Feature Extraction Research Based CSP In Brain Computer Interface

BCI(Brain-Computer Interface) is a brand-new control path of foreign exchage ofinformation, which is achieved by the establishment of connection between human brainand the outer environment via computers or other electronic equipments, independent ofperipheral nerve and musculature, and it is a means for the patients, who suffer frompartial loss of muscle control, to communicate with the outside world. It is anon-muscular communication system, which enables a direct communication betweenhuman intention and the outer environment. Since feature extraction of EEG signal is thecritical issue of BCI, this thesis focused on the intensive study of the method of featureextraction of EEG singnal, trying to enable the correct analysis of BCI system in theextraction of the user’s intention via brain as well as the EEG signal of the surroundingtissues.Firstly, this thesis put forward the method of feature extraction of the spatialfrequency pattern, which aimed at avoiding the defeciency of the great impact of bandparameters on the classification performance of CSP mathod, and it tries to figure out theoptimal spatial pattern in accordance with the influnce of frequency mode on motorimagery. With the utilization of the contradiction between equilibrium time resolution ofcontinuous wavelet analysis and frequency resolution, a frequency mode is embeded onthe basis of the spatial pattern, and CSP method is extended to polytpes, realizing thesimultaneous optimization of the spatial pattern and the frequency mode.Secondly, an algorithm of common spatial patterns is studied basing on Kullback-Leiblerde.In order to extract non-stationary EEG signal feature, using KL distanceto define the maximized difference of the categories and the discriminant model of theminimized class difference, it selects the r andν of the optimal parameters, uses thecomplete training data to drill KLCSP fiiltering vessel, and evalutes with the test data. Itis prooved that this algorithm is completely driven by data, free of any append records.Finally, a feature extraction method of the common spatial patterns, which based ontensor calculus, is put forward. This method is a supervised tensor method, which utilizes the advantage of tensor’s simultaneous multidimensional processing, and tensor anaysisis applied to CSP algorithm. It diagonalizes High-dimensional covariance tensor ofmulti-class EEG，and it tries to figure out the maximized discriminat model of twosamples or multiclass samples. It is prooved by the experimental results as well asanalyses that common spatial patterns, which is based on tensor calculus, decreases thenumber of parameters needed by data modeling and enhances classified performance.