基于CNN-LSTM的脑电情感四分类研究

为深入研究脑电信号时空特征之间的关联，解决因手动提取特征导致的脑电情感识别准确率较低问题。将卷积神经网络（Convolutional neural network, CNN）和长短时记忆网络（Long short- term memory, LSTM）相结合，构造出了CNN-LSTM模型。首先，提取了5个频段的5个不同特征：功率谱密度(PSD)、差分熵(DE)、差分不对称(DASM)、理性不对称(RASM)和差分熵差分(DCAU)。其次，将特征输入CNN-LSTM模型，在DEAP数据集中的效价和唤醒两种情感维度上展开四分类实验。最后，将堆栈自编密码器(SAE)，卷积稀疏自编码器(CSAE)，深度置信网络(DBN)分别与LSTM组合，构造SAE-LSTM，CSAE-LSTM，DBN-LSTM三种混合模型同CNN-LSTM进行分类准确率比较。实验结果表明，DE特征的分类识别效果在五种特征中占最优，β和γ频段上所有特征的识别准确率远高于其他频段，尤其是γ频段。CNN-LSTM模型获得了最高的平均分类准确率92.48%，充分证明了CNN-LSTM模型的有效性。

Four classification of EEG emotion based on CNN-LSTM

In order to deeply study the correlation between space-time features of EEG signals and solve the problem of low accuracy of EEG emotion recognition caused by manual feature extraction. The convolutional neural network (CNN) and long short term memory (LSTM) are combined to construct a CNN-LSTM model. First, five different characteristics of five frequency bands are extracted: power spectral density (PSD), differential entropy (DE), differential asymmetry (DASM), rational asymmetry (RASM) and differential caudality (DCAU). Secondly, the features are input into the CNN-LSTM model, and four classification experiments are carried out on the two emotional dimensions of valence and arousal in the DEAP dataset. Finally, the stack self encryptor (SAE), convolutional sparse self encoder (CSAE), and depth confidence network (DBN) are combined with LSTM respectively to construct three hybrid models SAE-LSTM, CSAE-LSTM, and DBN-LSTM for classification accuracy comparison with CNN-LSTM. The experimental results show that the classification and recognition effect of DE features is the best among the five features. The recognition accuracy of all features in Beta and Gamma bands is much higher than that in other bands, especially in Gamma band. The CNN-LSTM model achieves the highest average classification accuracy of 92.48%, which fully proves the effectiveness of the CNN-LSTM model.