The Application Of Modern Signal Processing Methods In Eeg Classification

The basic principle of the brain-machine interface is analyzing EEG in certain circumstances, and then computer can judge the people'mind in specific state. Due to the low signal to noise ratio of the EEG, the brain-machine interface which is feeing one of the main problems, that is low classification rate. In order to improve it, we must study the effective feature extraction and classification methods.This paper is for the EEG pattern recognition which consists of three main components:feature extraction, feature selection and classification. In the aspect of feature extraction, because of EEG's characteristic of nonlinear and non-stationary, using Wavelet Package Decomposition to extract relevant frequency band of EEG The average of time, frequency band energy, relatively wavelet energy and wavelet entropy are seen as the eigenvalue of EEG The classifiers are support vector machines and last learning machine. This thesis firstly optimizes the classifier parameters based on the genetic algorithm, obtains the recognition accuracy. In order to get a higher correct classification, increases feature selection on the basis of the previous, thus get an optimization based on the best feature subset selection and classifier parameter optimization. This paper contains three feature selections: feature selection based on Fisher criterion, the mean-standard deviation and genetic algorithm. The joint optimization, which can not only reduce the dimension of feature, also can improve the classification accuracy.The data of the BCI competition2003are analyzed with these methods to test and verify the accuracy and efficiency. Comparing the different methods of this paper, the results show that the joint optimization based on genetic algorithm can get better recognition accuracy, more than90%.