计算机科学与探索 ›› 2022, Vol. 16 ›› Issue (9): 2050-2060.DOI: 10.3778/j.issn.1673-9418.2103079

• 人工智能 • 上一篇    下一篇

融合注意力的多维特征图卷积运动想象分类

李珍琦1,2, 王晶1,2,+(), 贾子钰1,2, 林友芳1,2   

  1. 1.北京交通大学 计算机与信息技术学院,北京 100044
    2.北京交通大学 交通数据分析与挖掘北京市重点实验室,北京 100044
  • 收稿日期:2021-03-23 修回日期:2021-06-03 出版日期:2022-09-01 发布日期:2021-06-23
  • 通讯作者: + E-mail: wj@bjtu.edu.cn
  • 作者简介:李珍琦(1996—),女,河北唐山人,硕士研究生,主要研究方向为深度学习、时间序列分析与挖掘、脑机接口等。
    王晶(1987—),女,江苏扬州人,博士,副教授,博士生导师,CCF会员,主要研究方向为时间序列分析与挖掘、脑机接口、异常检测等。
    贾子钰(1993—),男,河北唐山人,博士研究生,主要研究方向为深度学习、时间序列分析与挖掘、脑机接口等。
    林友芳(1971—),男,福建武平人,博士,教授,博士生导师,CCF高级会员,主要研究方向为数据挖掘、机器学习、强化学习、复杂网络、智能技术与系统等。
  • 基金资助:
    中央高校基本科研业务费专项资金(2021JBM007);国家自然科学基金(61603029)

Attention-Based Multi-dimensional Feature Graph Convolutional Network for Motor Imagery Classification

LI Zhenqi1,2, WANG Jing1,2,+(), JIA Ziyu1,2, LIN Youfang1,2   

  1. 1. School of Computer and Information Technology, Beijing Jiaotong University, Beijing 100044, China
    2. Beijing Key Laboratory of Traffic Data Analysis and Mining, Beijing Jiaotong University, Beijing 100044, China
  • Received:2021-03-23 Revised:2021-06-03 Online:2022-09-01 Published:2021-06-23
  • About author:LI Zhenqi, born in 1996, M.S. candidate. Her research interests include deep learning, time series analysis and mining, brain-computer interface, etc.
    WANG Jing, born in 1987, Ph.D., associate professor, Ph.D. supervisor, member of CCF. Her research interests include time series analysis and mining, brain-computer interface, anomaly detection, etc.
    JIA Ziyu, born in 1993, Ph.D. candidate. His research interests include deep learning, time series analysis and mining, brain-computer interface, etc.
    LIN Youfang, born in 1971, Ph.D., professor, Ph.D. supervisor, senior member of CCF. His research interests include data mining, machine learning, reinforcement learning, complex network, intelligent technology and system, etc.
  • Supported by:
    Fundamental Research Funds for the Central Universities of China(2021JBM007);National Natural Science Foundation of China(61603029)

摘要:

运动想象(MI)作为脑机接口(BCI)的重要应用,是运动康复训练的重要支撑。由于脑电的电极分布并非天然的欧式空间,对运动想象进行准确分类具有很大的挑战。而且现有方法仅仅考虑了脑电信号(EEG)中某一维度或者某两维度的信息,无法全面捕获脑电信号在时、频、空三个维度存在的内在特征。同时,脑电信号各维度上的动态关联强度影响了分类的鲁棒性。针对上述问题,提出了一种新颖的融合注意力的多维特征图卷积网络(AMFGCN)。首先,根据电极节点分布的非欧空间特性设计出图结构,充分表示电极间的空间相关性。其次,提出时-空、频-空的双分支框架,同时表示脑电信号在时域、频域和空间域上的信息。最后,通过融合注意力机制、图卷积和时间/频谱卷积从图表示中学习脑电信号的空间表示、时间依赖性和频率依赖性,并自适应捕获各维度上的动态关联强度。在四个公开脑机接口数据集上进行了实验,结果表明AMFGCN模型提高了分类性能,优于其他现有的运动想象分类方法。

关键词: 运动想象(MI), 注意力机制, 图卷积网络, 多维特征, 脑电信号(EEG)

Abstract:

Motor imagery (MI), an important application of brain-computer interface (BCI), is critical for sports rehabilitation. Since the spatial position of electrodes is not Euclidean, accurate classification of MI is extremely challenging. Existing methods only consider the information of a single dimension or two dimensions in electroence-phalogram (EEG) signals, and cannot fully capture the inherent features of EEG signals in three dimensions of temporal, spectral and spatial. Meanwhile, dynamic correlations of EEG in each dimension affect the robustness of classification. To solve above problems, a novel model, named attention-based multi-dimensional feature graph convolutional network (AMFGCN), is proposed. Firstly, to model non-Euclidean spatial positions of electrodes and fully show spatial correlation between electrodes, a dedicated graph structure is designed. Secondly, a dual-branch framework of time-space and frequency-space dimensions is proposed, simultaneously decoding EEG signals from temporal, spectral and spatial domains. Thirdly, via fusing attention mechanisms into graph convolution and temporal/ spectral convolution, temporal, spectral and spatial correlations are then dynamically captured respectively. Experiments on four public BCI datasets illustrate superior performance of AMFGCN, which outperforms state-of-the-art competitive methods.

Key words: motor imagery (MI), attention mechanism, graph convolutional network, multi-dimensional feature, electroencephalogram (EEG)

中图分类号: