Research On Multi-task Motor Imagery EEG Classification Algorithm

As the intelligent needs of the society continue to evolve,the application areas of the Brain-Computer Interface(BCI)are no longer limited to medical clinical auxiliary rehabilitation,but also extended to industrial,military,people's daily life,entertainment and other aspects,which constantly opened up a new situation for the BCI field and even the artificial intelligence field.However,it is difficult for human EEG signals to be resolved due to their nonlinear and non-stationary characteristics.In additon,the processed results may be unstable according to the specificity of different subjects.In summary,in order to improve the accuracy and reliability of motor imagery EEG pattern recognition,two methods are proposed for the binary-class and four-category MI EEG tasks.1.For the binary-class EEG recognition task,firstly,the EEGs are decomposed based on the Tunable Q-factor Wavelet Transform(TQWT)method,and the sub-band signals are extracted to facilitate subsequent feature calculation.Secondly,the motor imagety EEGs are windowed in order to simulate online environment,and followed by calculating their energy,autoregressive model coefficients(AR)and fractal dimension as time-frequency and nonlinear features.Finally,a linear discriminant analysis(LDA)classifier is designed for classifying.The recognition rate and mutual information are used as indicators for classifier evaluation.Two Graz datasets of BCI2003 and 2005 are employed to verify the proposal.The maximum accuracy of classifying EEGs is 88.11% and the maximum mutual information is 0.95,which performs better than other recent literature results using the same datasets.2.As for the four-category EEG recognition task,primarily,the EEGs are filtered by sliding window method to find out the best ERD/ERS performance time for each subject.Subsequently,the cross-correlation sequences of EEGs are calculated using the "one-to-rest" strategy based on Cross-Correlation(CC)method.In order to preserve the characteristics of the sequences while reducing the feature dimension,mean,median,mode,standard deviation,maximum and minimum of the sequences are calculated as statistical features.Finally,Multiple Discriminant Analysis(MDA),Na?ve Bayes(NB),Support Vector Machines(SVM),and Decision Trees(DT)classifiers are applied to classify these EEGs.The recognition rate and kappa coefficient are used to as the classifier evaluation index.The Graz dataset of BCI2008 is used to verify the proposal four-category MI EEGs classification task of left-hand/right-hand/legs/tongue.The maximum accuracy of classifying EEGs is 86.6% and the maximum kappa coefficient is 0.87,which makes it possible to apply such methods in online enviroment.In summary,this paper not only designs an algorithm for the classic binary-class EEG recognition task,but also obtains some reliable classification results.In addition,considering the scalability of the classification algorithm in practical applications,four-category classifications are also designed.The pattern recognition algorithm of EEG recognition task can theoretically be extended to multi-class EEG recognition tasks,which provides a solution for the development of multi-task BCI system.