Applications Of Machine Learning In Brain-Computer Interface

BCI(Brain Computer Interface)is a communication or control system which can establish direct real-time interaction between human brain and external devices.BCI system can directly communicate or control the external devices by analyzing the EEG(Electroencephalo-graph)signals without using the peripheral nerves and muscles,which creates new pathways for paralyzed patients with motor ability severely impaired.In this thesis,the EEG signal processing in the BCI systems based on machine learning is mainly studied.For BCI in the left and right movement being not the same,there exist problems,such as long time training,low classification accuracy,and large error of predicting the trajectory of the left-hand and right-hand.So in this thesis,major researches are focused on the research on improving the BCI system control based on the corresponding algorithms in machine learning.The main contributions and innovations include the following aspects:(1)As left-hand and right-hand move at different speeds in experiment,an enhanced CSP(Common Spatial Pattern)based feature extraction and machine learning classification algorithm is proposed as well as the corresponding data classification model.The enhanced CSP method can achieve relatively high classification accuracy by extracting only four features with training time significantly reduced.While,six features are necessary to be extracted in the traditional method to achieve the same accuracy compared with the enhanced CSP method.In this part,the decision trees,Bayesian classifier,and SVM(Support Vector Machine)are used in this thesis for data classification,and the accuracy of each algorithm is relatively high.But considering the limitations of a single classification,an adaptive classification algorithm based on LDA(Linear Discriminant Analysis),decision trees,Bayesian and SVM is proposed.The data presented by the Shanghai Jiaotong University with left-hand and right-hand different movement and the data derived from the left-hand and right-hand imaginary movement experiments by the Medical University of Berlin are used to verify the classification accuracy of the proposed algorithm,and it is proved that its accuracy is higher than a single classification algorithm.(2)In the experiment of predicting the finger movement of macaque,a forecast model,based on continuous wavelet transform and a prediction model of neural networks,is proposed.By this way,the trajectory errors by using the proposed model is smaller than the results using the original ways.