Focused on the world-wide issue of improving the accuracy of emotion recognition, this paper proposes an electroencephalogram (EEG) signal feature extraction algorithm based on wavelet packet energy entropy and auto-regressive (AR) model. The auto-regressive process can be approached to EEG signal as much as possible, and provide a wealth of spectral information with few parameters. The wavelet packet entropy reflects the spectral energy distribution of the signal in each frequency band. Combination of them gives a better reflect of the energy characteristics of EEG signals. Feature extraction and fusion are implemented based on kernel principal component analysis. Six emotional states from a public multimodal database for emotion analysis using physiological signals (DEAP) are recognized. The results show that the recognition accuracy of the proposed algorithm is more than 90%, and the highest recognition accuracy is 99.33%. It indicates that this algorithm can extract the feature of EEG emotion well, and it is a kind of effective emotion feature extraction algorithm, providing support to emotion recognition.
针对提高情感识别正确率这一国际开放问题，本文提出了一种基于小波包熵和自回归模型相结合的脑电信号特征提取算法。自回归过程能最大程度逼近脑电信号，用很少的自回归参数提供丰富的谱信息。小波包熵反映脑电信号在各个频带中的谱能量分布情况。将二者结合，能够更好地体现脑电信号的能量特征。本文基于核主成分分析方法，实现了脑电信号特征提取融合。课题组采用情感脑电国际标准数据集（DEAP），选取 6 类情感状态以本文算法进行情感识别。结果显示，本文算法情感识别正确率均在 90% 以上，最高情感识别正确率可达 99.33%。本文的研究结果表明，该算法能够较好地提取脑电信号情感特征，是一种有效的情感特征提取算法。.
Keywords: auto-regressive; electroencephalogram; emotion recognition; kernel principal component analysis; support vector machine; wavelet packet entropy.