Research On Transfer Learning-based Frequency Modulated Visual Brain-computer Interfaces

Under the same external stimulus,there are certain similarities and differences between the EEG responses of different subjects.Because steady-state visual evoked potential(SSVEP)has a high signal-to-noise ratio,the brain-computer interface(BCI)based on SSVEP has attracted great attention from researchers in fields such as neural engineering.The SSVEP signals are dominated by external stimuli,so the SSVEP signals induced by the same stimulus in different subjects show strong similarities.Transfer learning reuses the shared knowledge between two similar domains,and transfers the learned knowledge in the source domain to the target domain,thereby reducing the training cost of the model.This paper uses transfer learning to improve the generalization ability and practicability of the BCI system based on SSVEP.According to the similarity between the EEG data of each subject,this article transfers the data of the source subject to the target subject.And according to the difference between the EEG data of each subject,the source subjects are selected.Two different methods are proposed to select the source subjects,and the results are compared and analyzed.The first method is selecting source subjects based on distance metrics.By calculating the distance between the source subject and the target subject,the degree of similarity between the two is obtained,and the average of the distances of all subjects is used as the threshold for selecting source subjects.The second method is selecting source subjects based on accuracy metric.This paper constructs two different selection algorithms based on accuracy metric.Among them,the screening algorithm 1 first transfers the EEG data of all source subjects to the target subject individually,and then uses the average value of the obtained accuracy as the threshold,and selects subjects in the source domain to obtain a subset of the source subjects.The selecting algorithm 2 first constructs a large number of different source subject subsets,and then transfers the data of all subjects in each subset as training data to the target subject,constructs a classification model of the target subject.The subset with the highest recognition accuracy is the final subset of source subjects.The selection algorithm 2 is combined with the improved extended canonical correlation analysis(CCA)to construct a new CCA(S-CCA)algorithm based on subject selection.S-CCA excludes the training phase of the target subject,thus increasing the applicability of BCI based on SSVEP.In addition,In addition,S-CCA is compared with three different training-free methods.An SSVEP dataset consisting of10 subjects was used for the comparative study.At the same time,the recognition accuracy and the information transfer rate are used as the measurement of the system performance.In all cases,experimental results show that S-CCA is better than the other three methods which are all training-free methods.This study shows that selecting source subjects can effectively improve the performance of SSVEP BCI based on transfer learning.