Study On Feature Extraction And Classification Of Motor Imagery Potential

Brain-Computer Interface (BCI) is a new focus of biomedical engineering in recent years. BCI is a human-computer interactive technology which control peripheral equipment by collecting electroencephalograph (EEG) when the human brain underlying the mental task, without using the muscle and peripheral nerve tissues. Therefore, BCI, which can bring huge convenience to the disabled, has great practical value. Besides, BCI can also be used in many other fields, such as military, entertainment and so on. But nowadays, BCI is still at the phase of experimental research, recognition accuracy and transfer rate of EEG are the key issues which constraints BCI technology from the laboratory to the practical application.In this paper, based on the analysis of the physiological characteristics of EEG, we focus our study on the feature extraction and classification method for motor imagery EEG, trying to improve the effectiveness of the features and the accuracy rate of classification.Firstly, we use the weighted rhythmic component extraction (WRCE) method for motion imagery potential feature extraction. WRCE linear weighted the data of all channels by the purpose of enhancing the focused frequency band, and suppressing the other frequency band. Because WRCE don’t filter out any frequency component of the signal, so, this method has overcome the dropout shortcoming of the filter method. By comparing the classification accuracy of WRCE feature and common spatial pattern (CSP) feature, WRCE made batter classification result. Secondly, base on the research of classification method, we use BP neural network (BP-NN) and Fisher linear discriminant analysis (Fisher-LDA) for the classification respectively, the classification result show that BP-NN do not perform well for the small training data set. At the same time, the BP-NN need more classification time than Fisher-LDA for the same dataset, prove that Fisher-LDA is a better choice for BCI which require highly for real-time property.In this study, we use Data set Ⅲ of BCI competition2003for the validation of the algorithm researched in this paper, and this data set was provided by Graz University of technology. The results show that the method in our study achieved87.85%of accuracy rate in classification of the data set. and this result was relatively satisfied.