Brain–computer interface (BCI) technology facilitates information exchange between the brain and external devices, offering new approaches to human–computer interaction and medical rehabilitation. However, distribution shifts caused by inter-subject variability in electroencephalogram (EEG) signals limit classification performance and generalization. To address this issue, DWDANet, an EEG classification framework integrating multi-scale spatiotemporal feature modeling with dynamically weighted domain alignment, is developed. Multi-scale convolutions extracted spatiotemporal features, while a Transformer encoder modeled global dependencies. Multi-source domain alignment with dynamic weighting enhanced collaborative modeling across source domains and improved generalization. In within-subject evaluations, DWDANet achieved accuracies of 87.22%, 90.32%, and 68.82% on the BCI IV-2a, BCI IV-2b, and SHU-MI datasets, respectively. In cross-subject experiments, it achieved accuracies of 70.61%, 81.43%, and 64.38% on these datasets, respectively, outperforming other existing methods on all three. These results indicate that DWDANet could provide an effective approach to improving motor imagery EEG classification performance and cross-subject generalization.
