基于多尺度特征选择与空间通道重构卷积的运动想象脑电解码方法

Motor Imagery EEG Decoding Method Based on Multi-Scale Feature Selection and Spatial-Channel Reconstruction Convolution

摘要: 针对运动想象(Motor Imagery, MI)脑电信号(EEG)普遍存在的信噪比低、非平稳性强以及传统卷积神经网络在特征提取过程中易产生大量时空冗余信息的问题，本文提出了一种融合多尺度局部特征选择与特征重构机制的深度学习解码模型。首先，该模型在浅层特征提取后引入多尺度局部特征选择模块，通过并行的多尺度深度卷积捕获不同感受野下的特征，并利用可学习的通道注意力权重实现特征的自适应加权，以增强特征的判别性。其次，为了进一步抑制任务无关的冗余信息，模型引入了空间与通道重构卷积(SCConv)特征优化模块，通过空间重构单元(SRU)和通道重构单元(CRU)对特征图进行压缩与重组，从而显著提升特征表达的有效性。在大型公开数据集OpenBMI上的实验结果表明，该模型在运动想象任务中的平均准确率达到72.95%，优于EEGNet、Conformer等主流对比方法。消融实验进一步证实了多尺度特征选择模块与SCConv特征优化模块在提升模型鲁棒性和解码性能方面的关键作用。

Abstract: To address the challenges of low signal-to-noise ratio (SNR) and strong non-stationarity prevalent in Motor Imagery (MI) Electroencephalogram (EEG) signals, as well as the issue of substantial spatiotemporal redundancy generated by traditional Convolutional Neural Networks (CNNs) during feature extraction, this paper proposes a deep learning decoding model that integrates multi-scale local feature selection with a feature reconstruction mechanism. First, a multi-scale local feature selection module is introduced following shallow feature extraction. This module utilizes parallel multi-scale depthwise convolutions to capture features across different receptive fields and employs learnable channel attention weights to achieve adaptive feature weighting, thereby enhancing feature discriminability. Second, to further suppress task-irrelevant redundant information, a Spatial and Channel Reconstruction Convolution (SCConv) feature optimization module is incorporated. By utilizing a Spatial Reconstruction Unit (SRU) and a Channel Reconstruction Unit (CRU) to compress and restructure feature maps, this module significantly improves the effectiveness of feature representation. Experimental results on the large-scale public OpenBMI dataset demonstrate that the proposed model achieves an average accuracy of 72.95% in left- versus right-hand motor imagery tasks, outperforming mainstream comparison methods such as EEGNet and Conformer. Ablation experiments further verify the critical roles of the multi-scale feature selection module and the SCConv feature optimization module in enhancing model robustness and decoding performance.