Classification Of Motor Imagery EEG Signals Based On Common Spatial Pattern

The Motor Imagery Brain-Computer Interface(MI-BCI)uses pattern recognition technology to determine the intentions of subjects from the collected EEG signals and then uses the intentions to control artificial limbs or other devices.It is of great significance for disabled people to interact with the outside environment again.Electroencephalogram(EEG)is a popular non-invasive method for brain signal acquisition,the collected signals by which have the properties of low signal-to-noise ratio(SNR),being rich in artifacts,and low spatial resolution.Therefore,accurate classification of motor imagery EEG signals is the key to the MI-BCI system.In MI-BCI,Common Spatial Pattern(CSP)is widely used.By finding a set of spatial filters of EEG signals,CSP maximizes the variance difference between the two classes of samples after spatial filtering,to obtain a highly separable frequency band energy feature.In this thesis,the time and frequency optimization of MI-EEG signal classification based on CSP is discussed.In the MI-BCI system,the separability of features obtained by CSP algorithm is affected by the time window of EEG signal.Specifically,there are differences in brain activity,neural reaction speed,and so on between subjects.So a fixed time window is not suitable for feature extraction from EEG signals.Although some researchers have found this problem,the research on the influence of the time window size on the performance of the MI-BCI system is still insufficient.In this thesis,a method named Multi-scale Time Group Common Spatial Pattern(MTGCSP)based on sparse group LASSO is proposed to extract CSP features corresponding to specific subjects from time windows of multiple sizes for MI task classification.The experimental results reveal that the MI activity in EEG signals can be more accurately detected by extracting and integrating CSP features from multiple time windows of multi-scale window size,thus improving the performance of the MI-BCI system.In addition to time window selection,another important problem is the frequency band of the MI-EEG signal.Although it has been suggested that MI activity in the human brain is concentrated in certain frequency bands,some subjects show band specificity in EEG signals.Plenty of research has been proposed to improve the classification performance of the MI-BCI system by optimizing the frequency band.However,most of the work ignores the correlation between adjacent frequency segments.In this thesis,frequency constrained sparse common spatial pattern(FCSCSP)based on fused LASSO is proposed for feature selection under constrained frequency band variation.After experimental verification,the comprehensive performance of FCSCSP is improved compared with other similar methods.