Research And Optimization On CSP-based Combined Feature Extraction

Brain-computer Interface provide a controlling pathway between human brain and environment based on electroencephalogram,linking human brain with computer or other devices,enlarging humans' ability.However,the BCI signal source is a wake,nonlinear and instable signal varied with the time changing.As a result,an effective feature extraction method is the key to improve the pattern classification accuracy.This thesis firstly summarizes the modern EEG analysis methods in motor imagery,and then emphasizes different feature extraction algorithms.Especially,many improved theories and combinatorial algorithms based on principle classic methods have been described in detail.According the review,since normal Common Spatial Pattern method is restricted to the abundant input channels and lacking frequency information,this thesis proves a novel feature extraction method combined the Common Spatial Pattern method with the Empirical Mode Decomposition.Firstly,The EMD-CSP method was proposed to decompose the EEG signal into a set of stationary time series called Intrinsic Mode Functions(IMF).These IMFs were analyzed with the band-power to detect the valuable IMFs with characteristics of sensorimotor rhythms(5-28Hz),and then the improved CSP filter based on the energy variance was attached to the feature extraction of screening IMFs.Finally,Once the feature vector was built,the classification of MI was performed using Support Vector Machine(SVM).The results obtained show that the EMD-CSP allows the most reliable features and that the accurate classification rate obtained is 92% and the maximum one is 93.8%,which confirm the feasibility and availability of this method in portable further BCI systems.Additionally,the EEG experiment based on Motor Imagery is designed with two different task paradigms to select the better one.In order to achieve optimized feature sets,on the one hand,the advanced CSP filter is further improved with S transition,and this method is respectively combined with EEMD,Bi-spectrum analysis,achieving union features and increasing the final classification rates.On the other hand,parameter optimization in classifiers,choices between various classifiers and feature dimension reduction are used to shorten the time cost during classification process.The research results not only verify the feasibility and efficiency of these combined-feature sets based on CSP and other algorithms with BCI data sets and experimental data sets,but also show that when using the LDA classifier,the Bispectrum-CSP feature is the best optimization among the other ones with the highest classification accuracy and the shortest pattern recognition time.