Modified Fast ICA for Blind Signal separation

Abstract

The Fast ICA or fixed-point algorithm is one of the most widely known algorithm for Blind Signal Separation (BSS) in terms of accuracy speed and computational complexity. Two versions of the algorithm are made available in literature: a one unit (deflation) algorithm and a symmetric algorithm. Both algorithms provide four standard contrast functions namely Skew, Pow3, Gauss and Tanh for selection to have better performance. In general, the selection of the contrast function depends on the data and the application. In Biomedical applications Electro Cardio Graph (ECG) data is problematic due to heart rate variation and several interferences such as muscle activity, respiration , thermal/electronic noise and noise from electrode-skin contact that corrupt the signal. We propose modified Fast ICA algorithm employing different data-adaptive contrast functions that are obtained from the underlying pdf of source signal distribution. We test the performance of modified Fast ICA algorithm for ECG application ECG data set is simulated and then added with four different types of noise distributions, white Gaussian, Flicker, Impulse and Rayleigh. We compare performance across the standard contrast functions in Fast ICA and modified Fast ICA using Signal Mean Square Error (SMSE) and Signal Noise Ratio (SNR) of estimated sources.. We show that the performance of modified Fast ICA is superior to standard contrasts in case of noisy ECG data .