Recognition Method Of Multi-Task Motor Imagery For Brain-Computer Interface

Brain-Computer Interface(BCI)is a human-computer interaction technique,which can create a connection between brain and external devices.BCI can decode the intention of human brain into instructions for controlling external machines.Because of the extremely high application value and broad application prospects in the fields of medicine,military aviation,entertainment and life,BCI technology has become the research focus of the academic community in recent years.At present,the common BCI implementation methods are mainly P300 potential,motor imagery and steady-state visual evoked potential,among which motor imagery can generate EEG signals spontaneously without external stimulation,and is regarded as the most popular BCI implementation technology for now.The existing realization methods of motor imagery have the disadvantages of low classification accuracy,small number of categories,and poor system adaptability.To resolve above problems,this thesis conducts the research on recognition method of multi-task motor imagery for BCI.The main contents are listed as below:(1)In terms of data time period extraction,this thesis employs the time-domain parameters combined with Fisher distance to resolve the problem of different optimal imagination time period for each subject,in which the time-domain parameters of the original EEG waveform for each subject are calculated,and then the time period with largest Fisher distance is selected,and finally the optimal motor imagery time is selected automatically.(2)In terms of channel selection,two methods are introduced in this thesis based on maximum ratio of energy and entropy and Pearson correlation coefficient.The former expresses the information quality of each channel by calculating the ratio of the average energy and entropy though several experiments for different channels,in which the importance of the channel can be measured by the ratio.The latter can extract the correlation coefficient between different channels,and the channel with the larger coefficient represents the main channel.As for the threshold selection,the method combining mean and minimum values is used to automatically select the number of channels to reduce redundancy and computational complexity.(3)As for feature extraction,this thesis uses the filter bank common spatial pattern with multi-bandwidth for frequency band and spatial filtering as to resolve the problemof different subjects with different optimal band.Then,the selection method of optimal individual feature based on mutual information is used to automatically select the optimal sub-bands for different subjects.To determine the number of features,4,8 and 12 features are selected.Experiment result shows that the best performance can be achieved by using 12 features.(4)In the stage of classification,this thesis converts the four-class problem(left hand,right hand,foot,tongue)into six one-versus-one multi-class structures.Then,Linear Discriminant Analysis,Support Vector Machines and Naive Bayes are utilized as the classifiers,respectively.The performance comparison by experiment analysis are clarified.The 2a dataset of the fourth BCI competition for the four-class motor imagery classification experiment is employed in this thesis.It is demonstrated that the Pearson correlation coefficient has better performance in channel selection.Meanwhile,selection of 12 features with Support Vector Machine as classifier can achieve the optimum classification results.Experiment results show that the highest Kappa value is 0.81,and the average value is 0.61,which proves the feasibility of the proposed method used in this thesis.It should be stressed that the method employed in this thesis has the advantages of high transmission rate and small time delay,which is very suitable for online implementation of BCI.