基于双分支多特征融合网络的运动想象脑电信号分类方法

DBMF-Net: Motor Imagery EEG Signals Classification Method Based on a Dual-Branch Multi-Feature Fusion Network

摘要: 基于运动想象的脑机接口(MI-BCI)为重度瘫痪患者提供了独立于外周神经和肌肉的人机交互途径，在运动功能康复领域具有重要应用价值。然而，脑电信号的会话间变异性大、判别性能量特征提取困难等问题，给MI-BCI系统的信号解码带来了严峻挑战。针对上述问题，本文提出一种双分支多特征融合网络(DBMF-Net)，通过多分支结构联合学习EEG信号的一阶时空特征、二阶协方差特征及时频域特征。具体而言，该模型利用多尺度卷积与Transformer编码器提取时空特征，引入基于黎曼几何的SPD协方差嵌入机制建模通道间二阶统计关系，并通过短时傅里叶变换分支捕获时频域判别信息，最终将三类特征融合用于运动想象分类。在BCI Competition IV 2a和2b数据集上的实验结果表明，DBMF-Net的平均分类准确率分别达到85.87%和89.39%，优于现有基准方法，验证了该方法在被试内会话间脑电解码任务中的有效性与鲁棒性。

Abstract: The motor imagery-based brain-computer interface (MI-BCI) provides an interaction pathway independent of peripheral nerves and muscles for individuals with severe paralysis, holding significant application value in the field of motor function rehabilitation. However, challenges such as large inter-session variability of EEG signals and the difficulty in extracting discriminative energy features pose serious obstacles to signal decoding in MI-BCI systems. To address these issues, this paper proposes a Dual-Branch Multi-Feature Fusion Network (DBMF-Net), which jointly learns first-order spatiotemporal features, second-order covariance features, and time-frequency domain features of EEG signals through a multi-branch structure. Specifically, the model extracts spatiotemporal features using multi-scale convolution and a Transformer encoder, introduces an SPD covariance embedding mechanism based on Riemannian geometry to model second-order statistical relationships between channels, and captures frequency-domain discriminative information through a short-time Fourier transform branch. Finally, the three types of features are fused for motor imagery classification. Experimental results on the BCI Competition IV 2a and 2b datasets demonstrate that DBMF-Net achieves average classification accuracies of 85.87% and 89.39%, respectively, outperforming existing baseline methods and validating the effectiveness and robustness of the proposed method in intra-subject inter-session EEG decoding tasks.