The Recognition Of Multi-class Motor Imagery EEG And Its Application In The BCI

Brain-Computer Interface(BCI) technology doesn’t depend on normal brain and muscle output brain pathway. It can provide a new man-machine interactive in order to help the physical disabilities who have normal thinking, the related research has important scientific significance and application value. BCI is based on electroencephalogram(EEG) signal that collects from brain, BCI can realize direct communication and control between brain and peripheral equipment. The scalp EEG signal reflects the change of mental state in real time, so it has been widely used in non-invasive BCI system. The key of BCI is characteristic parameters that reflect subjective consciousness of subjects, and convert them into system instruction. However, multi-channel EEG signal have the low spatial resolution, signal-to-noise ratio and obvious individual differences, these negative factors bring great difficulties for us to effectively analyze and accurately decode EEG signal. Therefore, signal processing and feature extraction technology of EEG are important for BCI system.In order to improve recognition rate of motor imagery BCI(MI-BCI) system, the paper focuses on feature extraction of motor imagery EEG(MI-EEG) based common spatial pattern(CSP) and independent component analysis(ICA) algorithm. Some distinctive research works which have been finished are as follows:1. We design experiment paradigms of left hand, right hand and foot motor imagery, many subjects participate in this experiment and collect rich MI-EEG data, so it builds a good foundation for classfication method and feature extraction.2. We focus on classification of multi-class motor imagery which uses approximate joint diagonalization matrix(JAD) and the one-versus-the-rest(OVR) based CSP. When traditional JAD designs spatial filter, the important is to choose key feature vectors. Common method selects the eigenvectors corresponding to the highest score eigenvalues. However, according to this choice criteria, different classes may select the same eigenvectors, which will result in the failure CSP spatial filter and the lower classification accuracy. So a new method based traditional JAD is proposed, it can automatically select effective eigenvectors. Next, the classification of motor imagery based OVR is analyzed in detail and achieves good results.3. ICA and ICA detection filter are studied, a new method that the optimization design of ICA spatial filter(ICA-SF) with the multiple sub-band features of EEG signal is designed. Multi-channel MI-EEG after ICA-SF usually need to choose the optimal movement-related rhythm band in order to further improve classification accuracy, but individual differences results in different movement-related rhythm band for different subjects. Therefore we need to optimize MI-BCI system parameter, including frequency band, optimization design of spatial filter and classifier parameter learning. Then multiple sub-band optimal combination method is proposed. Experimental results reveal that the proposed sub-band method compared with single-band have a higher average classification accuracy.4. Features detection and optimization method of MI-EEG based ICA and genetic algorithm(GA) have researched. For different individual, we use GA to select event-related desynchronization(ERD) band which is induced by motor imagery. If we can pinpoint movement rhythm band for individuals, which obviously improve the recognition rate of MI-BCI. The non-stationary of EEG and individual differences have the disadvantage of seeking the optimal rhythm enhance band(REB), which affects the stability of ICA. Therefore a new classification method is proposed that the feature optimization of GA based on ICA. Not only the method has better reliability and practicability, but can be used for the design and implementation of online BCI.