Temperature-induced response reconstruction method based on DL-AR model and attention mechanism

• Published in: Structures (Oxford) Vol. 50; pp. 359 - 372
• Main Authors: Chen, Chengbin, Tang, Liqun, Lu, Yonghui, Zhou, Licheng, Liu, Zejia, Liu, Yiping, Jiang, Zhenyu, Yang, Bao
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.04.2023
• Subjects: Attention mechanism; Autoregressive model; Data reconstruction; Deep learning; Structural health monitoring; Deep learning; Attention mechanism; Autoregressive model; Data reconstruction; Structural health monitoring
• ISSN: 2352-0124; 2352-0124
• DOI: 10.1016/j.istruc.2023.02.044

Data loss is unavoidable for structural health monitoring (SHM) systems due to various potential factors, such as equipment aging, radio interference. Previous studies have revealed that it is more challenging to extract structural state information from incomplete data. Therefore, it is important to develop high-precision response reconstruction methods for missing data. In this study, a hybrid deep-learning and autoregressive model with attention mechanism (DL-AR-ATT) framework is proposed to accurately reconstruct structural responses considering data correlations. The proposed model consists of the deep-learning (DL) and autoregressive (AR) components, which can simultaneously model the linear and nonlinear structure of data. To adaptively weight the contribution of the two components to the final result instead of directly summing, the study introduces a custom weighting layer into DL-AR-ATT. Moreover, the attention mechanism is introduced to eliminate redundant information caused by the data from multiple sensors as input. To verify the validity and feasibility of the proposed method, long-term SHM data from a long-span steel box girder suspension bridge and a prestressed concrete continuous box girder bridge are used. The results show that the reconstruction performance of DL-AR-ATT is superior to that of the one-dimensional convolutional neural network (1D-CNN)-based and hybrid DL and AR (DL-AR) models. Finally, the weights of the modules of DL-AR and DL-AR-ATT are visualized to verify the ability of the model to capture data correlations, and to explain how the attention and custom weighting layers improve the performance of DL-AR-ATT.