Research On Brain Signal Classification Based On Transfer Learning And Riemannian Manifold

Brain-computer interface(BCI)is a technology that decodes brain signals collected by sensors into neural activity information,so as to realize the purpose of controlling external devices or information exchange.This technique has extensive research and application value in the fields of human rehabilitation and nerve function.Affected by individual differences and non-stationary of brain signals,the classification model trained on the existing data will lose accuracy in the prediction of other subjects,while the collecting of brain signal is very costly.Therefore,using limited small sample data to improve the model generalization ability has become the focus of research in the field of BCI.Based on the theory of transfer learning and Riemannian manifold learning,the problems of data domain distribution mismatch and small sample dataset existing in BCI are studied.The main research is as follows.This paper presents an improved subspace alignment algorithm.The algorithm of subspace alignment is to design the Bregman divergence between the source domain and the target domain as the objective function,and solve the corresponding transformation matrix to carry out alignment transformation between them.The drawback of this method is that it does not utilize the label information of the source domain data.Therefore,this paper proposes to draw the Fisher discriminant term into target function to improve the classification separability while reducing the gap of domain distribution.The algorithm can effectively improve the ability of the model to predict new target data.In a cross-subject experiment,the traditional methods and other transfer learning algorithms were compared,and the average classification accuracy was improved by 7.73% and 5.53%,respectively.In the case of small number of target samples,a transformation method of Procrustes analysis in Riemannian space is proposed to be applied.The method matches the data of source domain and target domain on the statistical distribution,and then extracts the Riemannian geometric correlation features for classification.The effectiveness of this method has been proved in the cross-subject EEG decoding experiment with comparison of transfer learning algorithm.To solve the problem of high dimensions of EEG data,the spatial filtering algorithm in Riemannian tangent space is applied.In this method,the dimensions of EEG data are reduced during feature extraction,and the main information of feature is retained,so as to prevent problems such as the performance degradation and long computation time.