Pattern Recognition Of EEG Based On Support Vector Machine

Brain-computer interfaces (BCI) provide a direct communication and control channel for sending messages and instructions from brain to external computers or other electronic devices. Using the non-muscular channel, subjects with severe neuromuscular function can directly express their thought and manipulate the external device without using human language and actions. This will greatly enhance the ability of these subjects to manage external event and will improve their living quality. Brain-computer interface technology is an interdisciplinary technology integrating neurology, signal collection, signal processing, pattern recognition and more other relevant technique.The dataset in the experiment was recorded from a subject during a feedback session. The subject needs to control a feedback bar by means of imagery left or right hand movements. So how to map the Electroencephalography (EEG) signals to directions of the movements is very important here. So data processing and pattern recognition was mainly concerned in our BCI research.In this thesis, we firstly introduced the rationale, basic framework and the development of the BCI and the major limitation in the BCI implementation and application. Secondly, the method of wavelet analysis, Support Vector Machine (SVM), and Fuzzy Support Vector Machine (FSVM) was reviewed and used in direction recognition of the motor imagery experiment. The wavelet was used to extract features from the EEG signal, which was collected in the motor imagery, then the SVM and FSVM is used to classify the EEG feature matrix. Thirdly, some appropriate wavelet and kernel function was chose for SVM classification, and the results of SVM classification were illustrated and compared with the result of Back Propagation Neural Networks (BP- NN). Lastly, how the size of training data set affects results of classification was analyzed, and the online classification was summarized.The results show that the EEG signal could be effectively classified by SVM. The accuracy of FSVM is better than simple BP NN when using small training data set. The approach of FSVM greatly improves the generalization ability of SVM classification and its application area, which also diminishes the affection of outliers and noises. The calculation rate of FSVM is so high that the classification could be used in the online BCI research in the future.