OUTPUT-ONLY SYSTEM IDENTIFICATION USING BLIND SOURCE SEPARATION

Abstract

The increasing complexity of engineering structures has raised the requirement of continuous monitoring for their proper functioning. Structural health monitoring needs reliable response measurements, robust system identification algorithm for its analysis to assess the state of the structure. The process involves postulation of a model for the structural system followed by the estimation of the model parameters based on the analysis of output and/or input data. With advancement in communication technology and affordable computing, the vibration-based system identification techniques have been receiving attention to this end. Earlier approaches to structural system identification were based on minimization of the prediction error of the response of identified model. Direct estimation of the system parameters in physical space (namely, stiffness, mass and damping) has also been studied within prediction error minimization framework, or within Bayesian framework. Other methods for system identification in modal space based on the estimation of frequency response functions in frequency domain, or impulse response functions in time domain have also been developed. Various single input single output, single input multi output, multi input multi output based modal identification techniques have been developed in the past. However in come cases the input excitation are not known or difficult to measure, (e.g. ambient vibration excitation) output only system identification are used where only response measurements are available for estimation of system parameters. We focus on output-only modal identification in this work. Random decrement (RD) technique, natural excitation technique (NExT), stochastic subspace identification (SSI) and frequency domain decomposition (FDD) are some of the most commonly used output-only modal identification procedures. The random decrement, natural excitation and stochastic subspace identification techniques are not suitable for identification of base excitation problems where all response measurements include a component of base motion. Frequency domain decomposition often leads to several spurious modes which are often very difficult to distinguish from true structural modes. Recently, blind source separation (BSS) procedures based on certain statistical properties of signals have been used for modal identification. i The objective of BSS procedure is to extract the source components or modal responses from the measured data for further estimation of modal parameters such as natural frequencies, modal damping and mode shapes. Statistical BSS procedures like independent component analysis (ICA), second order blind identification (SOBI), complexity pursuit (CP) and their derivatives involve identification of modal responses within a scaling factor. Because of this scaling of modal response, the estimation of mode shapes is a complex affair in these procedures. Moreover, these procedures are sensitive to the algorithmic parameters as well as the assumptions about the statistical properties of the signals considered for source separation are not always valid. Motivated with the limitations of statistical-BSS based procedure, a non-statistical timefrequency based synchrosqueezed transform (SST) procedure for BSS is proposed in this study. It is shown that using acceleration measurements there is a greater chance of identifying higher modes than with displacement/velocity measurements. The decomposition of recorded response time histories in the time-frequency plane improves the quality of identification and allows extraction of modal components (sources) in true strength/intensity to the extent of their contribution in making up the total response. This allows for developing a simple procedure for estimation of mode shapes and modal damping, which avoids tedious computations required for other formulations. The mode shapes are estimated from the ratios of the extracted harmonic components from the response measured at different floors. A simplified procedure of estimation of modal damping is proposed. Due to decaying ground motion amplitudes towards the end of an event, the tail portions of the recorded response primarily comprise of free vibration response and a simple curve fitting of an exponentially decaying envelope to the tail portion of extracted modal source provides the estimate of modal damping. For validating the performance of proposed modal identification several example (for different building model and different earthquake excitation) cases have been considered. This study, focusses on modal identification and tracking changes thereof as damage indicator. In vibration based damage identification techniques a prior knowledge of modal parameters is required, to serve as a benchmark for comparison. Often, the lack of availability of this type of data can make a method impractical for certain applications. To address this issue a moving time window based damage identification procedure is proposed. Modal identification is performed in a number of non-overlapping time intervals by using synchrosqueezed transform based blind identification scheme. The time windows are selected based on the energy distribution of the signal in the time-frequency plane. As the ii window moves forward, the changes in modal parameters of the structure are tracked to detect changes in the state of structure. The change in flexibility of the structure is tracked by using a rank-1 update reduced rank approximation. Numerical simulation studies are presented on UCLA Factor Building for different earthquake events to demonstrate the validity of the proposed method.