Research On Feature Selection And Classification Algorithm Of EEG Based On Motor Imagery

Brain-Computer Interface(BCI)can complete the process of EEG signals at the brain level based on the modern information technology method,and achieve human-computer interaction,it also provides new ideas for the disabled who can not communicate with the outside world normally.One of the challenges about this field is that how to classify the EEG signals effectively by using relevent algorithms.In order to improve the classification and recognition performance of multi class EEG signals,based on the two standard data sets of Berlin IVa and Graz Data Set 2a,this paper aims at researching and analyzing some relevent algorithms which could solve this problem.The main work is as follows:For the two kinds of motion imagination EEG signals of Berlin IVa,in order to solve these problems with high data dimension and low average classification accuracy during the processing of interpretation,this paper first uses respectively Mutual Information(MI),AR power spectrum estimation and CSP ommon spatial pattern to get the time,frequency and electrode parameters effectively.The experimental results show that the above three algorithms can effectively reduce dimension of the signal's features.Then,we use the linear discriminant analysis(LDA)to classify,and find out that the combination of AR-CSP algorithm for feature extraction can further improve the classification accuracy.Compared with the results of the third international brain-computer interface competition in 2005,the classification effect is second only to the first.For the four kinds of motion imagination EEG signals of Graz Data Set 2a,in order to improve the data signal to noise ratio,independent component analysis is applied to blind source separation;for non-stationary nonlinear characteristics,this paper uses wavelet analysis and CSP spatial pattern to get signal's features;and then uses SVM support vector machine to classify.Because of the large amount of workload and the shortcomings of the traditional SVM algorithm in the selection of parameters,this paper combines the genetic algorithm to achieve the fast selection of the penalty factor and the optimal value of the kernel parameters.Considering that GA search is time consuming,this paper implements the parallel computing and model preservation mechanism.Finally,there is a big progress between the former classification results and the new ones.Meantime,compared with the results of the competition,th is paper not only can measure the validity and feasibility of the algorithm from multiple angles,but also the KAPPA value of CSP-GA-SVM combination algorithm adopted in this paper is higher than the first prize of the competition in 2008,the results of Wavelet-GA-SVM classification is just behind the second.The major experimental procedures is given and the results are analyzed,which provide reference for the practical application of EEG signal processing.Finally,this paper based on the BCI2000 platfo rm for virtual cursor movement experiment.Some of the algorithms in this paper are applied in this experiment,and the result achieves the virtual controlling,which can provide reference for the practical application of EEG signal processing.