Motor Imagery EEG Signal Classification Based on Multi-Teacher Knowledge Distillation

The lightweight design of motor imagery electroencephalogram classification models is crucial to promote braincomputer interfaces towards practical applications. However, traditional lightweight methods often sacrifice the expressive power of the model, resulting in significantly degraded performance in motor imagery electroencephalogram decoding tasks with large individual differences. To overcome this limitation, this study proposes a novel architecture that combines multiteacher knowledge distillation with efficient neural networks. The framework enhances a lightweight student model using multi-source domain data by fusing deep spatiotemporal representations from multiple complex teachers, thereby providing rich complementary supervision signals. This strategy not only compensates for the limited representational power of the student model through knowledge transfer, but also significantly reduces the training resource requirements through structural optimization. Experimental results demonstrate the effectiveness of the proposed method on BCI Competition IV Dataset 2b and stroke patient dataset CBCIC. The results show that the best accuracy of $75.7 \%$ and $73.3 \%$ are achieved on dataset 2 b and CBCIC datasets respectively, which outperforms some state-of-the-art methods.