计算机工程与应用 ›› 2022, Vol. 58 ›› Issue (13): 146-153.DOI: 10.3778/j.issn.1002-8331.2012-0505

• 模式识别与人工智能 • 上一篇    下一篇

脑电信号多维特征融合与分类算法研究

朱永清,王文格   

  1. 湖南大学 机械与运载工程学院,长沙 410006
  • 出版日期:2022-07-01 发布日期:2022-07-01

Research on Multi-Dimensional Feature Extraction and Classification Algorithm of EEG Signal

ZHU Yongqing, WANG Wenge   

  1. College of Mechanical and Vehicle Engineering, Hunan University, Changsha 410006, China
  • Online:2022-07-01 Published:2022-07-01

摘要: 针对目前运动想象脑电信号特征提取单一,分类识别准确率低等现象,结合卷积神经网络分类器,提出了一种多维度特征加权融合的特征融合算法来提高运动想象脑电识别率。对预处理后的脑电信号进行小波包变换,提取其共空间特征、能量特征、边际谱熵特征以及非线性动力学特征,然后加权融合,使用卷积神经网络分类器分类。为验证算法的合理性,使用BCI-IV Dataset 2a数据集对提出的特征融合算法进行验证分析,结果表明,所提出的加权特征融合算法结合CNN分类器可以有效提高运动想象识别准确率。实验中,9位志愿者平均分类准确率达到75.88%,平均Kappa系数为0.70。

关键词: 脑机接口, 运动想象, 卷积神经网络, 小波包变换, 特征融合

Abstract: Aiming at the current phenomenon of single feature extraction of motor imagery EEG signals and low classification recognition rate, this paper combines the classifier of convolutional neural networks and proposes a feature fusion algorithm of multi-feature weighted fusion to improve motor imagery EEG recognition rate. It performs wavelet packet transform on the preprocessed EEG signals, extracts its common space features, energy features, marginal spectral entropy features and nonlinear dynamics features, then performs weighted fusion. Finally, the paper uses convolutional neural network classification to classify. To verify the rationality of the algorithm, it uses dataset which is the BCI-IV Dataset 2a to verify and analyze the proposed algorithm. The results show that the proposed weighted feature fusion algorithm with CNN classifier can effectively improve the recognition rate of motion imagination. In the experiments, the average classification accuracy of 9 volunteers is 75.88%, and the average Kappa coefficient is 0.70.

Key words: brain-computer interface, motor imagery, convolutional neural network, wavelet packet transform, feature fusion