Research On Pattern Recognition Of Motor Imagery Brain Computer Interface

Brain computer interface(BCI)technology has important value for the rehabilitation of patients with motor dysfunction.However,BCI technology still faces challenges such as recognition accuracy,recognition action kinds,and real-time performance in current applications.This paper focuses on the pattern recognition in motor imagery BCI,including the removal of ocular artifacts(OA),feature extraction and classification methods for motor imagery electroencephalography(EEG)signal.The main work includes the following aspects:(1)EEG signal is always affected by OA during signal acquisition.To remove OA from EEG,a method based on the combination of discrete wavelet transform(DWT)and independent component analysis(ICA)is proposed.Firstly,EEG signal is decomposed by DWT,the ?,?,?,?,and ? bands of EEG are reconstructed based on the decomposed wavelet coefficients,and the bands containing artifacts are selected by kurtosis value and Renyi's entropy.Then,ICA is used to remove artifacts in these bands,and finally reconstructs the EEG signal after the OA removal.Analysis results of international BCI competition data show that this method can effectively remove OA components in EEG signal..(2)Considering that EEG signal has the characteristics of non-linear and non-stationary,a left-right hand motor imagery EEG classification method with multi-domain fusion based on D-S evidence theory is presented.Firstly,the time domain feature,AR model feature and DWT feature of EEG signal are extracted respectively,and construct probabilistic output support vector machine(SVM)classification model based on the three festure sets.Secondly,the output of each SVM is fused using D-S evidence theory.Finally,determining sample category based on decision rules.Three BCI competition databases are utilized to verify the presented method,and results indicate that this method acquires higher classification accuracy of left and right hand motor imagery EEG signals and has strong individual adaptability.(3)Due to the difficulties of feature extraction and low recognition rate of multi-class EEG signals,a multi-class EEG classification method based on convolutional neural network(CNN)and SVM is proposed,which combines the advantages of self-extraction features of CNN and little sample learning of SVM.Firstly,the CNN for extracting EEG features is designed.Secondly,principal component analysis is adopted to reduce the dimension of the extracted features.Finally,using SVM to recognize action pattern.Four-class motor imagery EEG signals from BCI competition is utilized to verify the proposed method,results show that the proposed method obtain higher recognition rate in the classification of multi-class EEG signals,and the superiority of deep learning.(4)Aiming at the difficulty during ipsilateral hand actions recognition,this paper presents a fusion strategy of EEG and electromyographic(EMG)signal based on the idea of hybrid BCI.Firstly,4 kinds of action's EEG and EMG signal of 3 subjects are synchronously acquired and preprocessed.Secondly,time domain feature,AR model feature and DWT feature of EEG,and integrated EMG feature,sample entropy feature of EMG are extracted.Then,using principal component analysis and sparse auto-encoder to fuse these features.Finally,fusion features are utilized to train SVM for action recognition.Experimental results reveal that the effectiveness of fusion strategy of EEG and EMG in the classification of ipsilateral hand actions.