A novel transfer algorithm based on time convolution and EEGNet for motor imagery classification

EEG signals have great individual variability, and there are obvious individual differences in the frequency domain and amplitude of subject’s EEG signal. These characteristics have led to several problems for the cross-subjects application of brain-computer interface. To improve the accuracy for the cross-subject motor imagery experiment, we propose a transfer algorithm based on time convolution and EEGNet network. The CORAL loss function is introduced to the algorithm. The publicly available BCI-2a IV dataset is used to validate the proposed algorithm. Compared to other state-of-the-art algorithms using the BCI-2a IV dataset, the algorithm shows a better improvement in high rate of cross-subject classification of EEG signals. To validate the validity of the algorithm proposed for cross-subjects classification on the small-sample dataset, we design our own experiments on motor imagery EEG data acquisition.