Increasing the efficiency and efficacy of second‐order blind identification (SOBI) methods

Summary This paper proposes a new technique to increase the efficiency and effectiveness of second‐order blind identification (SOBI) methods by reducing the number of time‐lagged covariance matrices required to produce highly accurate mixing matrix estimates. The technique is based on randomly selecting the time‐lagged covariance matrices as opposed to choosing them sequentially, which takes advantage of a property of independence with regard to the selection of time‐lagged covariance matrices, while simultaneously reducing the correlation between selected covariance matrices. The proposed randomized approach is first applied to undamped and damped sinusoids to demonstrate its effectiveness. The randomized method is then applied to a three‐degree‐of‐freedom linear structure subjected to random excitation. Finally, the proposed method is incorporated into a modified version of the SOBI method and applied to perform identification of the UCLA Factor building from recorded earthquake responses. In each case, the performance of the randomized approach is compared with the traditional sequential selection of time‐lagged covariance matrices, and the randomized approach consistently demonstrates superior performance both in terms of accuracy and efficiency. Additionally, the proposed randomized SOBI approach was easily inserted into the modified SOBI method demonstrating its malleability for SOBI variants and other blind source separation methods. Copyright © 2016 John Wiley & Sons, Ltd.