Research On Theory And Algorithms Of Motor Imagery EEG Recognition

The brain-computer interface is dedicated to establish a new direct communication channel between the brain and the external devices.The brain-computer interface system has formed a new brain information transfer pattern,which has been widely applied in the fields of medical rehabilitation,intelligent control,entertainment,etcThe brain-computer interface based on motor imagery means that the subjects can control the external equipment by imagining their limb movements,and it has broad application prospects.But there are several challenges in processing motor imagery EEG signals,e.g.,weak signal,intense interference noise,high dimension,which directly affects the recognition efficiency.Therefore,a key problem in EEG recognition is that how to extract low dimensional features with high discriminability.Nonnegative matrix factorization(NMF)has been applied to EEG feature extraction and dimension reduction which provides meaningful local data.However,NMF is unsupervised and cannot make use of the piror information.As a result,the extracted features contain little class information and achieve poor recognition performance.In this paper,a series of semi-supervised NMF methods are proposed in combination with the label information for motor imagery EEG recognition.A large number of experiments are used to verify the effectiveness of proposed algorithm.The main contents are as follows1.Based on the characteristics of the motor imagery EEG,the spatial and spectral components of the EEG are extracted by the canonical correlation analysis method,and a spectral clustering technology is introduced to classify the extracted features.The spatial and spectral features extracted by this method are helpful to improve the recognition ability of EEG2.We study the recognition performance of motor imagery EEG based on NMF systematically.A label-constrained NMF(CNMF)method is introduced by incorporating the training samples.This method expects that the samples with the same class label have the same representation in the low-dimensional space,beneficial the recognition of EEG.In addition,a dissimilar regularization term is added to the CNMF model to make the EEG signals with different labels have dissimilar representations as much as possible in the low-dimensional space.The experimental results show that the recognition performance of motor imagery EEG can be improved by combining NMF with label information3.The CNMF method maps the training samples sharing the same label to the same point the low dimensional space,it ignores the original properties of data(e.g.fuzziness and complexity).To preserve the original features of the data,a fuzzy costrianed NMF(FCNMF)method is proposed in this paper.By incorporates the label information,this method assume that the training samples is subordinate to the labeled class completely,and a hard label matrix is constructed;And consider the relationship between the labeled samples and other classes,a soft label matrix is constructed to equally distributed the membership of training samples to other classes.Then,it incorporates the two label matrices as additional constraints on the NMF procedure,fuzzing the label constraint.Compared with the CNMF method,the extracted features by the FCNMF contain more original features.In the classification experiment with two public datasets,the proposed method has performed better than the existing studies in the literature,the results show that the proposed method is effective for motor iamgery EEG recognition.4.A label relaxation NMF method is proposed to mine the label information of the motor imagery EEG.If the similarity between the training sample with the labeled class is high,it will be given a large label weight in the label matrix,relaxing the label constraint.To prevent over-relaxation of labels,a dissimilarity regularizer term is incorporated into our NMF model for enlarging the margins of samples from different classes.Also the update rules,the convergence,and the time complexity of the proposed method are analyzed systematically.Compared with the traditional NMF and CNMF,the extracted feature using the label relaxation can better fit the original sample,and a better accuracy is can be obtained with a small number of training samples5.The recognition effect of the proposed method in multi-task motor imagery EEG is analyzed.Four-task motor imagery EEG of six subjects is collected,including the left hand,the right hand,the foot,and the tongue imagery.The recognition performance of the proposed method is tested by using the public and collected motor imagery EEG dataset.The experiment results show that the proposed method outperforms other compared methods.It is not only suitable for motor imagery EEG signals processing,but also can be applied to multi-task EEG recognition.