The Study Of Online Classification And Recognition Algorithm For BCI Based On Motor Imagery

Several classification and recognition methods were widely applied to analyzing motor imagery based brain computer interface(MI-BCI),such as linear discriminant analysis(LDA)and support vector machine(SVM)etc.However,these algorithms are mainly used in the offline analysis,which can not capture the time varying features of EEG and the changes of subjects' states.Thus,in this master thesis,we proposed a new online framework for the classification and recognition in MI-BCI.Moreover,based on this framework,we further developed the algorithms based on bagging and stacking fusion,respectively.The main works are described as follows:1.According to the time-variation of EEG signals and the states of subjects that were sensitive to inner factor and external environment during the MI task,we proposed an online classification and recognition framework in this study.The framework can track the states of subjects and the time-variation of EEG signal by updating and retraining classification model to identify current testing data,resulting in a higher classification accuracy.The update of training set mainly included two steps,i.e.,the expansion of training set and deletion of old samples.In addition,in view of the great influence on the performance of classification and recognition algorithm when the new samples were added in the training set,we developed a relatively reliable method for expanding the training set.The proposed method will select relatively reliable samples according to the threshold value acquired by fitting class-conditional probability of the classifier based on Gaussian mixture model(GMM).2.Though this study adopted a relatively reliable method to expand training set,we couldn't make sure that the new added samples are completely correct or not contain outliers,which might have a strong impact on classification and recognition algorithm based on single classifier.Therefore,some unreliable samples are inevitable and may generate negative impact.This work made attempts to solve the negative impact of unreliable samples through online classification and recognition algorithm based on stacking fusion for the first time,which can improve the generalization performance of classifier.3.Further,we tried another online classification and recognition algorithm based on bagging.Compared with the stacking based algorithm,the bagging based one couldnot only improve generalization performance but also decrease the occurrence of unreliable samples in training set because of the bootstrap sampling method adopted during the training of base learner.In addition,in order to improve the performance and reduce the time consumption of the bagging algorithm in prediction,this paper utilized genetic algorithm to select the best subset of base learner in bagging.Finally,we applied these algorithms to the IVa data-set of the third international BCI,IIa data-set of the fourth international BCI and MI-BCI data of our laboratory.The results showed that the performance of online classification and recognition based on stacking was superior to traditional SVM in the IVa and IIa data sets.However,the performance of online classification and recognition based on bagging was obviously superior to SVM in all of the three data sets.