Blind separation of dependent sources using multiwavelet based on M-GARCH model

Standard ICA (independent component analysis) cannot estimate dependent sources all alone. In these methods, assumptions on sources are non-Gaussianity and independency. In sub-band independent component analysis (SDICA), independency restriction on sources can be relaxed to only have a limited independent sub-bands. In this paper, we propose a new preprocessing to SDICA using multiwavelet based on multivariate GARCH modeling (M-GARCH) with time-varying temporal structure of sub-bands of mixed sources with limited independent sub-bands. By this modeling, time-varying temporal structure effects of sources on ICA performances are eliminated. Sub-bands are calculated by multiwavelet transformations which are suitable tools in signal representation and de-noising because of enhanced feature extraction capabilities and concentration of signal energy in small number of coefficients over wavelet. One subband ICA method is used as a comparison basis to our proposed method.