To address low decoding accuracy and high computational complexity in multi-task motor imagery EEG decoding models, a dual-branch network-based dynamic knowledge distillation model is proposed. The principle of “multi-scale temporal fusion first, followed by unified spatial processing”, is followed in this architecture and a multi-scale spatiotemporal feature extraction module is designed. A dual-branch structure combining Transformer and BiLSTM are employed as the teacher model, which effectively integrates the Transformer’s strength in global feature representation with the BiLSTM’s advantage in fine-grained temporal modeling, thereby enhancing the decoding accuracy of motor imagery EEG signals. During the knowledge distillation phase, a lightweight EEGNet is utilized as the student model. A dual-branch dynamic knowledge distillation strategy is proposed, where the optimal branch is adaptively selected to generate soft labels. This enables the student model to learn EEG feature representations adaptable to different individuals, allowing it to achieve performance close to that of the teacher model while maintaining a compact size. Experimental validation on the BCI Competition IV Dataset 2a demonstrates the effectiveness of the proposed method. The dual-branch teacher model achieves an average decoding accuracy of 91.32% for multi-task motor imagery. After knowledge distillation, the student model attains an average EEG decoding accuracy of 88.85%, representing an improvement of 6.63% compared to the student model without distillation. Furthermore, compared to the teacher model, the number of parameters in the student model is reduced by 99.78%, and the computational load is decreased by 89.02%. The results indicate that the proposed method maintains high EEG decoding accuracy while significantly reducing computational complexity, providing a practical and lightweight solution for brain-computer interface systems in resource-constrained environments.
WolpawJ R, BirbaumerN, McfarlandD J, et al. Brain-computer interfaces for communication and control[J]. Clinical Neurophysiology, 2002, 113(6): 767-791.
[2]
AltaheriH, MuhammadG, AlsulaimanM, et al. Deep learning techniques for classification of electroencephalogram (EEG) motor imagery (MI) signals: a review[J]. Neural Computing and Applications, 2023, 35(20): 14681-14722.
LuoRuipeng, FengMingke, HuangXin, et al. A review of research on EEG signal preprocessing methods[J]. Electronic Science and Technology, 2023, 36(4): 36-43. (in Chinese)
[5]
LotteF, CongedoM, LécuyerA, et al. A review of classification algorithms for EEG-based brain-computer interfaces[J]. Journal of Neural Engineering, 2007, 4(2): R1-R13.
[6]
SchirrmeisterR T, SpringenbergJ T, FiedererL D J, et al. Deep learning with convolutional neural networks for EEG decoding and visualization[J]. Human Brain Mapping, 2017, 38(11): 5391-5420.
[7]
AngK K, ChinZ Y, ZhangH, et al. Filter bank common spatial pattern (FBCSP) in brain-computer interface[C]//2008 IEEE International Joint Conference on Neural Networks (IEEE World Congress on Computational Intelligence), 2008: 2390-2397.
[8]
ManeR, RobinsonN, VinodA P, et al. A multi-view CNN with novel variance layer for motor imagery brain computer interface[C]//2020 42nd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC), 2020: 2950-2953.
[9]
TortoraS, GhidoniS, ChisariC, et al. Deep learning-based BCI for gait decoding from EEG with LSTM recurrent neural network[J]. Journal of Neural Engineering, 2020, 17(4): 046011.
[10]
XieJ, ZhangJ, SunJ, et al. A transformer-based approach combining deep learning network and spatial-temporal information for raw EEG classification[J]. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 2022, 30: 2126-2136.
[11]
SongY, ZhengQ, LiuB, et al. EEG conformer: convolutional transformer for EEG decoding and visualization[J]. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 2023, 31: 710-719.
[12]
HamidiA, KianiK. Motor imagery EEG signals classification using a Transformer-GCN approach[J]. Applied Soft Computing, 2025, 170: 112686.
[13]
HintonG, VinyalsO, DeanJ. Distilling the Knowledge in a Neural Network[PP/OL].arXiv(2025-05-02)[2025-11-14].2015, 1050: 9.
[14]
LiuY, JiaZ, WangH. EmotionKD: a cross-modal knowledge distillation framework for emotion recognition based on physiological signals[C]//Proceedings of the 31st ACM International Conference on Multimedia, 2023: 6122-6131.
[15]
WuJ, TangB, WangY, et al. A multi-level teacher assistant-based knowledge distillation framework with dynamic feedback for motor imagery EEG decoding[J]. Neural Networks, 2026, 194: 108180.
[16]
LawhernV J, SolonA J, WaytowichN R, et al. EEGNet: a compact convolutional neural network for EEG-based brain-computer interfaces[J]. Journal of Neural Engineering, 2018, 15(5): 056013.
[17]
VaswaniA, ShazeerN, ParmarN, et al. Attention is all you need[C]//Proceedings of the 31st International Conference on Neural Information Processing Systems, 2017: 6000-6010.
[18]
Siami-NaminiS, TavakoliN, NaminA S. The performance of LSTM and BiLSTM in forecasting time series[C]//2019 IEEE International Conference on Big Data (Big Data), 2019: 3285-3292.
[19]
TangermannM, MüllerK R, AertsenA, et al. Review of the BCI competition IV[J]. Frontiers in Neuroscience, 2012, 6: 55.
[20]
LotteF. Signal processing approaches to minimize or suppress calibration time in oscillatory activity-based brain-computer interfaces[J]. Proceedings of the IEEE, 2015, 103(6): 871-890.
[21]
IngolfssonT M, HerscheM, WangX, et al. EEG-TCNet: an accurate temporal convolutional network for embedded motor-imagery brain-machine interfaces[C]//2020 IEEE International Conference on Systems, Man, and Cybernetics (SMC), 2020: 2958-2965.
[22]
HuangJ, ZouJ, YuZ, et al. EEG-MCLANet: a multi-branch CNN-LSTM network for motor imagery EEG decoding[C]//2025 37th Chinese Control and Decision Conference (CCDC), 2025: 1972-1977.
[23]
LuoJ, ChengQ, WangH, et al. MI-MBFT: superior motor imagery decoding of raw EEG data based on a multibranch and fusion transformer framework[J]. IEEE Sensors Journal, 2024, 24(21): 34879-34891.
[24]
HuangQ, YangY, LiJ, et al. AMFTCNet: a multi-level attention-based multi-scale fusion temporal convolutional network for decoding MI-EEG signals[J]. Biomedical Signal Processing and Control, 2025, 108: 107916.
[25]
DuX, LiK, LüY, et al. Motor imaging EEG signal recognition of ResNet18 network based on deformable convolution[J]. Electronics, 2022, 11(22): 3674.
[26]
DuX, XiM, DingX, et al. Motor imagery EEG signal classification based on deformable convolution v3 and adaptive spatial attention mechanism[J]. Biomedical Signal Processing and Control, 2025, 99: 106905.
[27]
MouH, YangW, ZhangS, et al. Hierarchical window attention for motor imagery EEG classification[C]//2024 IEEE International Symposium on Biomedical Imaging (ISBI), 2024: 1-5.