Research On Feature Extraction Of Two-classMotor Imagery EEG

Brain computer interface(BCI)is an interactive system for the brain and the computer,which does not rely on human neural pathways,and communicates with the outside through electroencephalogram(EEG).Because of its wide application value,BCI has become a research hotspot in brain science,medical rehabilitation,automatic control and other fields in recent years.The EEG signals is a non-stationary random signal,extracting an effective data feature set is an important part of the whole brain-computer interface system.This thesis studies feature extraction algorithm of EEG in two-class motor imagery tasks,which mainly includes the following two aspects:(1)Based on two-class EEG of motor imagery and rest,the autoregressive(AR)model of modern power spectrum is used to estimate the spectrum of motor imagery signals.To solve the problem that the actual model order is often unknown,an new method was presented to select the order of AR model based on sampling rate,comparing the new method with other order selection methods,the event-related desynchronization/synchronization phenomenon(ERD/ERS)and the result of T test prove the advantages of the new method.(2)Selecting the left vs right hand motor imagery from BCI Competition,this thesis presents a combination of Empirical Mode Decomposition(EMD)and Adaptive AR(AAR)models to solve a series of problems in feature extraction algorithm of motor imagery data,such as inaccurate description of non-stationary signals,computational difficulties,et al.EEG signals were decomposed into a set of stationary time series called Intrinsic Mode Functions(IMFs),selects imf which have distinct characteristics in motor rhythms(5-28Hz),calculates parameters of AAR model of the selected imf to constructs feature vectors.Based on Linear Discriminant Analysis(LDA)task recognition.The experimental results show that the highest classification accuracy of the two individuals is 84.34%and 85.71%respectively,which is higher than the accuracy of using AAR model parameters as features.Mutual information and Kappa values also show that the method of EMD+AAR have better classification performance,which provides a theoretical basis for online BCI system.