Study On Classification Algorithms Of Brain-computer Interface For Motor Imagination

Brain-computer Interface(BCI)provides the direct interact between the human brain and external devices.Researches in it are significant not only in medical or biological engeering,but also in entertainment,life,military and other aspects.Among them,the BCI system for motor imagination is a typical spontaneous BCI system and has received extensive attention.However,it is still insufficient for the decoding accuracy of EEG signals calculated by the existing methods,and on the other hand,it takes too long time for training.Therefore,the BCI system is still in the preliminary experimental stage.In view of the above-mentioned two problems,this paper studies the relevant decoding algorithms of EEG signals in the motor imagination BCI system from the perspective of experimental analysis and practical a pplication.(1)Decoding of EEG signals in experimental analysis: Generated from the generation mechanism of EEG signals,EEG signals in motor imagination BCI system have specific spatial and frequency patterns.Currently,the mainstream EEG signal processing methods can be summarized as European space and Riemannian space.Among them,algorithms in European space have gradually developed from spatial filtering to spatio-spectral coupling filtering.While the algorithma in Riemannian space hasn’t be explored sufficiently.However,Riemannian space can better describes the internal connection between high-dimensional EEG signals.Therefore,this article proposes a spatio-spectral coupling filtering algorithm in Riemannian space(RSSF).At the same time,due to insufficient utilization of the eigenvalues in the existing riemannian distance,this paper proposes a robust classification algorithm based on the LASSO model.And the results in the self-built three-class dataset and public datasets,have proved the superiority of RSSF in EEG signal processing and classification.(2)EEG signal decoding in the practical application: As a typical spontaneous BCI system,the subject needs a long period of training before the application of the motor imaging BCI devices.Therefore,this paper proposes an active learning method based on data augmentation(DAAL)under small samples,as well as proposes corresponding data augmentation algorithms for Riemannian space and Euclidean space.Under the verification of multiple data sets(with only 5 samples in training set),the accuracy of traditional algorithms with DAAL in the offline system has improved by nearly 10% compared with thenselves.In the online system,the DAAL can still maintain its superiority when the active learni ng performance deteriorates severely.It further proves that DAAL is very robust.Secondly,as a learnable framework,DAAL can achieve better performance by combining different sampling methods.In summary,this paper proposes a spatio-spectral coupling filtering algorithm in Riemannian space for experimental analysis,and proposes an active learning method based on data augmentation under small samples for application scenarios.Experimental results show that the algorithm s have superior classification performances and lay the foundation for subsequent online system research.