A Transfer Learning Method for Motor Imagery EEG Signals Classification Based on CCSP and Riemannian Tangent Space Mapping

Motor imagery (MI) based Brain-Computer Interface (BCI) system controls external devices by decoding Electroencephalography(EEG) signals recorded over the scalp. However, due to the nonstationary of MI-EEG signals, traditional BCI system usually requires current users to perform huge amounts of training tasks to obtain enough labeled samples for classification model construction, which is time-consuming and laborious. Aiming at this problem, this paper proposes a transfer learning algorithm, which helps to train the classification model of target subject with the help of other auxiliary subjects' samples. Firstly, we use the KL divergence to calculate the similarity of covariance matrices between target subject and each auxiliary subject, and thus we can determine the weight of each auxiliary subject's covariance matrices to construct composite covariance matrices. Secondly, we can obtain feature vectors through the spatial filter calculated by common spatial pattern(CSP) method using composite covariance matrices and then we transform these feature vectors into Riemannian manifold. Thirdly, we use tangent space mapping method in Riemannian geometry to project covariance matrices onto the tangent plane. Finally, we use SVM and LDA classifier to classify the projection vector. We prove the effectiveness of the proposed method by testing its performance on data_set_IVa of BCI Competition III and EEG data of five subjects collected through eego™ mylab 64 ES-300 produced by ANT Neuro Technology.