基于深度自编码和LSTM循环网络的脑电情感识别

EEG Emotion Recognition Based on Deep Auto-Encoder and LSTM

针对脑电信号情感识别中情感特征信息挖掘不充分、识别准确率低的问题,提出深度自编码方法提取脑电信号情感特征,并结合长短时记忆(LSTM)循环神经网络实现维度情感分类.首先,基于DEAP维度情感生理数据集,分别在唤醒度和效价维度选取情感标签阈值,划分不同情感状态;然后,通过时间窗对脑电信号分段,设计栈式自编码网络挖掘脑电数据蕴含的情感信息,并依照时间顺序生成情感特征序列;最后,建立LSTM循环神经网络进行模型训练、交叉验证和测试,并通过正确率、精确度、召回率和F1-Score评价情感分类效果.仿真实验结果表明:在效价维度上,脑电信号情感识别平均正确率达到77.4%,F1-Score为80.4%;在唤醒度上,平均识别正确率达到73.7%,F1-Score为77.5%.该方法使维度情感模型获得了较好的情绪识别结果,可以为情感计算和人机情感交互提供借鉴.

Aiming to the problems of inadequate mining of emotion feature information and low recognition accuracy in EEG signal emotion recognition,this paper proposes a deep auto-encoder method to extract the emotion feature of EEG signal,and combines with the Long Short-term Memory(LSTM)recurrent neural network to realize continuous dimension emotion classification.Firstly,based on the DEAP dimension emotional physiological dataset,emotional label thresholds were selected in the dimensions of arousal and valence respectively to classify different emotional states.Then,the EEG signals were segmented through the time window,and the emotional information contained in the EEG data was mined by using the deep auto-encoder,and the emotional characteristic sequence was generated according to the time sequence.Finally,the LSTM recurrent network was established for model training,cross validation and testing,and the effect of emotion recognition was evaluated by accuracy,precision,recall and F1-Score.The simulation results show:in the valence dimension,the average accuracy of EEG emotion recognition and F1-Score reaches 77.4%and 80.4%,respectively.In the arousal dimension,the average recognition accuracy and F1-Score reaches 73.7%and 77.5%,respectively.This method has good recognition results for continuous emotion model and can be used for emotion computing and human-machine emotional interaction.