基于三维并行多视野卷积神经网络的脑电信号情感识别

Emotion Recognition of EEG Signals Based on 3D Parallel Multi-Field Convolutional Neural Network

利用脑电信号识别情感状态已经成为当前的研究热门。现有的情感识别方法通常提取二维信息作样本,却忽略了包含大脑不同区域重要特征的空间信息。针对这个问题,结合脑电通道间布局和原始脑电信号中的频率相关特征,提出了基于三维并行多视野卷积神经网络(Three-dimensional Parallel Multi-field Convolutional Neural Network, TPMCNN)的脑电情感识别新方法。首先将原始脑电信号划分成多频带,并提取每个频带的微分熵(DE)特征。接着将数据按照电极传感器的位置转变成三维特征矩阵。最后采用TPMCNN网络处理所得到的矩阵。实验结果表明,利用不同频带的微分熵特征构造的三维特征矩阵,能够有效地提取多通道脑电信号中与情感识别有关的特征,所提出的并行多视野卷积神经网络能够充分发挥出深度学习的优势。实验在公开数据集DEAP上进行二分类,在唤醒和效价的准确率分别达到了97.31%和96.72%,四分类的准确率达到了97.17%,证实了所提出的方法对脑电信号情感识别的优越性能。

The use of EEG signals to identify emotional states has become a popular research topic.Existing emotion recognition methods usually extract two-dimensional information as samples,but ignore the spatial information that containing important features of different re-gions of the brain.A three-dimensional parallel multi-field convolutional neural network(TPMCNN)based on the layout of EEG channels and the frequency-related features in the original EEG signal is proposed to address this problem.Firstly,the original EEG signal is divid-ed into multiple frequency bands,and the differential entropy(DE)features of each band are extracted.Then the data are transformed into a 3D feature matrix according to the location of the electrode sensors.Finally the resulting matrix is processed by using the TPMCNN.The experimental results showe that the 3D feature matrix constructed using differential entropy features of different frequency bands can effec-tively extract features related to emotion recognition from multi-channel EEG signals,and the proposed parallel multi-field convolutional neural network can fully exploit the advantages of deep learning.The experiments are performed on the publicly available dataset of DEAP for dichotomous classification,and the values of accuracy reach 97.31%and 96.72%for arousal and valence respectively,and 97.17%for four-classification,confirming the superior performance of the proposed method for EEG signal emotion recognition.