A novel spatial–temporal synergistic transformer for structural vibration response reconstruction

Structural vibration response reconstruction is of significant importance because data incompleteness critically undermines the accuracy of structural health monitoring. Existing deep learning-based reconstruction methods neglect the multi-semantic information of structural vibration responses that reflects the inherent vibration characteristics of structures, leading to reconstruction results that fail to represent the true structural behavior, thereby reducing their reconstruction accuracy. To address this problem, we propose a novel spatial–temporal synergistic transformer (STSFormer) network that effectively extracts the multi-semantic information through spatial–temporal synergistic representation. Specifically, we design a spatial and channel synergistic attention consisting of shared multi-semantic spatial attention and progressive channel self-attention, which enhances the ability to extract multi-semantic information and mitigates their discrepancies by three core designs including dimensional decoupling, lightweight multi-semantic guidance and semantic discrepancy mitigation. Furthermore, STSFormer adopts a symmetrical encoder–decoder architecture with skip connections, which strengthens the effective fusion of deep and shallow semantic features. On the real acceleration dataset of the Canton Tower, the average relative error (RE) of our method is reduced by approximately 11.5%, 16.1%, 8.9% and 2.6% compared with the average RE of MSDF, PMTC, SegGAN and SAGAN, respectively. On the real acceleration dataset of the Hardanger Bridge, the average RE of our method is decreased by 6.3% and 11.3% relative to MSDF and PMTC. Experimental results on multiple measured scenarios demonstrate that the reconstruction accuracy of the proposed method outperforms representative state-of-the-art models. Meanwhile, the proposed method maintains lightweight design and possesses favorable robustness and generalization performance. The source code of the model has been publicly released at: https://github.com/FanyangCHD/STSFormer-for-SHM .