基于MHSA⁃EDSCNet混合模型的运动想象多任务分类研究  

Multi⁃Task Classification of Motor Imagery Based on MHSA⁃EDSCNet Hybrid Model

作　　者：张圆 乔晓艳[1] ZHANG Yuan;QIAO Xiaoyan(College of Physics and Electronic Engineering,Shanxi University,Taiyuan 030006,China)

机构地区：[1]山西大学物理电子工程学院,山西太原030006

出　　处：《测试技术学报》2024年第6期652-660,共9页Journal of Test and Measurement Technology

基　　金：山西省回国留学人员科研资助项目(2020-009)。

摘　　要：运动想象脑电信号解码是脑机接口技术的关键环节。针对传统深度学习方法难以获得脑电全局信息,提出多头自注意力(MHSA)机制结合改进的深度可分离卷积网络(EDSCNet)模型,用于运动想象多任务分类。首先,通过滤波器组共空间模式提取不同子带共空间模式空域特征,准确获取运动想象脑电的细粒度特征信息;其次,利用一维卷积改进深度可分离卷积网络,进一步提取脑电局部空间信息和空间关联信息,并结合多头自注意力机制,更好地捕捉运动想象脑电特征的全局空间信息,增强特征表征能力,提高多任务分类准确率,同时可减少模型参数和计算量;最后,在BCI Competition IV2a运动想象脑电数据集对该模型进行验证和评估,并对左手、右手、双脚和舌头四类运动想象任务脑电特征进行可视化。结果表明:模型在两个运动想象四类任务数据集,分别获得95.35%和96.87%的平均分类准确率以及0.9379和0.9586的Kappa系数。模型特征可视化对大脑不同的运动想象任务能够显著区分,并且模型对所有被试表现出一致的性能。Decoding motor imagery-based electroencephalography(EEG)signals is a crucial step in braincomputer interface(BCI)technology.To address the challenge of capturing global EEG information,this study proposes a model that combines a multi-head self-attention mechanism with an improved depthwise separable convolutional network(EDSCNet)for multi-task classification of motor imagery.Firstly,spatial features of different subbands are extracted using a filter bank common spatial pattern technique to capture fine-grained characteristics of motor imagery EEG accurately.Secondly,an enhanced onedimensional convolutional network,the EDSCNet,is employed to capture local spatial information and spatial correlation of EEG signals.The model further incorporates a multi-head self-attention mechanism better to capture the global spatial information of motor imagery EEG,enhancing feature representation and improving multi-task classification accuracy,while reducing model parameters and computational com⁃plexity.Finally,the proposed model is validated and evaluated on the BCI Competition IV2a motor imag⁃ery EEG dataset,and the visualizations of EEG features for the left hand,right hand,both feet,and tongue motor imagery tasks are presented.The results demonstrate that the model achieves average classification accuracies of 95.35%and 96.87%,with Kappa coefficients of 0.9379 and 0.9586,respectively,on the two motor imagery datasets.The feature visualizations show distinct patterns for different motor imagery tasks in the brain.Moreover,the model exhibits consistent performance across all subjects.

关 键 词：脑电信号 深度可分离卷积 滤波器组共空间模式 多头自注意力 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]