【作者】 刘鹏；

【导师】 乔晓艳；

【作者基本信息】 山西大学 ， 电子与通信工程， 2021， 硕士

【摘要】 机器自动情绪识别因其在人机交互中的潜在应用而受到越来越多的关注。目前,情绪识别可通过多种信息来实现,如面部表情、语音语调、身体姿态以及人体各种生理信号等。相比较而言,皮层脑电作为中枢神经生理信号,调节与情绪相关的递质分泌和大脑的边缘系统,能够客观反映人的情绪状态变化,从脑电信号中提取有效特征并进行情绪识别在人机交互和精神心理疾病监测中具有重要意义。论文针对维度情感模型的构建以及脑电信号情绪识别中特征提取与分类等关键环节展开相关研究,分别采用了信号处理方法提取特征以及深度栈式自编码(Deep Stacked auto encoder,Deep-SAE)自动生成特征,设计栈式自编码深度学习算法以及长短时记忆循环网络分类模型,探索不同方法和模型对脑电情绪识别的有效性。论文主要研究内容如下:(1)构建维度情感模型,基于DEAP维度情感生理数据集,从不同角度提取了情感脑电信号特征。分别为:利用自回归模型(Auto Regressive,AR)功率谱估计方法,提取脑电θ、α、β、γ节律的功率谱密度;采用小波包分解提取脑电的小波包系数以及能量占比时频特征;通过非线性分析方法提取脑电的样本熵和小波包熵特征。(2)设计栈式自编码神经网络深度学习算法,对提取的脑电特征在效价和唤醒度两个情感维度进行了机器情绪识别。此方法在DEAP数据集上的仿真测试结果为:在效价维度,情绪识别平均正确率达到80.3%;在唤醒度维度,平均识别正确率达到81.5%,获得了较高的情绪识别正确率。在此基础上,还分析了脑电特征组合、样本数据均衡以及维度情感标签阈值对识别结果的影响。(3)在生理信号情绪识别中,为了自动挖掘脑电信号所蕴含的情感特征信息,论文提出了深度栈式自编码(Deep-SAE)方法,直接从脑电数据中自动解码深层的情感信息,并生成特征序列。通过对多导联脑电信号仿真测试,经过编码-解码过程,获得了多导联脑电数据有效解码的情感特征信息。(4)针对脑电信号情绪识别中分类算法导致的泛化能力差的问题,论文综合考虑了脑电信号具有的时序性和导联分布特点,对获取的不同导联脑电特征序列,设计长短时记忆(Long short-term memory,LSTM)循环网络进行模型训练、交叉验证与算法测试,通过正确率、精确度、召回率和F1-分数等指标评估模型的性能。仿真测试结果为:在效价维度,脑电信号情绪识别平均正确率和F1-分数分别达到77.4%和0.804;在唤醒度维度,平均识别正确率和F1-分数分别达到73.7%和0.775。该模型较其它方法获得了较好的情绪识别泛化能力。论文研究的脑电信号情绪识别维度情感模型构建、特征提取与分类算法,为机器自动情绪识别提供了新的思路,研究成果在情感机器人、医疗健康、心理学、脑-机交互、情境学习以及多媒体游戏开发等领域有着重要的科学与应用价值。更多还原

【Abstract】 Machine automatic emotion recognition has attracted more and more attention due to its potential application in human-computer interaction.At present,emotion recognition can be realized through a variety of information,such as facial expressions,voice intonation,body posture,and various physiological signals of the human body.In comparison,cortical electroencephalography(EEG),as a physiological signal of the central nervous system,regulates the secretion of emotion-related transmitters and the limbic system of the brain,and can reflect changes in human emotional state objectively.Extracting effective features from EEG signals and performing emotion recognition are of great significance in human-computer interaction,mental and psychological disease monitoring.This paper focuses on the emotion dimension model construct and the key links of EEG emotion recognition,such as feature extraction and classification.In order to explore the effectiveness of different method and model for EEG emotion recognition,signal processing method are used to extract features and Deep stacked auto-encoder(Deep-SAE)method for automatically generating feature respectively,and the classification method is designed with stacked auto-encoder deep learning algorithm and Long short-term memory(LSTM)recurrent network.The main research contents of this paper are as follows:(1)The dimensional emotional model was built.Based on the DEAP dimensional emotional physiology dataset,the characteristics of emotional EEG signals were extracted from different perspectives.The power spectral densities of θ,α,β and γ rhythms were extracted by AR(Autoregressive)model power spectrum estimation method.Wavelet packet decomposition was used to extract the wavelet packet coefficients and energy ratio time-frequency characteristics of the EEG.The characteristics of sample entropy and wavelet packet entropy of EEG were extracted by nonlinear analysis.(2)A stacked auto-encoder neural network deep learning algorithm was designed,and machine emotion recognition was performed on the extracted EEG features from the two emotional dimensions of valence and arousal.The simulation results of this method on the DEAP dataset show that: in the valence dimension,the average accuracy of emotion recognition reaches80.3%;In the arousal dimension,the average accuracy of recognition reaches81.5%,obtained a high accuracy of emotion recognition.On this basis,this paper analyzed the influence on results of emotion recognition in three aspects,including EEG feature combination,EEG data balance,and the dimensional emotion label threshold.(3)In order to automatically mine the emotional feature information from EEG signals in the emotional recognition of physiological signals.A Deep-SAE method was proposed,with this method,the deep emotional information can be automatic decoded directly from the EEG data,and the feature sequence can be generated consequently.Through the simulation test of multi-channel EEG signal,the emotion characteristic information of multi-channel EEG data is obtained using the method of encoding and decoding.(4)Aiming at the problem of poor generalization ability caused by classification algorithms in EEG emotion recognition.The temporal characteristics of EEG signals was considered,and the time feature sequence was generated from the EEG emotion features.Then,a long short-term memory recurrent network was designed for model training,cross-validation and testing.The performance of the model was evaluated through some indicators,including 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 0.804,respectively.In the arousal dimension,the average recognition accuracy and F1-Score reaches 73.7% and 0.775,respectively.Compared with other methods,this model has better generalization ability in emotion recognition.This paper provided a new idea for EEG signal automatic emotion recognition in the dimensional emotion models construction,feature extraction and classification recognition methods.Research results has important scientific and application value in emotional robot,medical health,psychology,brain-computer interaction,situated learning and multimedia game development.更多还原