计算机工程与应用 ›› 2019, Vol. 55 ›› Issue (9): 112-117.DOI: 10.3778/j.issn.1002-8331.1801-0337

• 模式识别与人工智能 • 上一篇    下一篇

面向脑电情感识别的改进多分类RVM模型研究

张雪英,王薇蓉,孙  颖,宋春晓   

  1. 太原理工大学 信息工程学院,太原 030024
  • 出版日期:2019-05-01 发布日期:2019-04-28

Research on Improved Multi-Classification RVM for Emotional Recognition of EEG Signal

ZHANG Xueying, WANG Weirong, SUN Ying, SONG Chunxiao   

  1. College of Information Engineering, Taiyuan University of Technology, Taiyuan 030024, China
  • Online:2019-05-01 Published:2019-04-28

摘要: 从相关向量机(RVM)和支持向量机(SVM)的相似性以及RVM的稀疏特性出发,将RVM应用于脑电信号(EEG)的情感识别中。针对一对一(OAO)和一对多(OAA)两种多分类方法各自的特点和不足,提出了一种全新的两层多分类模型(OAA-OAO),改进现有OAO算法中无效投票影响最终决策的现象。设计情感EEG信号识别对比实验,验证基于RVM的改进多分类算法在脑电信号情感识别中的应用。对于实验室采集的情感脑电信号,提取其非线性特征(功率谱熵、样本熵和Hurst指数)并采用主成分分析法进行降维。将OAA-OAO-RVM算法分别和OAO-SVM、OAO-RVM两种识别网络进行对比,分析RVM的识别性能以及OAA-OAO多分类算法的分类性能。结果表明,采用降维后的最优特征集合作为识别网络的输入向量得到的识别性能更高,且RVM表现出的性能优于SVM。同时,改进后的OAA-OAO算法较传统OAO模型的平均识别率提高了7.89%,证明OAA-OAO算法可有效去除一部分无效投票从而使分类精度得到显著提高,验证了此模型是一种有效的多分类模型。

关键词: 相关向量机, 支持向量机, 多分类, 脑电信号, 情感识别

Abstract: Based on the similarity of Relevance Vector Machine(RVM) and Support Vector Machine(SVM) and the sparse characteristics of RVM, RVM is applied to the electroencephalogram(EEG) signal emotional recognition. Aiming at the characteristics and deficiencies of the two methods of One-Against-One(OAO) and One-Against-All(OAA), this paper proposes a new two-layer multi-classification model(OAA-OAO) to improve the effect of invalid voting on the final decision-making in the existing OAO algorithm. A series of emotional EEG signal recognition experiments are designed to verify the effectiveness of improved multi-classification algorithm based on RVM in EEG emotion recognition. For the emotional EEG signal collected by the laboratory, this paper extracts the nonlinear features(power spectral entropy, sample entropy and Hurst index) and reduces the dimension by Principal Component Analysis(PCA). OAA-OAO-RVM is compared with OAO-SVM and OAO-RVM respectively, the recognition performance of RVM and the classification performance of OAA-OAO is analyzed. The results show that using the optimal feature set after dimension reduction as input vector of the recognition network has higher recognition rate, and the performance of RVM is better than that of SVM. At the same time, the average recognition rate of OAA-OAO is 7.89% higher than that of OAO, which proves that the OAA-OAO algorithm can effectively remove part of the invalid votes so that the classification accuracy can be improved obviously and this model is an effective multi-classification model.

Key words: Relevance Vector Machine(RVM), Support Vector Machine(SVM), multi-classification, electroencephalogram(EEG) signal, emotional recognition