Structural dynamic response reconstruction using self-attention enhanced generative adversarial networks

In structural health monitoring (SHM) of civil engineering structures, loss of measured structural responses inevitably occurs in practice, especially when structures encounter extreme loads. The loss of measurements severely undermines the completeness of the collected structural information and the reliability of structural condition assessment. Therefore, timely and accurate recovery of lost data is of paramount importance. This paper proposes a novel approach based on a self-attention mechanism enhanced generative adversarial network (SAGAN) for learning the intrinsic correlations between responses and reconstructing the lost data based on the accurately measured ones. SAGAN innovatively embeds the self-attention mechanism in the computational flow to facilitate the extraction of spatial and even temporal correlations among structural responses. In the experimental validations, the reconstructed responses of Guangzhou New Television Tower (GNTT) show great agreement with the true responses in forms of both time sequences and Fourier spectra. SAGAN is also versatile, demonstrating its effectiveness and robustness by reconstructing the responses competently under both ambient and typhoon excitations. In addition, by visualizing and analyzing the internal matrices and feature maps of SAGAN, it is found that the self-attention module benefits the learning of data features and improves the establishment of mappings between responses. The suitability of the proposed approach for SHM related tasks is validated by extremely consistent modal parameters. The identified natural frequencies from the reconstructed and the corresponding true responses are identical, and the Coordinate Modal Assurance Criterion (COMAC) value reaches to 99.98%.