Semi-supervised Extreme Learning Machine Algorithm For EEG Classification

Brain-computer interface(BCI)can allow its users to control external devices directly which are independent of peripheral nerves and muscles with brain activities.The development of BCI technology has brought changes to many fields,such as medical rehabilitation devices,text input,military application,gaming input devices and so on,and has a broad prospect.However,as a kind of interdisciplinary technology,there are still many problems in theory and key technology,which better models and methods are needed to promote its development.As the most commonly used signal in the analysis of BCI system,Electroencephalogram(EEG)can quickly show the information of brain’s physiological and emotional.However,the EEG signals are expensively labeled with individual difference,and it is difficult to be effectively used by supervised methods.Therefore,the semi-supervised extreme learning machine(SS-ELM)was used as a base classifier in this paper.In order to solve the problem of model promotion and safety problems of unlabeled sample,the dual learning and manifold learning were introduced in this paper which improves the classification and safety control ability of the model,and applies the proposed method to EEG recognition of motor imagery.The main work of this paper is given as follows:(1)In order to reduce the influence of unlabeled samples on the semi-supervised learning,the CR-SSELM method has been proposed which applies to EEG signal classification.Firstly,the dual learning was used to design the risk term of unlabeled sample,which is embedded to the objective function of SS-ELM.Then,the CR-SSELM model was trained through the constraints of manifold regularization terms.Finally,the EEG dataset Competition IV Dataset 2a was used to test,and the results show that the performance of CR-SSELM is still better than the ELM method in 4 of the 9 subjects when the performance of SS-ELM is worse than the supervised ELM method after the unlabeled EEG samples are added into the training.It is shows that the safe strategy of this paper has the ability of sample’s safety control.(2)In semi supervised learning,the performance of classifier depends on the quality of adjacency graph,which can effectively underly the description of sample feature’s distribution.However,the SS-ELM method only uses the similarity between samples to construct the adjacency graph,which may have some problems in the approximation of nonlinear functions and the tracking of sample distribution.Therefore,a balanced graph-based regularized SS-ELM was proposed in this paper.Firstly,the proportion of nonnegative weight vector between the label consistent graph and the sample similarity graph was set properly.Then,the high-quality graph was achieved,which is added to construct the manifold regularization when training the model.Finally,the high-performance classification model was obtained.The proposed method was evaluated on Competition IV Dataset 2a and BCI Competition III Dataset 4a.The experiment results show that the strategy of this paper can effectively improve the quality of the graph,and the performance of classification was significantly better than the SS-ELM algorithm.