Motor Imagery And Epileptic EEG Classification Based On Hierarchical Extreme Learning Machine

As the advantage of BCI(Brain-Computer Interface)is more and more prominent,much attention from researchers has been paid to its development.Because of the ease of access,EEG is favored by researchers and has become one of the most important way to obtain information of BCI.In order to improve the efficiency of EEG classification problem and to meet the requirements of accuracy and response speed in the practical application of BCI,this paper focus on the classification of EEG signals based on ELM(Extreme Learning Machine),and the specific research work is as follows:Firstly,this paper presents a method of EEG identification based on HELM(Hierarchical Extreme Learning Machine).Considering the characteristics of the complexity,non-stationarity and low signal to noise ratio of the EEG signals,PCA(Principal Component Analysis)and LDA(Liner Discriminate Analysis)are chosen to extract low-dimensional and discriminative features from EEG signals firstly.Finally,HELM classifier is used to classify and identify the features.The testing speed and classification accuracy of the network are improved by the deep level of feature reextraction and the sparsification for features.In the BCI Competition II data set Ia,this paper compares the proposed method with the state-of-the-art methods corresponding to the same data set.The experimental results show that the proposed method achieves higher accuracy and faster classification speed.Secondly,in order to improve the stability of HELM,this paper presents an improved algorithm,KHELM(Kernel Hierarchical Extreme Learning Machine).By using the kernel mapping idea from SVM(Support Vector Machine)and Kernel-ELM(Kernelbased Extreme Learning Machine),the kernel function is used to replace the output of hidden layer and weaken the influence of stochastic weights on the network output in the original HELM,so as to improve the stability of the network.In this paper,the classification performance of KHELM on BCI Competition II data set Ia is tested.The results show that the introduction of kernel function does improve the stability of the classifier,meanwhile,the proposed method can further improve the classification speed for motor imagery EEG signals.Thirdly,considering the convenience,the proposed methods in this paper are integrated into the EEG signal processing system V2.0,which is programmed with Matlab.Meanwhile,the methods are applied to the real epilepsy EEG data set.The experimental results show that HELM and KHELM have achieved better performance.