Feature Extraction And Classification Of EEG Signals Evoked During Motor Imagery

Brain Computer Interface(BCI) has totally overturned the usual way of communication between human and the world. It restores and expands hunman’s physiological functions and the ability of cognizance in a wide range. Thus the combination of human and machine is redefined. BCI could distinguish the specific cognition by collecting the electroencephalogram(EEG) produced by the brain. And it could transfer different cognitions into the signals that control the external devices and communication devices in real time.Motor imagery(MI) EEG is produced by human when the motions of limbs or the other objects are imitated and arisen repeatedly in mind without the actual physical movements. At present, the BCI systems based on MI EEG still have a huge space to be promoted in the matter of the classification performance. Therefore it is full of significance to study how to promote the accuracy rate of classification of the EEG, and decrease the error of classification.Based on the basic neurophysiological characteristics of MI, the paper studies the algorithms of feature extraction and classification recognition in detail. Respectively for less channels acquisition of EEG, multi-channels acquisition of EEG and multi-classification of EEG signals.First of all, based on the systematic understanding of the generating mechanism and characteristics of motor imagery EEG signal, the paper designs and implements some algorithms of feature extraction and classification to the two-class EEG collected from less channels. These algorithms include wavelet packet decomposition, common spatial pattern filter(CSP), bayes classifier, K neighbor classifier, linear discriminant classifier and support vector machine (SVM). The offline classification analysises are made on respectively of the BCI competition data and the autonomous acquisition of EEG data, and we get the ideal classification results.Secondly, as the increasing amount of data in the acquisition of multi-channel EEG,inspired by the traditional compressed sensing sparse matrix and dictionary designing, the paper proposes a classification processing method of MI EEG based on sparse representation classification algorithm. According to the inherent ERD/ERS characteristics of MI EEG,using CSP to extract reduced dimensional features, and calculating the power spectral density of the extracting frequency band, and then letting the feature sets consist of power spectrum density of the training samples be the dictionary to make a sparse representation of the test EEG. Finally, through the calculation of the reconstruction error of between the test EEG samples and all kinds of EEG dictionary, the test EEG sample category labels are ultimately determined. The results show that the sparse representation classification algorithm is very effective. Compared to a simple linear classifier, this method significantly increases the pattern classification precision of the multi-channel EEG acquisition.Finally, based on the research above, further study is made about the effect of EOG artifacts to the multi-classification EEG recognition.With the help of EEG analysis platform called EEGLAB, using independent component analysis(ICA) method on the original EEG to identify and remove EOG artifacts. Experiments show that this method can effectively improve the recognition accuracy of multi-classification. And the third chapter designing classification method is also applied in the classification,the results prove the feasibility of the classification method in multi classification EEG recognition.