注意力残差网络结合LSTM的EEG情绪识别研究  

Research on EEG Emotion Recognition of Attention Residual Network Combined with LSTM

作　　者：张琪[1] 熊馨[1] 周建华[1] 宗静 周雕 ZHANG Qi;XIONG Xin;ZHOU Jianhua;ZONG Jing;ZHOU Diao(School of Information Engineering and Automation,Kunming University of Science and Technology,Kunming 650500,China)

机构地区：[1]昆明理工大学信息工程与自动化学院,昆明650500

出　　处：《华东理工大学学报（自然科学版）》2024年第4期570-579,共10页Journal of East China University of Science and Technology

基　　金：国家自然科学基金(82060329)。

摘　　要：基于脑电信号的情感识别已成为情感计算和人机交互领域的一个重要挑战。由于脑电信号中具有时间、空间、频率维度信息,采用结合注意力残差网络与长短时记忆网络的混合网络模型(ECA-ResNet-LSTM)对脑电信号进行特征提取与识别。首先,提取时域分段后脑电信号不同频带微分熵特征,将从不同通道中提取出的微分熵特征转化为四维特征矩阵;然后通过注意力残差网络(ECA-ResNet)提取脑电信号中空间与频率信息,并引入注意力机制重新分配更相关频带信息的权重,长短时记忆网络(LSTM)从ECA-ResNet的输出中提取时间相关信息。实验结果表明:在DEAP数据集唤醒维和效价维二分类准确率分别达到了97.15%和96.13%,唤醒-效价维四分类准确率达到了95.96%,SEED数据集积极-中性-消极三分类准确率达到96.64%,相比现有主流情感识别模型取得了显著提升。Emotion recognition based on EEG signals has become an important challenge in the field of emotional computing and human-computer interaction.In order to obtain better emotion recognition performance,a key issue is how to effectively combine time,space and frequency dimension information in EEG signals.This paper proposes a hybrid network model(ECA-ResNet-LSTM)combining attention residual networks and long short-term memory networks.By integrating time,space and frequency information in EEG signals,this model can effectively improve the accuracy of emotion recognition.Firstly,the differential entropy features of EEG signals in different frequency bands after time-domain segmentation are extracted,and the differential entropy features extracted from different channels are transformed into a four-dimensional feature matrix.Then,the spatial and frequency information in the EEG signal is extracted through ECA ResNet,and attention mechanisms are introduced to redistribute the weights of more relevant frequency band information.LSTM extracts time related information from the output of ECA ResNet.Finally,the experimental results show that in DEAP dataset,the accuracy of awakening and valence dimension binary classification reaches 97.15% and 96.13%,respectively,and the accuracy of awakening valence dimension four classification reaches 95.96%;In SEED dataset,the accuracy of positive neutral negative three classification reaches 96.64%.Compared with the existing mainstream emotion,the classification accuracy of the recognition model has been significantly improved.

关 键 词：脑电信号 情感识别 微分熵 注意力机制 残差网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]