Research On Classification Of Multi-Task Motor Imagery EEG Signal

The motor imagery brain-computer interface(MI-BCI)completes the generation of control commands by decoding the motion intention of the human brain and ultimately realizes that the external device can act according to human thinking.However,the study on the classification of MI EEG signals still suffers from the problems of low classification accuracy,low recognition number,and high or fixed number of channels,which limit the application of the MI-BCI.To solve the above problems,thesis proposes a classification method for multi-task MI EEG signals,which overcomes the contradiction between the number of channels and the classification accuracy,and achieves higher classification accuracy while using fewer channels.This major research work is as follows:(1)In view of the fact that most of the current studies use fixed or all channels as the algorithm input,which cannot adapt to the neural information of the subjects,thesis proposes a channel selection strategy based on multi-domain information fusion.Firstly,the pre-processed EEG signals are segmented in the time domain and filtered by filter banks to extract different time-frequency domain sub-bands,and in each timefrequency sub-band,the channel contribution rate(CCR)is calculated according to the2 norm and the Frobenius norm,and then the optimal channel combination is extracted.The experimental results show that the proposed strategy in thesis can adaptively generate different optimal channel combinations according to the neural information contained in the subjects’ EEG.(2)Aiming at the problem that the EEG feature vectors extracted by traditional feature extraction algorithms have insufficient characterization capability and only contain single-domain information,thesis proposes an optimization extraction method of one versus one common spatial pattern based on the optimal ordering in the projection space,which is called OVO-CSSP.Thesis also proposes a multi-domain information fusion feature extraction algorithm,which combines OVO-CSSP and Mu and Beta rhythm window energy to extract the EEG signals in each time-frequency subbands.The EEG signals in each time-frequency sub-bands were extracted by combining OVO-CSSP with Mu and Beta rhythm window energy.The experimental results found that the extracted feature vectors have the advantages of strong characterization ability and rich feature information,and can achieve good classification results on the two four-classified motion imagery datasets selected in thesis.(3)The proposed algorithm is verified on the two BCI competitions,the experimental results showed that the average recognition accuracy of the 12 subjects was 77.6%,and the average kappa coefficient was 0.70.Compared with the classification results obtained by the OVO-CSP(One versus One Common Spatial Pattern)algorithm,the classification results obtained in thesis have increased by 24%,and compared with the OVO-FBCSP(One versus One Filter Banks Common Spatial Pattern)algorithm,the classification results have increased up 6.5%.The results of thesis are also higher than those of other recent papers using the same dataset.The proposed algorithm achieves better classification accuracy with fewer lead channels and provides an implementation solution for moving multi-task motion imagery BCI systems toward online applications.