Research EEG Recognition Algorithm Based On Motor Imagery

Brain-Computer Interface (BCI) as a new man-machine mutual means, completely subverts the original methods which human use to communicate with the outside world. It uses the EEG information which is "interpreted" by the computer and equipment to manipulate the outside world, abandoning the original human peripheral nerve and muscle system. The emergence of BCI brings practical value to a number of areas, therefore it becomes a hot topic of multidisciplinary in recent year.EEG recognition is a core part of BCI, we usually use three experimental paradigms in BCI application, they are Motor Imagery、P300and Steady-State Visual Evoked Potential(SSVEP). P300and SSVEP are induced EEG, but motor imagery is spontaneous EEG, so it has a broader applicable prospect. In this paper, we mainly study EEG recognition algorithm based on motor imagery. When human image different parts of the limb movements, there are different EEG characteristics between different parts of the somatosensory area of the motor cortex, we extract EEG features and classify EEG features, ultimately identify the different classes of motor imagery EEG. This paper makes study of the following aspects about motor imagery.(1) Two classes motor imagery EEG classification based on CSP-SVM. Here we mainly study of three EEG patterns:imaging left hand, right hand and foot movement, and classify two classes of them. We mainly study of feature extraction and recognition based on the multi-lead. The advantage of the multi-lead is that we can get richer feature information in a specific area of the brain, which can effectively improve the recognizability. Here we use common spatial pattern to extract EEG feature, and use SVM classifier to class EEG feature. This method is named CSP-SVM method. First we select the related frequency band, extract the feature by CSP, then get the effective energy distribution characteristics and finally use SVM classifier to classify the characteristics. By this method we get higher classified accuracy. The method is validated through the online BCI system.(2) Feature Extraction methods of left and right hand based on time-frequency correlation. The subject needs to wear more leads in the multi-lead method, it is more cumbersome to operate, so it isn’t convenient to BCI. Here we only use two electrodes to classify the EEG, it improves the convenience of the system. Because individual differences in EEG patterns, some characteristics are not enough, the classification accuracy rate is low, we need to select the time-frequency bands. Here we strike the energy spectrum of the time-frequency bands, use correlation theory to get task-related EEG characteristics, finally classify the feature of removing the strong correlated bands with linear discriminant classifier. The results showed that the classification accuracy using the features from selected frequency band and time segment is higher than that using the normal band power features.(3) Classification of four-class motor imagery EEG based on CSP-HMM. Classify four kinds of motor imagery EEG can help to improve the BCI system communication rate and improve the usefulness of the system. Recently there’re not enough classification methods for four kinds of motor imagery EEG classification. The related research methods are:extracting four kinds of motor imagery features by the CSP, selecting the features by LDA dimensionality, and finally using the Bayesian linear classifier to classifying the features; extracting four kinds of motor imagery AR coefficient features, selecting the features by PCA, finally using the HMM to classify the features. Based on the above two methods, the paper gives the EEG classification method based on CSP-HMM.4s data of motor imagery is selected and four2s sub-data sets are obtained by sliding time window with0.5s step size. CSP is used to extract features from the four sub-data sets respectively. Features from each class are used to train one HMM. Four different HMMs were obtained corresponding with the four classes. Test data would be measured by the four HMMs and will be classified based on the Maximum likelihood obtained from the four HMMs. The results show that CSP-HMM yields better performance than CSP-Bayes.