Abstract
Motor imagery electroencephalography (MI-EEG) decoding enables brain-computer interface (BCI) systems to convert self-generated motor intentions into external control commands. However, inter-subject variability and EEG non-stationarity cause distribution shifts that limit the cross-subject generalization of existing decoding models. Although existing domain generalization methods can learn shared representations without target-subject data during training, they mostly rely on feature alignment or data augmentation and insufficiently exploit complementary discriminative knowledge among source subjects. In response to this problem, this study proposes TSMD-Net, a Two-Stage Multi-Domain Distillation Network for cross-subject MI-EEG domain generalization. TSMD-Net distills knowledge from multiple domain-specific heads into a unified leader head. The proposed method adopts a progressive two-stage optimization strategy, where feature covariance alignment first establishes stable cross-domain representations, and multi-domain collaborative distillation with hyperbolic supervised contrastive learning then enhances cross-domain prediction consistency and class discriminability. Experiments were conducted on two motor imagery benchmarks, namely BCI Competition IV 2a and 2b, and on the larger-scale High Gamma Dataset for motor execution. TSMD-Net achieved mean accuracies of 70.21%, 77.82%, and 79.53%, respectively. These results indicate that TSMD-Net provides effective and generalizable cross-subject motor decoding across the three evaluated EEG datasets.