Spatial Feature Extraction Of EEG Signals Based On Motor Imagery

As an advanced technology in interdisciplinary,Brain-Computer Interface（BCI）enables brain to communicate with the outside world without the assistance of muscle tissue,which provides a new means of communication between brain and the outside world.Motor Imagery（MI）is a common BCI paradigm.BCI system takes the signal generated by brain during motor imagery as input,and the signal is processed into corresponding machine commands and expressed by external equipments.In the signal processing of BCI system,feature extraction and classification are the key steps to transform brain information into machine commands.As a result,this thesis focuses on the methods of feature extraction,and uses the extracted features for classification.Cooperative interactions among neuron groups scattered in adjacent brain regions emerge in brain cognitive acts.Therefore,this study speculates that there is certain relationship between the electroencephalography（EEG）signals of adjacent channels during motor imagery,and proposes a Channel Binary Pattern（CBP）method to explore local spatial information.By encoding the interactions between EEG signals in adjacent channels,CBP mines local spatial patterns and extracts local spatial features.However,since CBP ignores global information when extracting local spatial information of EEG signals,this study proposes a global spatial filter based on maximizing the variance of each class to extract global spatial information,and performs information fusion of local and global spatial features based on classifiers.This thesis implements experiments on a public dataset.Compared with another method of local spatial feature extraction,the classification results of CBP are significantly improved.And the average classification accuracy after information fusion is 8.8%higher than that of CBP alone.The classification results confirm the effectiveness of spatial information extraction at different scales（local and global）and the feasibility of information fusion.The available information in EEG signals is not only embodied in spatial relation,but also contained in temporal sequence.Due to the difference in subjective consciousness of subjects or external interference in the motor imagery period,different subjects would have different time periods unrelated to motor imagery.So,a Sparse Temporal Segment Common Spatial Pattern（STSCSP）method is proposed to eliminate the redundant information in time sequence and extract reliable time-space features for each subject.STSCSP divides the whole time period of MI task into several overlapping sub-periods,and extracts spatial features for each sub-period.Furthermore,7)_2,1 norm and7)_1,2 norm is used to mine time information from the spatial features.This survey implements experiments on three public datasets by STSCSP,and the classification accuracies are84.2%,90%and 91.5%,respectively.Experimental results show that STSCSP is superior to other comparison methods and capable of extracting time-space information effectively.