Evoked Potential Signal Extraction Method

Evoked potential (EP) is an electric signal generated by the central neural system, which is a response of various stimuli, like sound, light and electrical impulse. EP reflects the status and changes of the neural system. It is often seriously contaminated by ongoing EEG and other noises. At present, many signal processing techniques, such as independent analysis (1CA), wavelet analysis, high-order cumulants etc., have been introduced to extract the EP signal. ICA is the emphasis of this thesis.ICA is a new tool for blind source separation (BSS). It is originally developed to deal with the cocktail-party problem. It has caught the sight of a lot of researchers. The present algorithms can be roughly divided into tow categories-the algorithms based on the measures of nongaussianity and the ones based on information theory. This thesis reviews several main algorithms in detail and applied them to the extraction of EP signal. The result of the computer simulation shows in the separation of the EP and the noises.Among the algorithms based on information theory, the Maximum Likelihood, the Infomax and the Mutual Information are three basic criteria, and they are equivalent. All the three criteria can be related with the Kullback-Leibler (K-L) divergence, which is one of the basic concepts of information theory. The invariability for invertible transform of K-L divergence is the key of the equivalence. This thesis proves the equivalence theoretically.Combining the advantages of the two categories of algorithms, this thesis proposes a new algorithm named as FastInfomax for the blind source separation. The FastInfomax algorithm comprises of two steps. First, utilizing the rICA algorithm to obtain the initial weights; second, executing the online learning through the Infomax algorithm. The computer simulation shows that the Fastlnfomax algorithm can speed up the convergence significantly.