计算机科学与探索 ›› 2023, Vol. 17 ›› Issue (6): 1427-1440.DOI: 10.3778/j.issn.1673-9418.2108004

• 人工智能·模式识别 • 上一篇    下一篇

融合多尺度自注意力机制的运动想象信号解析

刘京,赵薇,董泽浩,王少华,王余   

  1. 1. 河北师范大学 计算机与网络空间安全学院,石家庄 050024
    2. 河北师范大学 软件学院,石家庄 050024
    3. 河北师范大学 数学科学学院,石家庄 050024
  • 出版日期:2023-06-01 发布日期:2023-06-01

Motor Imagery Signal Classification Based on Multi-scale Self-attentional Mechanism

LIU Jing, ZHAO Wei, DONG Zehao, WANG Shaohua, WANG Yu   

  1. 1. College of Computer and Cyber Security, Hebei Normal University, Shijiazhuang 050024, China
    2. College of Software, Hebei Normal University, Shijiazhuang 050024, China
    3. School of Mathematical Sciences, Hebei Normal University, Shijiazhuang 050024, China
  • Online:2023-06-01 Published:2023-06-01

摘要: 基于运动想象脑电信号的脑机接口(BCI)技术近年来发展迅速,并且与传统方法相比,深度学习取得了具有竞争力的效果。然而如何设计和训练一个端到端网络来充分提取运动想象脑电信号的潜在特征仍然是一个挑战。从脑电的时间和空间特征出发,提出了一个基于注意力机制的多尺度时空自注意力网络模型用于运动想象脑电信号四分类(左手、右手、脚、舌头/休息)。由于运动想象脑电信号的幅值与响应时间因人而异,无法确切确定与运动想象最相关的大脑区域,在空间上使用自注意力机制自动将较高的权重加权到与运动相关的通道,将较低的权重加权到与运动无关的通道来选择最佳通道;在时间上,使用并行多尺度TCN层提取不同尺度下的时间域特征信息,消除时间域上的噪声。多尺度融合模块融合提取的空间和时间域特征,最后输入到特征分类模块进行分类。提出的模型在BCI竞赛数据集IV-2a、IV-2b数据集和HGD数据集上分别达到79.26%、85.90%和96.96%的精度。与现有方法相比,该方法在单被试分类中具有更高的准确率。结果表明,该方法具有较好的性能、鲁棒性和迁移学习能力。

关键词: 运动想象, 注意力, 脑机接口, 深度学习, 多尺度

Abstract: Brain-computer interface (BCI) technology based on motor imagery electroencephalograph (EEG) signals has developed rapidly, and compared with traditional methods, deep learning has achieved competitive results. However, it is still a challenge to design and train an end-to-end network to fully extract the potential features of motor-imaging EEG signals. Based on the temporal and spatial characteristics of EEG, a multi-scale spatiotemporal self-attention network model based on attention mechanism is proposed to classify motor imagery EEG signals into four categories, such as left hand, right hand, foot and tongue/rest. Since the amplitude and response time of motor imagery EEG signal vary with the subjects, it is impossible to determine the most relevant brain region with motor imagery. Therefore, the self-attention mechanism is used to select the best channel by automatically weighting the higher weight to the motion-related channel and the lower weight to the non-motion-related channel in space. And parallel multi-scale TCN layer is used to extract the temporal feature at different scales and to eliminate the noise in time domain. The fusion module fuses the extracted spatial and temporal features, and finally inputs them to the classification module for classification. The model achieves accuracy of 79.26%, 85.90% and 96.96% on the BCI competition datasets IV-2a, IV-2b and HGD dataset. Compared with the baseline method, this method has a higher accuracy in subject classification. The results show that the method has good performance, robustness and transfer learning ability.

Key words: motor imagery, attention, brain-computer interface, deep learning, multi-scale