Research On Steady-state Visual Evoked Potential Brain-computer Interface Based On ALPHA Information Transfer Method

Electroencephalography(EEG)has become the most widely used input signal of brain computer interface(BCI)because of its ease of use,non-invasiveness and low equipment price.Because Steady state visual evoked potential(SSVEP)has the characteristics of high information transfer rate and high signal-to-noise ratio,BCI system based on SSVEP has received extensive attention from researchers.Affected by the background EEG activity or other noise environment,the SSVEP signal of the same subject has the characteristics of non-stationary in different periods or when using different equipment to perform the same task,which limits the plug-and-play of SSVEP-BCI in real life.Aiming at the non-stationarity of SSVEP signals caused by cross-day and cross-electrode,this paper uses the Align and Pool for EEG Headset Domain Adaptation(ALPHA)method of transfer learning to alleviate the problem.This study designs SSVEP-BCI character spelling systems with 12-targets and 40-targets.Sixteen subjects participated in the experiment.Each subject carried out SSVEP-BCI character spelling experiment based on dry electrode and wet electrode on three different days.The experimental tasks performed each time are the same,first 12-target system and then 40-target system.This study uses the method of transfer learning across day and electrode to realize zero training of SSVEP-BCI character spelling system.Firstly,for the non-stationarity of SSVEP signal,the influence of cross day and cross electrode on the non-stationarity of SSVEP signal is analyzed from the perspectives of the SSVEP response and covariance matrix.The similarity results show that both cross day and cross electrode will cause the non-stationarity of SSVEP signal.Secondly,for individuals who use BCI system for a long time,the similar information in individuals(i.e.across days or across electrodes)can improve the performance of BCI and reduce training by using transfer learning.In order to verify the feasibility of intra-individual transfer learning,four transfer directions are analyzed offline,including two dimensions of day and electrode: cross-day wet-to-dry transfer,cross-day wet-to-wet transfer,cross-day dry-to-dry transfer and within-day wet-to-dry transfer.It is found that the method based on transfer learning can maintain stable performance through zero training.Compared with the fully calibrated methods,the method based on transfer learning can achieve better or similar performance in different transfer directions.Finally,based on the consideration of practicability and high performance and according to the results of offline analysis,cross-day wet-to-dry transfer is selected for the online experimental verification.Online experiment further verified that the method of transfer learning.Nine of the sixteen subjects participated in the online experiment of cross-day wet-to-dry transfer.The results showed that the average accuracy of transfer method was 85.97 ± 5.60 % in cross-day wet-to-dry transfer,and the average accuracy of dry electrode based on the fully calibrated method was 77.69 ± 6.42 %.Through transfer learning,the same user can maintain better performance with new period or new electrode.This study lays the foundation for facilitating the long-term usage of the SSVEP-BCI and advances the frontier of the dry electrode-based SSVEP-BCI in real-world applications.