Multiple least mean kurtosis adaptive filters for blind source separation

Abstract

In this paper, a novel use of adaptive filters for blind source separation is presented. The known independent component analysis algorithm separates signals from their mixtures based on the observation that a mixture of statistically independent signals is more Gaussian than the separate signals. Similarly, an adaptive filter, that is designed to minimize the Gaussianity of its output, relies on the same hypothesis. The proposed adaptive filter uses all the mixture signals, the observation signals, as its inputs—one as the main input, and the rest as reference inputs. The filter is iteratively modified, using gradient descent, such that the measure of non-Gaussianity of its output is maximized, leading to the separation of one source signal at its output. To separate the N source signals from the given N mixtures, N such adaptive filters are used. The proposed method has been successfully applied to the blind separation of multiple signals.