Multi-channel response reconstruction using transformer based generative adversarial network

Abstract

Accurate measurement data are a basic prerequisite for effective structural health monitoring (SHM). However, data loss are inevitable in the long-term monitoring of large-scale structures. To solve this problem, this research proposes a transformer-based generative adversarial network (GAN) to reconstruct lost measurements from observed measurements. The generator of GAN is an encoder-decoder structure using transformer as the backbone combined with discrete wavelet transform. Skip connections are used between the encoder part and decoder part to promote multi-scale information flow. A novel discriminator is designed to assess the reality of wavelet spectra of reconstructed samples. To deceive the discriminator, the generator must generate samples that are accurate over the full frequency band. The developed model is used to reconstruct linear responses of a footbridge under pedestrian excitations and nonlinear responses of a suspension bridge under typhoon events. Experimental results demonstrate that lost responses can be reconstructed accurately, even when a large proportion of data are lost. The effectiveness of the proposed method is further verified by comparing the reconstruction accuracy of the proposed model with those of other three state-of-the-art models. The results demonstrate that an improved performance of applying the proposed approach for dynamic structural response reconstruction is achieved and validated with in-field testing data under ambient and extreme excitation conditions.