面向运动想象解码的双尺度时空特征融合网络

Dual-Scale Spatiotemporal Feature Fusion Network for Motor Imagery Decoding

摘要: 卷积神经网络(Convolutional neural networks, CNNs)已在运动想象(Motor Imagery, MI)脑机接口(Brain-Computer Interface, BCI)领域取得了广泛应用和良好效果。然而，基于单一尺度卷积结构的模型难以从脑电信号中充分挖掘多样化信息；而现有的多尺度CNN虽然能够提取不同尺度的特征，但通常仅通过简单拼接的方式进行融合，难以实现多尺度信息之间的深度协同建模。为了解决上述问题，相关研究提出了双尺度时空特征融合网络(Dual-Scale Spatiotemporal Feature Fusion Network, DSSFFN)。该方法通过双尺度卷积分支以提取脑电信号不同尺度的特征，并通过Transformer模块对来自不同尺度的特征进行融合，从而提升特征的判别性。本文采用BCI竞赛IV的2a数据集进行实验，同时与多个前沿的运动想象算法进行比较。结果显示，DSSFFN在数据集上平均准确率为72.65%，高于所有的对比模型，展现出了DSSFFN模型在运动想象脑电识别任务上的优异性能。此外，研究还通过消融实验分析了双尺度时空卷积分支与Transformer模块对整体性能的贡献，验证了各个关键模块的有效性。同时比较了不同融合模型对模型影响，进一步验证了Transformer模块融合双尺度的有效性。

Abstract: Convolutional neural networks (CNNs) have been widely applied in motor imagery (MI) brain-computer interface (BCI) systems and have achieved promising performance. However, models based on single-scale convolutional structures are often insufficient to fully exploit the diverse and rich information embedded in EEG signals. Although existing multi-scale CNNs can extract features at different scales, they usually fuse multi-scale features through simple concatenation, which makes it difficult to achieve deep and effective collaboration among features from different scales. To address these issues, related research proposes a Dual-Scale Spatiotemporal Feature Fusion Network (DSSFFN). This method employs dual-scale convolutional branches to extract EEG features at different scales, and further integrates the features from different branches using a Transformer module, thereby enhancing the discriminative capability of the feature representations. In this study, experiments are conducted on the BCI Competition IV-2a dataset, and the proposed method is compared with several state-of-the-art MI recognition algorithms. The results show that DSSFFN achieves an average classification accuracy of 72.65% on this dataset, outperforming all competing methods, thus demonstrating its superior performance in motor imagery EEG classification tasks. Furthermore, ablation studies are performed to analyze the contributions of the dual-scale spatiotemporal convolution branches and the Transformer fusion module, which verifies the effectiveness of each key component in the proposed network. In addition, the impacts of different fusion models on the proposed model’s performance were compared, which further verifies the effectiveness of the Transformer module in fusing dual-scale features.