Studying On The Algorithms Of Feature Extraction And Classification In Brain-Computer Interface

Brain-computer interface (BCI) is one kind of techniques on which more and more interest has been put by the international researchers recently. It has the potential to offer humans a new and innovative communication option with their environment. These systems rely on the acquisition and interpretation of the commands encoded in neurophysiological signals without using the conventional muscular output pathways of the central nervous system (CNS).Being a key issue for implementation of a BCI system as well as improving its performance, signal processing techniques are the main subjects in the plan for basic research program of NUDT "human-machine integrative system based on multi-modal measurement combination". The main research goal of this paper is to increase the classification accuracy of BCI system in the literature, the state of art is then reviewed and summarized firstly, and characters of different EEG components along with various signal processing methods are analyzed and compared. With the datasets of BCI competitions II and III, those methods and aspects which affect the classifier accuracy for EEG signals are emphatic analyzed and studied, and improved our techniques which have been applied in the BCI competition III, by introducing support vector machines (SVM) to design the classifier, better classification results are obtained. For the P300 speller paradigm of BCI Competition, the result on testing set revealed a classification accuracy of 3% higher than what we submitted to the competition, whereas for Motor imagery paradigm, the accuracy was the same as the top score. Moreover, SVM is superior to linear discriminant analysis (LDA) in the sense of "curse of dimensionality" and "overfitting". Another catergory of signal which can be made use for BCI system called ECoG, is also studied in this paper. ECoG is better than EEG in the senses of spatial resolution, signal noise ratio and bandwidth. Less complicated algorithm combining only common spatial pattern (CSP) and SVM without using continuous wavelet transform revealed accuracy for classification on the testing set 6% higher than what we submitted to the competition, which exceed the top score ever achieved in the competion. Detailed algorithms and results of the above mentioned approaches are described in the paper. Perspective on future BCI system and our later work are discussed.