Nonlinear Identification of Civil Structures using Time and Frequency Domain Data Features

Vibration based structural health monitoring (SHM) approaches are becoming more popular in recent years for damage identification and response prediction of civil structures. A common approach in vibration based SHM is model calibration using recorded data. Calibrated models can be used for damage diagnosis as well as response prediction of linear and nonlinear structural systems. Among different types of models, finite element (FE) models offer advantages to mitigate the modeling errors since they can incorporate the information from geometry and material behavior of structures. Most of the research in this area involves updating linear FE models of structures, but linear FE models cannot reliably be used for damage prognosis since civil engineering structures behave highly nonlinearly under moderate to high amplitude loadings such as earthquakes. The objective of this research is to develop a framework for reliable nonlinear structural identification for robust response prediction and damage identification. The unidentifiability issue of nonlinear FE models is specifically tackled by implementing simpler models, more informative data features, and advanced optimization and stochastic simulation methods. The first part of the thesis is focused on developing a method for identification of time-varying modal parameters using recorded data. The deterministic stochastic subspace identification method is applied on short windows of input-output data for estimating the time-varying modal parameters of nonlinear systems, and it is shown that the identified values are more accurate than the most common output-only methods such as the wavelet transform. In the next parts of the research the recorded time-domain and extracted time-varying frequency-domain data are used for calibration of nonlinear models for two complex real-world structures, namely a three story infilled frame and a seven story shear wall building. The effects of modeling errors and used data features on the performance of the calibrated models are studied. In the last part of this thesis, the nonlinear model calibration is performed in a probabilistic framework. The maximum a-posteriori values of the modeling parameters and their uncertainties are estimated and the effects of using different types of data features on the estimation uncertainty and accuracy of the models are studied.