Learning Structures : fusing deconvolution-based seismic interferometry with Bayesian inference for structural health assessment

Abstract

Monitoring vibration responses of civil structures is crucial to the assessment of their health status and reliability against natural hazards. In this study, we present a two-step computational methodology for structural identification and damage detection via fusing the concepts of seismic interferometry and Bayesian inference. Firstly, a deconvolution-based seismic interferometry approach is employed to obtain the wave-forms that represent the impulse response functions (IRFs) with respect to a reference excitation source. Using the deconvolved waveforms, key structural characteristics that correspond to the current state of the structure (e.g., shear wave velocity) can be extracted. Changes in these features can be used as a qualitative damage metric (e.g., to determine if the structure is damaged). We study the following two different damage detection methods that utilize shear wave velocity variations: (1) the arrival picking method (APM) and (2) the stretching method (SM). Secondly, a hierarchical Bayesian inference framework is employed to update a finite element model minimizing the gap between the predicted and the measured time histories of the IRFs. We employ a sequential Markov Chain Monte Carlo (MCMC) sampling to obtain a baseline structural model. Through the comparison of the model parameter distributions with the baseline information, we show that the damage localization and quantification is possible. We initially test our procedure utilizing the synthetic records of a 10-story shear type building. Despite high noise contamination, identification results realized through our approach for both stiffness and damping parameters show good correlation with their true values. For further deployment, we analyze the shake-table experiment dataset that contains various damage scenarios. We show that the variations in the shear wave velocity can be used for qualitative/quick damage detection, and that the velocity reduction is more evident for the more severely damaged states. We then update our FEM by the presented Bayesian learning framework by utilizing the extracted IRFs of the experimental structure. Induced damage, i.e. bolt-loosening on the first floor, affects the posterior distributions quite noticeably. Finally, the structural damage detection problem is addressed by studying an experimental data set of full-scale seven story building slice, that was progressively damaged via previously recorded historical earthquake records utilizing the Network for Earthquake Engineering Simulations (NEES) shake-table. Our results indicate that the developed framework is promising for monitoring structural systems. It allows for non-invasive determination of structural parameters.