have focused on integrating numerical simulation data with experimental data, employing transfer learning methods for damage detection. However, existing studies primarily rely on time-domain data or its transformed domain data for damage feature extraction and identification, often overlooking the spatial attributes of critical structural nodes (based on sensor placement) and their influence on damage detection outcomes. This study addresses this gap by proposing a Spatiotemporal features Learning Network model (SfLN) that utilizes the time-domain response data of critical structural nodes alongside their corresponding spatial distribution characteristics. The model employs a one-Dimensional Convolutional Neural Network (1DCNN) to extract common time-domain features from cross-domain data, followed by a multi-layer Graph Convolutional Network (GCN) to capture the spatial features associated with the "graph" attributes of critical nodes. The effectiveness of SDI through transfer learning under the fusion of spatiotemporal features is investigated. Validation through a case study involving a tower steel structure demonstrates the efficacy of the proposed method, achieving a damage classification accuracy of 97.684% and an F1 Score of 97.674% on target domain data testing—an improvement of 13.895 p.p. and 14.404 p.p., respectively, compared to results obtained from direct training on target domain data. Further validation through three comparative experiments underscores the advantages of the proposed approach, highlighting the significant contribution of the spatial distribution characteristics of critical structural nodes to damage feature extraction and task transfer. This work represents a valuable exploration into the integration of spatiotemporal features attributes for transfer learning in SDI.

Keywords: Structural damage identification, spatial characteristics of critical nodes, spatiotemporal features learning network, Transfer Learning