Blind Source Separation for Signal Processing Applications

Blind Source Separation (BSS) is a statistical approach to separating individual signals from an observed mixture of a group of signals. BSS relies on only very weak assumptions on the signals and the mixing process (hence the “blind” descriptor) and this blindness enables the technique to be used in a wide variety of situations. Research in the field of Blind Source Separation has resulted in the development of a family of algorithms, known as Independent Component Analysis (ICA) algorithms, that can reliably and efficiently achieve blind separation of signals. Within Blind Source Separation research there are two important problems that are generally considered: instantaneous BSS and convolutive BSS. The difference between these two is based on the nature of the signal mixing process; in essence instantaneous BSS separates signals that are mixed without introducing time delays whereas convolutive BSS can achieve separation when time delays are involved. In this thesis, the mathematical foundations of both instantaneous and convolutive BSS are developed. Once this mathematical framework has been established, the emphasis of the thesis moves to experimental results obtained with ICA techniques. The two primary applications of BSS addressed in this thesis are to acoustic signal mixtures (both instantaneous and convolutive) and to biomedical signal processing. Biomedical applications of BSS are of particular importance to this thesis and, in particular, a novel application of BSS to electromyography (EMG) is proposed and examined. Experimental results demonstrate the effectiveness of BSS in achieving signal separation from EMG data which could potentially have important applications in clinical EMG testing.