Study On Blind Source Separation And Application In Brain Signals

BSS (Blind Sources Separation) is one of the multiple signal processing method. It has many potential applications especially in communication systems, biomedical engineering, medical imaging, speech enhancement, signal analysis and process control, and feature extract, etc.The work in this dissertation consists of studying blind source separation and its application especially in Brain Signals.BSS may be divided into two categories: Independent component analysis (ICA) based and spatio-temporal structure based. Two basic hypothesizes about sources of non Gaussian distribution and statistically independence are used directly or indirectly in BSS. In the study of sequential ICA, analogy measure of two random variables is presented. Then non-linear function is used to measure the non-Gaussianity of sources. Furthermore, a gradient approach based learning rule and a fixed iteration algorithm are derived. Through the geometrical explanation based on the algorithm of maximizing kurtosis absolute value, it is prove that only one Gaussian signal is permitted in sequential ICA, while extracting the residual sources is not effected. The probability property of separating sources by analyzing the process is given out. In order to avoid the same source re-extraction, the equivalence of deflationary orthogonalization and direct subtract is discussed. It is point out that is impossible to avoid the error cumulation in sequential ICA.In parallel ICA, three statistically independent based ICA approaches, the Maximum likelihood, the Infomax, and the Minimum Mutual Information, are studied by comparing their iteration equations. The results show that these approaches are totally the same although they are derived from different criteria and cost functions. The source of type and the robust of cost function should be considered instead of its variance.A theoretical framework of sequential ICA and parallel ICA is presented based on literature [51], which relies on the dependence, correlation and non Gaussianity of random vectors. Two theorems on the theoretical framework are proposed and proved. Two groups of algorithms, sequential ICA and parallel ICA, are unified under this framework. In summary, when the de-mixture process is finished, the results are I(y) = 0 and = G(y), butwith different processing.An existing adaptive algorithm based on spatio-temporal decorrelation sources is improved to save calculation time. A new algorithm based on sources signal non-stationary isproposed, which has less calculation comparing with existing algorithms. A BSS algorithm based on fractional statistical is proposed to separate a stable distributed signals.Artifacts are removed from 16-channel EEG recorders based on the analogy measure ICA.The dissertation also proposes a method that preprocesses the EP signals contaminated by a stable noises with Infomax and compares with the latency change detection results between the normal DLMS, DLMP and the Infomax preprocessing based methods (p-DLMS, p-DLMP). It is clear from computer simulations that p-DLMS and p-DLMP algorithms detect the latency changes more effectively and accurately than DLMS and DLMP do even when the SNR is as low as -20dB. Further more, the performance of p-DLMS and p-DLMP algorithms are unchanged with different a(0