Study On Feature Extraction And Classification Methods Of Motor Imagery Eeg In Brain-computer Interface

The brain-computer interface technology has become a hot spot for researchers at home and abroad. It is a new assistive and control method for the disabled people, but also provides a special way of communication and entertainment for normal people. The control signal of brain-computer interface technology consists of spontaneous EEG and evoked EEG. Because the rhythm feature is obvious, the current evoked EEG-based brain -computer interface technology is relatively mature, but additional stimulation system is needed, so it is relatively little convenience when applied to practical; but for research based on spontaneous EEG, EEG control signals are generated spontaneously. Because of no stimulation, it is easier to the practical application, but the rhythm feature is not obvious, which proposed high demands for feature extraction and classification algorithms. At present, most researches are still in the laboratory stage, the classification accuracy is not high and can rarely used in practice. Therefore, it is very significant for research on feature extraction and classification based on spontaneous EEG brain-computer interface.Motor imagery EEG is a kind of spontaneous EEG. The EEG data used in this paper were acquired from NEUROSCAN system in the laboratory. Three types imagined movement of the motor imagery EEG including holding spring-grip using the left hand and the right hand, stepping on accelerograph using the right foot, were studied. In this paper, firstly, EOG interference signal, which is unavoidable in the experiment, was filtered out by using SCAN4.3 software of NEUROSCAN system. After filtering out EOG interference signal, motor imagery EEG was filtered by using a combination of FFT and IFFT method, which is proposed by author. Then, motor imagery EEG data were carried on average energy analysis, the feasibility of filtering is verified.Two feature extraction methods, which are based on AR model spectrum estimation and discrete wavelet analysis, are analyzed. The spectra indicates that the ERD phenomenon is not clear when AR model-based power spectrum estimation method were used, so a discrete wavelet analysis-based feature extraction method was designed by author in this paper. Using this method, two characteristics of signal, which are energy value and combination of energy value and wavelet coefficients, were extracted. These characteristics, as input of classifier, were used to select the adequate classifier.Three kind of classification methods were described in the classifier design, which are BP neural network classifier, self-organizing neural network classifier and particle swarm optimization-support vector machine. The self-organizing neural network classifier and particle swarm optimization-support vector machine were designed by author in this paper to use as classification method of motor imagery EEG. Classification results show that for the classification feature selection of EEG, using the combination method of the energy value and the wavelet coefficients is better than the single energy values, the former has higher accuracy. For the classification method selection, self-organizing neural networks and particle swarm optimization-support vector machine classification method is better than BP neural network classifier, both the highest classification accuracy rate reaches to 80%. Particle swarm optimization- support vector machine classification has slightly better result than self-organizing neural networks, but the complexity of network was increased. Self-organizing neural networks is simple and easy to implement, and can classify different characteristics of subjects automatically, has strong adaptability. Therefore, the self-organizing neural network classifier is adopted by author of this paper, and on this basis, self-organizing neural network's initial weights set algorithm were changed, the classification accuracy is further improved.Final conclusion: The feature extraction and classification method to three types imagined movement of the motor imagery EEG including holding spring-grip using the left hand and the right hand, stepping on accelerograph using the right foot, were researched by designing the experiment using NEUROSCAN system in this paper. The EEG signal is pre-processed by using the combination of FFT and IFFT method. In feature extraction aspect, the discrete wavelet analysis was used to extract the combination feature of the energy value and wavelet coefficients of EEG signal. In the classification choice aspect, the improved self-organizing neural networks were adopted, and the result shown is better.