[A modified fast independent component analysis and its application to ERP extraction].

Feature extraction of event-related potentials (ERP) plays an important part in both basic and clinical researches for cerebral neurophysiology. ICA is a method for separating blind signals based on signal statistic characteristics. In this paper, the fundamental principle, the discrimination condition and the practical algorithm of Independent Component Analysis are discussed. Then, a fast Independent Component Analysis algorithm (Fast ICA) is introduced. But like Fast ICA, its convergence is dependent on initial weight. We bring in a revision factor into the algorithm; thus the new algorithm could implement convergence on a largescale. In this paper, the revision factor is calculated by gradient. By modifying kernel iterate course, several iterations of Fast ICA are merged into one iteration of Modified Fast ICA, so the convergence of ICA will be accelerated. Finally, Modified ICA is applied to ERP extraction. The simulation shows that the convergence speed can be increased by using the improved algorithm.