计算机工程与应用 ›› 2022, Vol. 58 ›› Issue (4): 186-191.DOI: 10.3778/j.issn.1002-8331.2008-0404

• 模式识别与人工智能 • 上一篇    下一篇

基于图卷积网络的运动想象识别

许学添,蔡跃新   

  1. 1.广东司法警官职业学院 信息管理系,广州 510520
    2.中山大学孙逸仙纪念医院 耳鼻喉科 听力学与言语研究所,广州 510120
  • 出版日期:2022-02-15 发布日期:2022-02-15

Motor Imagery Recognition Based on Graph Convolution Network

XU Xuetian, CAI Yuexin   

  1. 1.Department of Information Administration, Guangdong Justice Police Vocational College, Guangzhou 510520, China
    2.Institute of Hearing and Speech-Language Science, Department of Otolaryngology, Sun Yat-sen Memorial Hospital, Sun Yat-sen University, Guangzhou 510120, China
  • Online:2022-02-15 Published:2022-02-15

摘要: 运动想象识别将大脑的神经活动信号转为编码输出以实现意念控制,是脑机接口的一个重要研究方向。近年来深度学习算法的应用进一步提高了运动想象识别的准确率,但是当前基于深度学习的运动想象分析都将多路脑电信号作为二维矩阵信号,忽视了不同节点的空间关联信息。为了解决这个问题,将图卷积网络算法应用到运动想象分类中,通过多个节点脑电信号的相关系数建立脑电图结构,提取脑电信号的时频域特征信息作为输入,再经过图卷积网络进行节点特征聚合以学习谱域特征,最后通过全连接层输出分类结果。该方法在BCI Competition IV Dataset 2a数据集上取得80.9%的准确率和0.74的kappa系数,相比其他方法可以充分学习时、频、谱域的特征信息,取得更好的识别效果,为运动想象脑机接口提供一种新的思路和方法。

关键词: 图卷积网络(GCN), 运动想象, 深度学习, 时频特征

Abstract: Motor imagery recognition is an important research direction of brain-computer interface(BCI), which converts brain neural activity signals into code to realize mind control. In recent years, the application of deep learning algorithm further improves the accuracy of motor imagery recognition. However, the current motor imagery analysis based on deep learning takes multi-channel EEG signals as two-dimensional matrix signals, and ignores the different nodes spatial association information. To solve this problem, the graph convolution network(GCN) is applied to the classification of motor imagery. The EEG structure is established by the correlation coefficients of multiple nodes’ EEG signals, and the time-frequency feature information of EEG signals is extracted as the input of GCN. Then, node time-frequency features are aggregated by graph convolution network to learn spectral domain features. Finally, the classification results are output through full connection layer. This method achieves 80.9% accuracy and 0.74 kappa coefficient on BCI Competition IV Dataset 2a. This method can fully learn the feature information of time, frequency and spectrum domain, and achieve better recognition effect. It provides a new idea and method for brain-computer interface of motor imagery.

Key words: graph convolution network(GCN), motor imagery, deep learning, time-frequency feature