基于增广协方差矩阵的新型运动想象脑机接口分类框架：带自注意力机制的深度黎曼几何学习

The rapid development of motor imagery brain-computer interface (MI-BCI) technology has introduced a novel mode of human-machine interaction, drawing significant interest. However, due to the nonlinearity and nonstationarity of motor imagery signals, this communication method faces challenges such as low recognition accuracy and poor stability, which severely hinder the practical application of MI-BCIs. Riemannian geometry techniques, leveraging their non-Euclidean properties, have shown promise in addressing motor imagery signal pattern recognition. Traditional Riemannian geometry methods, however, still suffer from limitations, such as requiring manual feature extraction and resulting in suboptimal classification performance. This paper proposes Aug-SPD-AttentionNet, a deep Riemannian geometry learning framework for MI-BCI classification that combines lagged augmented covariance representations, a learnable high-dimensional embedding layer, and self-attention operations defined on the SPD manifold. The proposed framework employs multiple deep Riemannian learning blocks to extract manifold-aware features and uses a Riemannian self-attention mechanism to capture long-range dependencies among different SPD representations. Aug-SPD-AttentionNet was evaluated on BCI Competition III Dataset IVa, BCI Competition IV Dataset I, and BCI Competition IV Dataset 2a, achieving mean classification accuracies of [Formula: see text], [Formula: see text], and [Formula: see text], respectively. Ablation experiments further demonstrate that the proposed components collectively improve classification performance, increasing the mean accuracy across the three evaluated subjects from 70.60 to 82.62%, corresponding to a 12.03-percentage-point improvement over the baseline SPDNet.