基于通道自注意图卷积网络的运动想象脑电分类实验  

Experimental classification of motor imagery EEG based on graph convolutional networks with channel self-attention

作　　者：孟明[1] 张帅斌 高云园[1] 佘青山[1] 范影乐[1] MENG Ming;ZHANG Shuaibin;GAO Yunyuan;SHE Qingshan;FAN Yingle(School of Automation,Hangzhou Danzi University,Hangzhou 310018,China)

机构地区：[1]杭州电子科技大学自动化学院,浙江杭州310018

出　　处：《实验技术与管理》2025年第2期73-80,共8页Experimental Technology and Management

基　　金：浙江省普通本科高校“十四五”教学改革项目(JG20220228);浙江省“十四五”研究生教学改革项目(2023-173);浙江省产学合作协同育人项目(2023-87);国家自然科学基金项目(62271181)。

摘　　要：该文将运动想象脑电分类任务设计成应用型教学实验。针对传统图卷积网络(graph convolutional neural networks,GCN)无法建模脑电通道间动态关系问题,提出一种融合通道注意机制的多层图卷积网络模型(channel self-attention multilayer GCN,CAMGCN)。首先,CAMGCN计算脑电信号各个通道间的皮尔逊相关系数进行图建模,并通过通道位置编码模块学习通道间关系。然后将得到的时域和频域特征分量通过通道自注意图嵌入模块进行图嵌入,得到图数据。最后通过多级GCN模块提取并融合多层次拓扑信息,得出分类结果。CAMGCN深化了模型在自适应学习通道间动态关系的能力,并在结构方面提高了自注意机制与图数据的适配性。该模型在BCI Competition-Ⅳ2a数据集上的准确率达到83.8%,能够有效实现对运动想象任务的分类。该实验有助于增进学生对于深度学习和脑机接口的理解,培养创新思维,提高科研素质。[Objective]Graph convolutional neural networks(GCNs),which are deep learning algorithms for processing graph-structured data,have recently garnered significant attention in the field of electroencephalogram(EEG)-based brain–computer interface(BCI)systems given their exceptional capabilities in capturing the topological features of signals.Most existing motor imagery EEG classification methods based on GCN models exclusively focus on the construction of static graphs,which could potentially ignore the multidomain dynamic changes that occur within deep graph structures.These dynamic changes are crucial for accurately decoding complex brain activities.To address this limitation,a novel channel self-attention multilayer GCN(CAMGCN)is proposed to enhance the GCN model’s ability to learn dynamic relationships between different EEG channels for motor imagery EEG classification experiments.[Methods]The classification experiments in this study employed the BCI CompetitionⅣ2a dataset,which included nine subjects with four types of motor imagery tasks.Preprocessing steps,which included filtering,artifact removal,and baseline correction,were applied to the original EEG signals.Then,differential entropy features were extracted from specific frequency bands of the preprocessed signals to describe their nonlinear characteristics.Meanwhile,a one-dimensional convolutional neural network was used to extract time domain features to characterize the temporal characteristics of the signals.To construct the CAMGCN model,an adjacency matrix was calculated to model the topology of the graph on the basis of the Pearson correlation coefficient between each channel of the EEG signal.The relationships between different channels were then learned using a channel location encoding module.The feature components in the time and frequency domains were embedded by a channel self-attention graph embedding module to obtain the graph data.Subsequently,multiple GCN modules were used to extract and fuse multilevel topological information,which w

关 键 词：脑机接口 脑电图 图卷积网络 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]