| This dissertation is concerned with vibration-based seismic damage detection of tall building structures, based on the measurement dynamic signals directly, rather than on structure modeling and numerical simulation.; For the purpose of studying earthquake response behaviours and vibration-based seismic damage detection, a scaled tall building model structure simulating one typical high-rise residential building in Hong Kong is elaborately fabricated and tested on shaking table. Several earthquake records simulating different soil site condition are designed and exerted on the model structure. The excitation magnitude of such earthquakes is enhanced successively. Following each level of earthquakes, visual inspection on incurred seismic damage is conducted and white-noise exciting tests are performed, providing the basis for the subsequent damage detection.; Generally speaking, vibration-based damage identification usually needs to develop a mechanics model of the investigated structure, for deriving the required dynamic characteristics. However, it is a laborious work to develop a mechanics model for highly complex building structure, which also cannot fulfill immediate post-earthquake damage assessment requirement. Further, apart from modal identification error, structure model introduces additional modeling error, which increases uncertainty gay in model-based damage identification. Therefore, a variety of model-free damage identification or evaluation approaches, based on the measured signals directly, are developed in this study.; Firstly, an approach of damage identification is developed, employing both excitation and response data under white-noise exerting. Considering the availability of excitation information recorded during shaking table tests and possible modal identification errors, the identification is based on the computed frequency response functions (FRFs) directly, rather than on ultimate modal parameters derived from FRFs. Taking obtained FRFs as input, neural networks are constructed for damage location and severity identification. However, high dimensionality of FRF hurdles the neural network training convergence. Using principal component analysis (PCA) technique, feature extraction on the FRFs is executed, where the functionality of PCA for information compression as well as measurement noises filtering is investigated in detail. This part of study mainly explores damage identification potential of using non-frequency or non-mode shape dynamic characteristics, when both excitation and response data are available. (Abstract shortened by UMI.)... |
