Markov and semi-Markov switching of source appearances for non-stationary independent component analysis

摘要

Independent Component Analysis (ICA) is currently the most popularly used approach to blind source separation (BSS), the problem of recovering unknown source signals when their mixtures are observed but the actual mixing process is unknown. Real-world signals often have such difficult non-stationarity that each source signal abruptly appears or disappears, which potentially degrades the performance of ordinary ICA especially in noisy situations. To address such non-stationary cases, we have proposed a non-stationary ICA method, called Switching ICA, based on a special type of hidden Markov model. In this article, we present new experimental comparison of our method to some existing methods, with a full treatment of parameter estimation including those for the parameters that have previously been fixed. In simulation experiments using artificial source signals, the proposed method exhibited performance superior to existing methods, especially in the presence of noise. We also propose a simple semi-Markov extension of the original Markov one, to avoid unrealistic assumption implied in the Markov model, that is, the probability of state duration decreases exponentially with its length. The semi-Markov model is demonstrated to be more effective for robust estimation of the source appearance.