Research On The Method For Frequency Recognition Of Steady-State Visual Evoked Potential BCIs Based On Transfer Learning

The brain-computer interface(BCI)bypasses the body’s own neural pathways and converts brain signals into computer instructions,creating a direct peripheral equipment link.A steady-state visual evoked potential(SSVEP)-based BCI can obtain higher classification accuracy when training data is sufficient,or it can also use low accuracy to suppress the training phase at the cost.This dilemma between performance and practicality limits the practical application of brain-computer interface based on steady-state visual evoked potential.Although some researches attempted to conquer this problem,they have not yet established a highly recognized method.In this thesis,we proposed a canonical correlation analysis(CCA)-based transfer learning framework for improving the performance of an SSVEP BCI and reducing its calibration effort.Based on the canonical correlation analysis algorithm,three spatial filters were estimated using the training data of a target subject and a set of source subjects,and two template signals are estimated using the training data of the first two.Using these three spatial filters,the test signal of the target subject from a single experiment is filtered separately with two template signals.Six correlation coefficients between the filtered test signal and template signals are calculated,and combined into a feature signal.Frequency identification is performed using template matching method.To reduce the individual discrepancy between subjects,an accuracy-based subject selection(ASS)algorithm is developed for screening those source subjects whose EEG data are more like those of the target subject.The proposed Intra-and Inter-Subject Canonical Correlation Analysis(IISCCA)integrates both subject-specific models and subject-independent information for frequency identification of SSVEP signals.The performance of ASS-IISCCA was evaluated on a benchmark data set with 35 subjects and compared with the state-of-the-art algorithm task-related component analysis(TRCA).The results show that ASS-IISCCA can significantly improve the performance of SSVEP BCIs with a small number of training trials from a new user,thus helping to facilitate their applications in real world.