【摘要】 基于脑电信号(Electroencephalogram, EEG)的情绪分析和识别研究一直是脑科学领域的热点，利用复Morlet小波变换构建脑电张量，结合其低秩、稀疏的特点，提出一种EEG多域特征提取方法——稀疏正则的低秩逼近Tucker分解算法。该算法提取样本所特有的多域特征——核心张量和样本共有的因子矩阵，对情绪脑电进行分类和分析，克服了传统Tucker分解计算效率低，易导致维度爆炸的缺陷。实验结果表明，在MODMA数据集上，以核心张量作为多域特征进行EEG样本分类，对不同情绪刺激下抑郁症患者(MDD)的平均识别率为88.9%,且运算效率较传统Tucker分解提高约16倍。利用表征空间特征的因子矩阵对活跃脑区进行分析，实现对脑区空间层面动态变化的对比，发现MDD患者与正常对照组在效价与唤醒度的敏感度上的差异。SLraTucker分解算法能够有效提取EEG的多域特征，为分析情绪脑电以进行相应诊断提供了新的方法和思路。更多还原

【Abstract】 Emotion recognition based on electroencephalogram(EEG)has always been a hot topic in the field of brain science. Complex Morlet wavelet transform is used to construct EEG tensor. Considering its low rank and sparse characteristics, we propose a multi domain feature extraction method of EEG called sparse regulation for low rank approximation Tucker decomposition(SLraTucker). This algorithm extracts the unique multi domain features which is the core tensor and the common factor matrix to classify and analyze the emotional EEG. It overcomes the limitations of traditional Tucker decomposition, such as low computational efficiency and easy to cause dimension explosion. The experimental results show that on the MODMA data set, the core tensor is used as the multi domain feature to classify the EEG samples. The average recognition rate of major depressive disorder(MDD)patients under different emotional stimulation is 88.9%,and the decomposition efficiency is about 16 times higher than that of traditional Tucker decomposition. The mode-1 factor matrix which represents the spatial features is used to analyze the active brain regions, and the dynamic changes of brain regions are compared at the spatial level. The differences in the sensitivity of valence and arousal between MDD patients and health controls are found. The SLraTucker decomposition algorithm can effectively extract the multi domain features of EEG,which provides a new method and idea for the analysis of emotional EEG to make corresponding diagnosis.更多还原