DG-MTLNet: A Domain Generalization Motor Imagery Classification Model With a Novel Temporal Generation Component

Motor imagery is a core paradigm in brain–computer interface (BCI) research. Despite notable progress, the accuracy of electroencephalography (EEG) decoding remains constrained by pronounced intersubject physiological variability. Domain generalization (DG) methods offer a feasible solution as they do not require target subject data during training. However, existing DG approaches are limited by insufficient modeling of temporal dependencies, inadequate exploitation of feature diversity, and overly simplistic training strategies. To address these issues, we propose a novel temporal generation (TGen) component that integrates a temporal convolutional network with a variational autoencoder to capture temporal dependencies in EEG signals. It employs a dual-branch design combining random convolutions with covariance regularization to enhance feature diversity. Building upon TGen, we further construct the DG model, named DG-MTLNet, by incorporating cross electrode convolution and multitask learning. Evaluations on the public BCI competition IV datasets 2a and 2b, the large-scale public datasets OpenBMI and EDSP, and the diverse dataset MI-GS, demonstrate that DG-MTLNet outperforms the best baseline models, achieving mean accuracy improvements of 2.50%, 3.11%, 3.23%, 1.54%, and 3.47%, respectively.