In a recent decade, blind source separation (BSS) has been received a great attention because its applicability to various field such as wireless communication, medical engineering, image restoration and speech enhancement. BSS is equivalently called ICA (independent component analysis) to extract independent signals from a given mixture space. This dissertation presents analysis on extant algorithms and proposes new methods to overcome difficulties not solved with the existing algorithms.
Chapter 1 introduces definition, mathematical description, applicable fields and research trend of BSS. Chapter 2 reviews the methodologies of existing approaches. Chapter 3 presents stability conditions and asymptotic error covariance of the general BSS algorithms. Chapter 4 provides four new methods to cope with troubles under different environments and assumptions: (1) source signals have arbitrary distributions, which compel us to adjust the nonlinear functions, (2) measurement signals are less than the source signals, (3) the number of source signals increases or decreases and (4) some of source signals are nonstationary. In Chapter 5, we propose a new fast algorithm for convolutive mixtures of nonstationary signals such as speech signals based on block processing and FFT.