A Fusion Transfer Learning Method of Motor Imagery EEG Signals Based on Riemannian Space

Abstract

Due to the low signal-to-noise ratio and non-stationary characteristics of EEG signals, traditional brain-computer interface systems often need to perform long-term training for the current subject to obtain enough labeled samples in order to establish reliable classification models. In general, the training of the current subject's classification model can be aided by using labeled EEG samples from other subjects through transfer learning, thereby reducing or even eliminating calibration time. At present, most transfer learning algorithms applied in brain-computer interfaces still require a small number of labeled EEG samples from the current subject, and the classification accuracy is not high. The Riemannian space-based motor imagery EEG signal transfer learning fusion algorithm proposed in this paper does not require the current subject's labeled EEG samples and has a high classification accuracy. This algorithm adopts the covariance matrix as the EEG data characteristics, the source domain and target domain in Riemannian space alignment data distribution, through the joint distribution fit method to reduce the data distribution of the differences between different objects, and is suitable for the target domain classification model. Experimental results show that the algorithm can significantly improve the cross-session and cross-object classification accuracy of motor imagery EEG signals.