Research On EEG Signal Domain Adaptation And Classification Based On Transfer Learning

The development of brain-computer interface technology has subverted the way people and the outside world exchange information to some extent.The brain-computer interface system does not rely on peripheral nerves and muscles,and can realize information exchange and control operations of smart devices only by recognizing the subject’s subjective intentions.Motor imagery Electroencephalogram(EEG)signals have been widely used in the field of brain-computer interface due to their non-invasiveness and easy acquisition.As a new type of human-computer interaction,the brain-computer interface system based on motor imagery EEG signals has great application prospects in rehabilitation,life and entertainment,vehicle transportation and military security.Due to the distortion of temporal and local information of brain signals and differences between subjects,the cross-session and cross-subject classification of motor imagery EEG signals remains challenging to tackle.This thesis proposes a rotation alignment domain adaptation method with Riemannian mean to reduce the difference in data distribution between different subjects.The method uses covariance matrix to represent data feature of the EEG signals,and achieves data alignment by rotating the symmetric positive definite matrix in Riemannian space.In this process,the proposed algorithm extends the traditional transfer component analysis to a matrix form in order to function in the Riemannian framework.This method does not require data labels and is unsupervised.Experimental results show that the algorithm proposed in this thesis improves the cross-session and cross-subject classification accuracy of motor imagery EEG signals.Aiming at the cross-domain classification problem in the field of brain-computer interface,this thesis proposes a coordinate alignment algorithm based on Riemann space and a novel parameter transfer method.The proposed methods reduce the difference in data distribution between different subjects,and make the target classification parameters as close as possible to the classification parameters of the source domain subjects which have a similar feature distribution to the target subject.Therefore,when the target subject has only a small amount of labeled data available,good classification accuracy can still be achieved.Through experimental comparison,the optimal classification performance can be obtained by combining the two algorithms.In addition,an intelligent wheelchair steering system is designed based on the proposed algorithms,which has broad application prospects in the field of intelligent medical care.