利用卷积回声状态网络实现脑电情感识别  

EEG-Based Emotion Recognition by Using Convolutional Echo-State Network

作　　者：晁浩[1] 马庆敏 刘永利[1] CHAO Hao;MA Qingmin;LIU Yongli(College of Computer Science and Technology,Henan Polytechnic University,Jiaozuo 454003,China)

机构地区：[1]河南理工大学计算机科学与技术学院,焦作454003

出　　处：《北京邮电大学学报》2022年第2期36-43,共8页Journal of Beijing University of Posts and Telecommunications

基　　金：国家自然科学基金项目(61872126);河南省高等学校重点科研项目(19A520004);河南省高校基本科研业务费专项资金项目(NSFRF1616)。

摘　　要：针对多通道脑电(EEG)的情感识别,提出了一种卷积回声状态网络(CESN)模型。首先构造EEG信号的特征矩阵序列;然后通过卷积操作提取各个样本的高层抽象特征,形成一维特征向量序列;利用具有自反馈功能的蓄水池结构,捕获向量序列的动态时序信息;最后用岭回归来实现情感识别。在情感分析专用生理信号数据集上进行实验的结果表明,EEG信号的动态时序性蕴含着与情感状态相关的区分性信息,所提的CESN模型能够有效地挖掘这种信息,并用于情感分类,解决了卷积神经网络中因使用反向传播算法而导致的局部最优和训练时间过长的问题。A convolutional echo-state network(CESN) model is proposed for the emotion recognition task based on electroencephalogram(EEG) signals. First, a feature matrix sequence of EEG signal is constructed. Then, high-level abstract features are extracted from the feature matrices via convolution, and one dimension feature vectors are formed. After that, a reservoir with self-feedback function is employed to extract the dynamic temporal information from the feature vector sequence. Finally, the emotion recognition task is realized by ridge regression. The experiment is carried out on the database for emotion analysis using physiological signals datasets. The experiment result shows that the EEG signal segments contain temporal information related to emotion, and the CESN model can mine and utilize the information effectively. In addition, the proposed CESN model can circumvent the problems of the local optimization and long training time, which are caused by back propagation algorithm in convolutional neural network.

关 键 词：多通道脑电信号 情感识别 卷积回声状态网络 特征融合 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]