【作者】 李志鹏；

【导师】 原桂彬；

【作者基本信息】 哈尔滨工业大学 ， 仪器科学与技术， 2017， 硕士

【摘要】 情绪作为人脑的高级功能,在很大程度上影响着人类学习、工作、生活的各个方面。正确的识别人类情感可以使人工智能更好的服务于人类,因此情绪识别成为人工智能领域的一个十分重要的研究方向。目前情绪识别主要围绕人的外在行为特征和客观生理信号展开。相比人的外在行为特征,客观生理信号具有着更好的自发性和客观性。脑电信号作为生理信号一种,是由中枢神经系统产生,与人类情感联系最为紧密。对于脑电信号的情感识别研究,大多数实验根据全通道EEG信号展开,全通道EEG信号的采集不便于后续便携式设备的开发。此外情感识别分类算法多是采用浅层分类器,由于浅层分类器学习能力的限制,导致识别率存在瓶颈。本文针对以上两方面存在的问题围绕脑电信号的预处理、特征提取、通道选择、分类算法改进等做了以下几方面的研究:实际采集的EEG信号十分微弱,包含着伪迹和环境噪声干扰,本文首先通过预处理对EEG信号的噪声进行去除,获取纯净的脑电信号。针对EEG信号具有非平稳性,本文采用小波变换和信息熵的组合特征提取方法获得了992维的EEG组合特征。针对EEG采集实验需要采集大量脑电通道数据,导致实验设备的便携性差,数据处理复杂的问题。本文采用了BP+DEMATEL的脑电通道选择方法,选出了最优的脑电通道组合,并采用SVM对该算法的可行性进行了验证。经过通道选择,脑电数据由原先的62通道降到了8通道,极大的减少了实际采集EEG通道数量,并获得与情感相关的脑区位置,为后续的便携式可穿戴设备的研制奠定了基础。针对目前浅层分类器对于EEG信号情感识别率较低的问题,本文提出了一种新型的基于深度学习组合神经网络的脑电信号识别算法。该算法通过两个双层的RBM对脑电数据进行深度挖掘并采用识别率较高的GRNN进行数据分类,取得了87.1%左右的识别率。通过大量的对比实验验证了该算法对于EEG信号有着识别率高且抗干扰性强的特点。此外,降维方法对于情感识别的准确率也有着较大影响,本文对比了PCA及m RMR的降维方法,通过理论推导及实验分析说明了两种降维方法各自的特点。采用实验室自主研发的脑电采集装置,设计了基于视频诱发情感的脑电采集实验采集了脑电实验数据,并采用本文提出的情感分类算法对采集的EEG数据进行分类,根据实验的分类效果验证了算法的有效性。更多还原

【Abstract】 As a senior function of human brain,emotion has a great influence on human study,work,and all aspects of life.Recognizing human emotion correctly can make artificial intelligence serve human being better.Therefore,emotion recognition has been a very important research direction in the field of artificial intelligence.Up till now,emotion recognition mainly focuses on people’s external behavior characteristic and objective physiological signal.Compared with the external behavior characteristic,the objective physiological signal has better spontaneity and objectivity.As a physiological signal,EEG signal which is generated by the central nervous system,has most closely contact with human emotion.The experiments based on emotion recognition of the EEG signal is often implemented by the full channels EEG signals,and the full channels acquisition of EEG signal has limited the development of subsequent portable equipment.In addition,most emotion recognition classification algorithms use shallow classifier,which leads to the bottleneck of recognition rate.In this paper,based on the above two existent issues,around the brain electrical signal preprocessing,feature extraction,channel selection,and classification algorithm improvement,several aspects of research is proposed as follows:EEG signal is a kind of small-signal,and always include artifacts and noise jamming.In this paper,firstly,the noise signal is removed by preprocessing in order to get the pure EEG signal.To solve the problem that EEG signal has non-stationary characteristics,992 dimensions features are obtained by combining feature extraction method of wavelet transform and information entropy.Aiming at the problems of the poor portability of experimental equipment and complex data processing generated by the full channels EEG acquisition experiment,BP plus DEMATEL EEG channel selection method is proposed to select the optimal combination of EEG channel,and the SVM is adopted to verify the feasibility of the algorithm.The proposed channel selection method declines down the original 62 channels EEG data to 8 channels,greatly reduces the number of the actual EEG channels acquisition.In addition,the position where brain regions is associated with emotion most closely can be obtained,which laid the foundation for the subsequent development of portable wearable devices.To solve the problem that EEG signals emotion recognition rate based on the shallow classifier is low,this paper presents a new type of EEG signal recognition algorithm by combining deep learning and neural network.The algorithm uses two double-layer RBM to deeply mine the EEG data and utilizes GRNN for data classification,and achieves a recognition rate of about 87.1%.A large number o f comparative experiments make sure that the algorithm for the EEG signal has strong anti-interference and high recognition rate.In addition,dimensionality reduction methods also have great influence on the accuracy of emotion recognition.In this paper compares dimension reduction methods such as PCA and m RMR to illustrate the characteristics of them by using theoretical deduction and experimental analysis.EEG acquisition experiments based on video evoking emotion are designed by our research laboratory.The emotion classification algorithm is used to classify the collected EEG data,and experimental results testify the effectiveness of the algorithm.更多还原