Abstract
Motor imagery-based brain-computer interface (MI-BCI) technology establishes a connection between human intention and external devices in active rehabilitation. However, obtaining a mass of labeled EEG data is often difficult due to the strict requirement of experimental environment and the necessity for highly cooperative subjects, which makes the application of few-shot learning of EEG classification particularly important. Therefore, we propose a method that combines few-shot learning with triplet metric learning, aiming to maintain strong generalization capabilities of the model with limited samples. First, we pretrain a base model using large auxiliary dataset, and then fine-tune it with a small number of labeled samples from the test subjects to obtain a specific model. During the training process, metric learning between anchor samples and positive/negative samples are employed to gradually converge similar samples, creating clearer class boundaries. Then the feature information of the samples is enhanced through an attention mechanism to obtain their essential features. The proposed framework was evaluated using two publicly available datasets and obtained classification accuracies of 68.29% and 84.40%, respectively, representing enhancements of 1.04% and 1.28% over existing state-of-the-art methods. In conclusion, experimental results indicate that our proposed approach can improve the effectiveness of MI-BCI rehabilitation training.