【作者】 李洪伟；

【导师】 李海峰；

【作者基本信息】 哈尔滨工业大学 ， 计算机科学与技术， 2018， 硕士

【摘要】 脑是人体中最复杂的器官,对大脑神经机制的研究是人工智能的基础,是实现人工智能的重要途径。情感是人的一种综合状态,计算机具备智能的标志之一是具有识别情感的能力。目前,对情感的认知与识别多是同通过视频、图像或短时声音完成的。通过长时间音乐欣赏来探究情感的研究较少。因此本课题让被试进行长时间的音乐欣赏,诱导被试产生相应的情感,记录情感产生过程中的脑电信号。对该脑电信号,分析人在音乐欣赏过程中关于情感的脑认知机制,并将脑认知机制用于情感识别工作中,达到提高情感识别率的目的。本文完成的工作主要有。通过长时音乐欣赏的脑认知实验采集相应的脑电数据。传统的脑认知实验大多都是基于短时声音信号的,缺乏基于长时音乐诱发情感的脑电信号数据。因此,设计基于长时间音乐欣赏的脑认知实验用于采集相应的脑电信号,并对数据进行了预处理,用于后续的情感分析与识别。针对长时音乐欣赏诱发的脑电信号难以分析的问题,提出了音乐事件点这一概念,进而使用事件相关电位技术分析长时音乐欣赏诱发的脑电信号。经过统计学和认知科学验证,证实通过音乐事件点在长时音乐欣赏诱发的脑电信号中提取到了ERP成分。使用该方法分析了不同情感状态下ERP波形的差异。发现在长时间音乐欣赏过程中,ERP的N1,P2成分在不同情感状态下差异显著,这说明在听觉情感的早期,不同情感对应的脑活动已经出现明显差异。对脑电信号的时频特性进行分析,探索音乐欣赏过程中脑的情感认知规律,并将发现的认知规律用于情感识别中。通过分析发现与情感相关的主要脑区为额区、中央区和枕叶区,与情感相关的脑电频段为delta频段、theta频段、alpha频段和gamma频段,并根据相关脑区和频段进行特征提取和优化,在二分类下达到71.2%的识别率,在三分类问题下,也得到了较高的识别率。针对传统脑网络研究中忽略了脑的实时性和动态性的问题,提出了动态脑网络用于分析长时音乐欣赏诱发的脑电信号。在长时间音乐欣赏过程中,脑的连接性会不断地发生变化。我们使用互信息对脑电信号的不同频带构建动态脑网络,观察脑网络随时间的变化并用于情感识别。我们使用脑网络进行情感分类,四分类下情感识别率达到了67.3%,超过了目前的最高识别率。对整个课题进行总结,本课题提出音乐事件点与动态脑网络两种创新方法用于分析脑电信号。通过探究脑认知规律,将脑认知规律应用于实际分类问题中,达到提高识别率的效果。经过验证,这种研究思想确实可行,通过脑认知机制提高计算机的识别率是实现人工智能的重要手段。更多还原

【Abstract】 The brain is the most complex organ in the human body.The study of the brain’s neural mechanisms is the basis of artificial intelligence and an important way to realize artificial intelligence.Electroencephalogram(EEG)has been an important means for people to explore the brain since its discovery.Emotion is a kind of integrated state of a person.One of the signs that computers possess intelligence is the ability to recognize emotions.Analyzing and identifying emotions has become an important cross-disciplinary research topic that spans multiple fields.In this topic,the subject is asked to evoke the corresponding emotions through long-term music appreciation,and the EEG signals are recorded during the evoked course.Then we analyzed the EEG signals to analyze the brain recognition mechanism in the process of human music appreciation and used in emotion recognition to achieve the purpose of improving the recognition rate of emotion recognition.The work done in this paper mainly includes:Designed a brain-cognitive experiment based on long-term music to collect corresponding EEG data.Traditional brain-cognitive experiments are based on short-term sound signals,lacking EEG signals based on long-term music-induced emotions.Therefore,we designed the brain cognitive experiments by ourselves,to collected the corresponding EEG signals,and preprocessed the data,and constructed an EEG signal data set based on long-term music-induced emotions.It is difficult to analyze the EEG signals evoked by long-term music,we propose the concept of music event point so that event-related potential technology can be used to analyze the EEG signals evoked by long-term music.After statistics and cognitive science verification,we are convinced that we extract ERP components from EEG signals evoked by long-term music.Subsequently,we used this method to analyze the differences in ERP waveforms under different emotional states.In the process of music appreciation for a long period of time,the N1 and P2 components of ERP differ significantly in different emotional states.This shows that at the early stage of auditory emotions,the brain activities corresponding to different emotions are significantly different.Analyzed the time-frequency characteristics of EEG signals to explore the cognitive mechanisms of the brain during music appreciation and used it in actual emotion classification.We found that the main brain regions related to emotion inducing are the central area and the frontal area.The main emotion-related EEG frequency bands are delta band,theta band,alpha band,and gamma band.And feature extraction and optimization are performed according to relevant brain regions and frequency bands.Under the two classifications,the recognition rate of 71.2%.This is the highest recognition rate.Under the three classification problems,the recognition rate is also higher.Propose the concept of dynamic brain networks to analyze the EEG signals evoked by long-term music.The traditional brain network research neglected the real-time and dynamics of the brain.In the process of long-term music appreciation,the connectivity of the brain changes constantly.Therefore,the characteristics of the corresponding brain network will continue to change.We use mutual information to construct a dynamic brain network o and observe changes in the brain network over time.We found that through the dynamic brain network,emotions can be effectively divided.We use the brain network to classify emotions.Under the four categories,the emotion recognition rate reached 67.3%,exceeding the current highest recognition rate.Summarized the whole topic.This topic proposes music event and dynamic brain networks to analyze the EEG signals.Explores the law of brain cognition and applies the law of brain cognition to practical classification problems to achieve the effect of improving the recognition rate.After our verification,this research idea is indeed feasible.Increasing the computer’s recognition rate through the brain cognitive mechanism is an important means of realizing artificial intelligence.更多还原