Lost data reconstruction for structural health monitoring by parallel mixed Transformer-CNN network

Abstract

Transformer-based and convolutional neural network-based deep learning methods have been extensively utilized to reconstruct lost data for structural health monitoring. However, the lack of inductive bias in Transformer and the fixed receptive fields of convolutional neural network (CNN) limit their local and global modeling capabilities, respectively. The traditional method of combining the two in series creates a sequential relationship and interdependence, negatively impacting reconstruction accuracy. To address this problem, this paper proposes a novel parallel mixed Transformer-CNN network. By parallel connecting the shifted window transformer and densely connected convolutional block, both can process structural vibration responses simultaneously and independently, leveraging the locality of convolution to address the inductive bias of Transformer. Experiments on real acceleration data from the Canton Tower validate its effectiveness. Compared with models using only Transformer or CNN, the reconstruction errors are reduced by 56% and 17%, respectively. The modal parameters extracted from the reconstructed data are highly consistent with those from the true data, and our model has good robustness to lost rates and noise.