Structural floor acceleration denoising method using generative adversarial network

Abstract

Structural floor acceleration recorded by the safety network of buildings is essential for detecting and assessing the state of structural damage after earthquake disasters. However, the deployment of high-quality sensors on each floor is not always practical. An effective solution is the deployment of low-quality sensors using denoising methods to suppress the measured noisy signal. In traditional filtering-based denoising methods, careful manual selection of filter parameters is required, which is not only complex to implement, but also incapable of dealing with records with high-level noise. To address this problem, a Generative Adversarial Network (GAN) denoising method called DeGAN is proposed. The results for the testing set revealed that DeGAN was more efficient at denoising high-level noise compared to the Discrete Wavelet Transform (DWT)-based method. Furthermore, a new dataset which contains simulation noise and real noise, and a set of shaking table experimental data were utilized to evaluate the denoising performance and robustness of DeGAN. The results demonstrated that DeGAN outperformed the DWT-based method, UNET method and ResNet method in terms of the SNR of the denoised data.