Research On Multi-Class Motor Imagery EEG Feature Extraction And Classification And Its Application

In recent years,the incidence of stroke in China has been increasing,seriously threatening the health of elderly people.Rehabilitation training based on motor imagery Brain-Computer Interface(BCI)can promote the nerve repair of sensorimotor cortex in stroke patients.Therefore,it is of great significance to carry out the research on motor imagery BCI to restore motor function in stroke patients.However,due to the characteristics of non-stationarity nonlinearity and individual differences in electroencephalography(EEG)signals,there are still many difficulties in the analysis of EEG,especially for multi-class situation.The purpose of this paper is to study the application of multi-class motor imagery EEG classification.And a channel selection method based on Common Spatial Pattern and a multi-domain fusion feature extraction algorithm are proposed,which can effectively reduce feature dimensions and fully extract features,and finally realize the four-class classification of single joint motor imagery EEG signals.The main work is organized as follows:(1)In view of existing curse of dimensionality of multi-channels EEG signals,this paper proposes a channel selection method based on Common Spatial Pattern,which uses the weight value of each channel in the projection space to calculate the contribution rate,and then selects some channels to reduce the computational amount of the algorithm.Then,in order to make up for the lack of time-frequency information of Common Spatial Pattern,this paper proposes a multi-domain fusion feature extraction algorithm,which integrates Improved Local Characteristic-scale Decomposition(ILCD)and Adaptive Common Spatial Pattern(ACSP).ILCD algorithm is used to decompose the selected signals and calculate the time-frequency energy characteristics to form the feature vectors group in the time-frequency domain;ACSP algorithm is used to construct a spatial filter and extract the feature vectors group of the spatial domain.Finally,the multi-domain features are fused,and the highest accuracy rate of four-class classification on BCI2008 Competition Data Sets 2b was 88.7%.The average Kappa coefficient of this algorithm is 0.71,which is better than the results of BCI2008 Competition winners and other algorithms using same dataset.(2)In order to further apply the classification algorithm of four-class motor imagery EEG to the rehabilitation training of stroke patients and meet the actual needs of effective control of different movements of single joint,this paper designs a four-class motor imagery EEG classification experiment of shoulder joint:"forward extension","backward extension","abduction" and "adduction".Firstly,the experimental paradigm is designed and the data is collected;Secondly,the artifacts are removed;Then Time Domain Parameters,Power Spectral Density,Common Spatial Pattern and ILCD-ACSP algorithm are used to extract the features of the signals,and Linear Discriminant Analysis and Support Vector Machine are used for classification.The results show that the combination of ILCD-ACSP feature extraction method and Support Vector Machine classifier achieves a classification accuracy of 63.8%,which is the optimal value in all the combinations.