Research On Feature Extraction And Classification Algorithms Of Motor Imagery-based Brain-Computer Interface Based On EEG Using A Small Training Set

In motor imagery(MI)-based BCI based on electroencephalogram(EEG),when the subject spontaneously executes a specific imagined movement,the evoked EEG signals are used to directly control the external electronic device in real time.This BCI technique is mainly applied in the field of medical rehabilitation to help people recover their damaged motor functions.Moreover,it plays an important role in robot control,military,and entertainment,etc.At present,long calibration time is needed to get recognizable EEG signals for the classifier.However,the mental fatigue caused by long-term training may decrease the performance of the BCI system.To shorten the calibration time,by means of the combination of theory and experiment,cross-subject transfer learning feature extraction and semi-supervised feature learning were respectively proposed and testified using the small training sets with classification accuracy 70%as benchmark.Main contributions are as follows:(1)Considering high inter-subject invariability,we focused on the changes of the spatial filters before and after transferring the training samples from the specific source subject into the training samples from the target subject,researched the impacts of the training sets from best or all source subjects on the target subject.Based on cosine similarity with weighted sources,an iterative regularized common spatial patterns(RCSP)feature extraction approach(iRCSP-CSW)and a non-iterative RCSP feature extraction approach(RCSP-CSW)were respectively proposed in a supervised way.The experimental results showed,when the percentage of the training set from the target subject was 20%,the proposed algorithms could obtain better robustness and extract more distinctive feature vectors than the common spatial patterns(CSP)approach and other traditional RCSP approaches.(2)To avoid high computation load of cross validation in RCSP approaches,it is assumed that the reference point in the Riemannian manifold can adjust the difference between the subjects.Therefore,a generic subject TL feature extraction approach and a specific subject TL feature extraction approach were respectively proposed in an unsupervised way based on the Riemannian tangent space.The proposed approaches can not only avoid cross validation,but also generate the tangent vectors without using the labels of the training samples.For commonality,the generic subject approach set the same weight for the tangent vector sets from different subjects.For specificity,the specific subject approach used the sequential forward search(SFFS)method to select the tangent vector sets from the best source subjects for the target subject.The experimental results showed,when the percentage of the training set from the target subject was 10%,the specific subject approach could get better robustness and inter-class separability than the competing approaches.(3)Considering the effective use and unknown distribution of the testing set,based on the graph-based model and the transductive support vector machine(TSVM)model which belong to the semi-supervised learning algorithms,an improved TSVM approach(ITSVM)and an improved self-training TSVM approach(IST-TSVM)were respectively proposed.First,a limited training set and a large testing set from the target subject were used to generate the comprehensive features which consisted of the CSP features and the geometric features.Then,the concave-convex procedure approach was used to solve the optimization problem of the objective function under a new constraint which can tackle unknown distribution of the testing set.In addition,IST-TSVM made full use of the testing set in the phase of feature extraction and the phase of feature learning based on a new confidential rule.The experimental results showed,when the size of balanced training set was 10 or the size of unbalanced training set was 20,IST-TSVM obtained better efficiency,accuracy and robustness than support vector machine(SVM)and the traditional TSVM approaches.(4)EEG signals from six subjects who imagined movements of left-hand or right-hand were collected.Based on the BCI illiteracy phenomenon,six subjects were divided into bad subjects and good subjects with classification rate 70%as the boundary.The transfer learning and semi-supervised learning approaches in this thesis were compared and evaluated in terms of the training set size,distribution and the subject group.The experimental results showed,using small balanced training sets,iRCSP-CSW totally outperformed the competing approaches for bad subjects.In addition,using small balanced or unbalanced training sets,semi-supervised IST-TSVM showed its superiority for good subjects.The experimental results and analysis provided a quantitative basis for further study of different calibration time reduction schemes for different kinds of subjects and laid a good foundation for practical application of motor imagery.