Study On Transfer Learning-based Motor Imagery Brain-computer Interfaces

The state of the brain changes all the time,so the electroencephalography(EEG)signal is nonstationary.Before using a brain-computer interface(BCI),new users have to go through lengthy training and collect EEG data to construct the classification model of a BCI.This limits its practicability.Transfer learning is an effective way to reduce training time in BCIs.By transferring the previous user’s EEG data to a new user as training data and constructing his/her classification model,the new user’s training time can be reduced largely or even eliminated.However,how to effectively transfer data from previous users to a new user poses a huge challenge.In the BCI system based on transfer learning,the EEG data of different users often have different feature distributions due to the large individual differences among subjects.If the data of previous users are directly transferred to a new user and used to create its classification model without any transformation,the accuracy and robustness of the classification model will be compromised greatly.In this paper,a method based on hybrid Riemannian and Euclidean space data alignment(REA)is proposed.A mixed reference matrix is obtained by regularizing two reference matrices yielded in Riemannian and Euclidean spaces respectively.The EEG signals are aligned so that the data of the source user and the target user have similar feature distribution.The results show that the REA method is better than the two single-space alignment methods in reducing individual differences,and the classification model is better in performance.To solve the problem of how to find the common features of EEG signals of different subjects in the process of applying transfer learning to the BCI system,this paper proposes a method of selecting related subjects based on a modified sequential forward floating-point search algorithm.This method sequentially adds or deletes subject data from the source subject subset.As long as the classification result of the current subset is better than that of the subset evaluated at this level,the subset will be retained;otherwise,the subset will be deleted.This process continues until the classification accuracy of the target subjects is no longer improved.Compared with the four existing algorithms,the performance of the sequential forward floating-point search algorithm is much better than that of the algorithm without transfer,and there is no loss of information.Finally,this paper combines the REA algorithm with the related subject selection algorithm to construct a transfer learning model of MI-BCI.In this paper,a REA algorithm is used to align the data of source subject and target subject respectively,to change the distribution of EEG signal of source subject.Then,the sequential forward floating-point search algorithm is used to select the source subject.Finally,the classification model of the target subject is constructed in Riemann and Euclidean space by using the selected source subject data.The results show that the combination of the two methods has better classification performance than a single method,so the proposed method based on transfer learning will promote the practical application of BCI based on MI.