Bayesian wavelet packet denoising for structural system identification

Abstract

Non-parametric system identification has been widely applied in structural health monitoring and damage detection based on measured response data. However, the presence of noise in the measured data significantly affects the accuracy of structural system identification. A dilemma is that it is not possible to know with any measure of certainty whether and how much the measured data are corrupted by noise. This paper develops a Bayesian discrete wavelet packet transform denoising approach and investigates the effects of noise in the measured data on structural system identification. The denoising approach is based on the integration of Bayesian hypothesis testing and wavelet packet analysis. It avoids the arbitrary selection of threshold required in classical wavelet thresholding methods and considers the uncertainty of noise, thus resulting in more accurate denoising result. Both original and denoised data are used to investigate the effect of noise on structural system identification through error analysis, R2 statistic, and p-value analyses. The methodology is validated using both simulated data and experimental data. A non-parametric system identification method, the fuzzy wavelet neural network model, and experimental data from a 5-storey test steel frame and a 38-storey test concrete structure are employed to investigate the effect of noise on system identification. A comparative study demonstrates that the proposed denoising approach outperforms the wavelet soft thresholding methods. The results of this research provide a robust methodology to denoise the measured data for accurate structural system identification. Copyright © 2006 John Wiley & Sons, Ltd.