Research On Channel And Feature Selection Algorithm Of MI-BCI

Brain Computer Interface(BCI)is a system that delivers human electroencephalogram(EEG)to the computer for signal recognition and processing and sending out corresponding control instructions,so as to realize the communication between the human brain and the computer or other electronic devices.There is redundant or irrelevant to motor imagery tasks information between EEG channels.At the same time,high-dimensional features are prone to overfitting problems.This paper studies the channel and feature selection algorithms of EEG signals based on correlation,sparse representation and Fisher ratio for the motor imagery(MI)based BCI to improve the classification accuracy of motor imagery EEG.The main research work of this paper is as follows:(1)Aiming at the situation that motor imagery EEG signals contain irrelevant or redundant information channels related to MI tasks.A channel selection method based on correlation and sparse representation was proposed for EEG classification in this paper.Firstly,the Pearson correlation coefficient of each channel of the training sample was calculated to select the significant channels.Then the filter bank common spatial pattern features of the region where the significant channels were located were extracted and spliced into a dictionary.The number of non-zero sparse coefficients obtained from the dictionary was used to characterize the classification ability of each region,and the significant channels contained in the significant regions were selected as the optimal channels.Finally,the common spatial pattern(CSP)and support vector machine(SVM)were employed for feature extraction and classification respectively.In the classification experiments of MI task with BCI competition III dataset Iva and BCI competition IV dataset I,the average classification accuracy reached 88.61% and83.9%,which indicates the effectiveness and robustness of the proposed channel selection method.(2)In view of the fact that decomposing EEG signals into subbands related to brain activity tasks will increase the feature dimension and easily lead to over fitting problems.A feature selection for MI method based on optimal channels and frequency bands was proposed for EEG classification in this paper.Firstly,the optimal channels were selected according to the correlation coefficient,and the 4-40 HZ broadband was decomposed into multiple overlapping subbands by bandpass filter.The subbands with higher power spectrum density were chosen for feature extraction.Next,the pair-wise relevance was utilized to remove subbands features with less difference.And then the screened subbands features were combined with features extracted from the broadband signal.The Fisher ratio was exploited to carry out further feature selection.Finally,SVM was trained to classify the selected features.An experimental study was implemented on the above datasets.The average classification accuracy reached 89.33% and 84.08%,which shows the rationality and effectiveness of the proposed feature selection method.(3)In this paper,we collected the left and right hand motor imagery EEG data of eleven healthy volunteers,and the EEG channel and feature selection method proposed were implemented on the EEG data processed by the eeglab toolbox.The average classification accuracy reached 60.37% and 64.53%,which shows that the proposed method can improve the classification accuracy of the MI-BCI effectively.