Study On Feature Extraction And Classification Method Of Motor Imagery

Brain-Computer Interface(BCI)is an emerging human-computer interaction technology,and Electroencephalogram(EEG)signal through brain activity to complete the communication and control between human and machine.Motor Imagery BrainComputer Interface(MI-BCI)can convert neural activity signals in the brain into control signal that can be recognized by computers according to limb movement or imaginary limb movement,only rely on brain activity to achieve control task in the absence of external stimuli.Feature extraction and classification of EEG signal are the key to transform human intention into control signal.The thesis aims to improve the classification accuracy of left hand,right hand,foot and tongue four types motor imagery,and the identification method is the study focus.The main contents are as follows:(1)The acquisition and preprocessing of motor imagery EEG signal are realized.This thesis designs a motor imagery EEG signal acquisition experiment,obtains four types of motor imagery EEG signal dataset,and performs the preprocessing of filtering,epochs extraction,baseline correction,interpolated compensation,and artifact removal on the collected data in turn.For artifact removal,the thesis uses Principal Component Analysis(PCA)to reduce the dimension of signal to determine the number of independent components to be decomposed,then uses Independent Component Analysis(ICA)to further analyze the artifact component and remove it.(2)Aiming at the problem of small amplitude and low signal-to-noise ratio of EEG signal,this thesis proposes a feature extraction method combining Discrete Wavelet Transformation(DWT)and Common Spatial Patterns(CSP).First,the frequency band close to the motor imagery activity frequency range is selected by DWT,and in the selected frequency band the energy average of the detail information of EEG signals is taken,and then superimposed as the time-frequency domain feature.Second,the optimal spatial filter is established using the CSP method,and the signal variance value is transformed to obtain the space domain feature.Finally,the serial strategy is used to fuse the time-frequency domain and spatial domain feature,then timefrequency-space multi-domain fusion features were obtained,and the maximumminimum standardization method is used to narrow the numerical gap in the data.(3)For the classification of four kinds of motor imagery signals,Support Vector Machine(SVM)is used as the classification method,and grid search is used to automatically find the optimal parameters of the classifier.The feature vectors containing multi-domain information are input into the SVM classifier,and the classification result is further analyzed,which verifies the method proposed in this thesis has good classification performance.