计算机工程与应用 ›› 2022, Vol. 58 ›› Issue (8): 175-184.DOI: 10.3778/j.issn.1002-8331.2010-0263

• 模式识别与人工智能 • 上一篇    下一篇

集成胶囊网络的脑电情绪识别

谌鈫,陈兰岚,江润强   

  1. 华东理工大学 化工过程先进控制和优化技术教育部重点实验室,上海 200237
  • 出版日期:2022-04-15 发布日期:2022-04-15

Emotion Recognition of EEG Based on Ensemble CapsNet

CHEN Qin, CHEN Lanlan, JIANG Runqiang   

  1. Key Laboratory of Advanced Control and Optimization for Chemical Processes, Ministry of Education, East China University of Science and Technology, Shanghai 200237, China
  • Online:2022-04-15 Published:2022-04-15

摘要: 为了充分提取脑电信号多频带的时频信息和保留导联空间分布的位置信息,提出了一种基于集成胶囊网络的情绪识别模型。对预处理过的脑电信号进行小波包特征提取,并将Theta、Alpha、Beta、Gamma四个频带的小波系数能量值填充于根据导联空间分布映射的稀疏矩阵中,拼接构成多频带特征矩阵,通过胶囊网络对特征数据进行训练,对不同频带的胶囊网络构建集成学习模型。实验结果表明,Gamma频带比低频带对情绪识别的敏感度更高,融合了多频带和空间信息的特征能够显著提升情绪识别的精度,最终集成胶囊网络将脑电情绪分为二类和四类的识别精度可以达到95.11%和92.43%,相比同类研究有一定的提升。

关键词: 情绪识别, 脑电信号, 多频带特征矩阵, 胶囊网络(CapsNet), 集成学习

Abstract: In this paper, an emotion classification model based on ensemble CapsNet is proposed in order to obtain the characteristics of multiband information of EEG signals and the spatial information of channel distribution. Firstly, the wavelet packet transform(WPT) is employed to decompose the pre-processed EEG signals into sub-frequency bands(Theta,Alpha, Beta, and Gamma) and then wavelet coefficient energy values are computed as the features in the four frequence bands. Secondly, the features are spliced to form a multiband feature matrix which are trained through the CapsNet and then generate multiple capsule models. Finally, the ensemble CapsNet model is constructed by the ensemble method. The experimental results show that high frequency features have more obvious effect on emotion recognition than those on low frequency bands, and the feature in Gamma band has the best emotional representation among all the independent frequency bands. Furthermore, the emotional feature integrating multiband information has a higher identification accuracy than that of each single band. The proposed method can achieve an average accuracy of 95.11% and 92.43% for two-category and four-category identification tasks respectively, which presents a certain improvement compared with similar research.

Key words: emotion recognition, EEG, multiband feature matrix, CapsNet, ensemble learning