Increasing the Efficiency of Blind Source Separation Methods for Improved Modal Parameter Estimation

Abstract

One of the most challenging cases for performing parameter estimation on a structure is in the absence of input data. These systems are commonly referred to as “output-only” and the identification of their modal parameters is known as operational modal analysis (OMA). One of the newer arrivals to OMA is a family of methods referred to as blind source separation (BSS). These methods are powerful tools for OMA because they limit assumptions required of measured responses and do not involve transformations to another domain. Second-order blind identification (SOBI) is a popular method within the BSS family, and this method has shown great promise in identification of modal parameters for structural systems. However, SOBI methods operate using several output covariance matrices computed over a series of time lags, and larger systems or longer recordings quickly become cumbersome due to the size and number of matrices required for accurate identification. A new technique is proposed that increases the efficiency of SOBI methods by reducing the number of time-lagged covariance matrices required to produce highly accurate estimates of modal parameters. The technique is based on randomly selecting the time-lagged matrices as opposed to choosing them sequentially. The random selection of covariance matrices greatly reduces the correlation between matrices, which is an issue inherent to sequential selection. The proposed randomized approach is applied to a series of systems, ranging from simple sinusoids to a sample structural model. In each case the performance of the randomized approach is compared to the traditional sequential selection of time-lagged covariance matrices and the randomized approach consistently demonstrates superior performance both in terms of accuracy and efficiency of identified mode shapes. doi: 10.12783/SHM2015/164