Design of extraction algorithms based on second order statistics exploiting beamforming techniques

Signal extraction methods are becoming increasingly popular due to lower computational demands and less restrictive requirements than source separation algorithms. Many existing signal extraction algorithms extract interesting signals based on some known features of the sources. However, immediate extraction of the desired signal is not guaranteed, leading to inefficient and ad hoc deflation techniques. We present a design strategy for efficient signal extraction algorithms. First, by incorporating some amount of prior information in the form of a guess of either the autocorrelation function or the mixing column of the desired source, immediate identification of the desired extraction filter is guaranteed. Second, for a parameterized mixing system new techniques for the design and evaluation of signal extraction algorithms have been developed. These techniques are used to ensure immediate extraction of the desired signal by exploiting knowledge on physical parameters. The design procedure is flexible in the use of a priori information and leads to extraction algorithms that are robust to noise, deal with incomplete prior information, and handle modeling errors. Furthermore, the extraction algorithms can be used to identify extraction filters with different objectives. The design procedure and the properties of the extraction algorithms are evaluated by examples and experiments