Adaptive Blind Source Separation in Multi-Channel Systems

Diploma thesis (596 KB pdf)

The purpose of Blind Source Separation (BSS) is to recover a set of latent independent source signals from observable signals that are generated in a mixing process as superpositions of these very source signals. For this task, only the mixtures are available, whereas both the statistical properties of the original source signals and the details of the mixing process are unknown.

The field of application of BSS includes not only tasks in the area of biomedical sciences, computer vision, and telecommunications (blind channel equalization), but also the separation of acoustic signals.

This diploma thesis starts with discussing the information-theoretic concepts, the methods for parameter estimation, and some issues from nonlinear optimization theory that are needed in BSS. In this context, the statistical model of Independent Component Analysis (ICA) is introduced as a powerful tool for performing BSS, relying solely on the statistical independence of the source signals. Based on several different approaches (maximization of non-Gaussianity, maximum likelihood estimation, minimization of mutual information, diagonalization of the cumulant tensor), various implementable algorithms for adaptively estimating the ICA model are derived, compared and contrasted.