CASIA OpenIR  > 毕业生  > 博士学位论文
隐秘信息检测中脑电信号分析方法的研究
刘祺
2017-05-26
学位类型工学博士
中文摘要危害公共安全、妨碍社会正常秩序的涉贪涉恐涉黑犯罪是目前最为严重的犯罪行为,作案手段极其复杂,打击难度大,对公民和国家造成了极大损失。在实际侦查中,这类犯罪的作案人往往具备丰富的经验和较强的心理素质,贯穿整个过程的大脑思维活动是这些犯罪的显著特征。因此,与传统的犯罪心理侦查方法相比,利用能够反映大脑生理功能和思维状态的脑电信号对作案人试图隐藏的案件相关信息进行检测是一种更加科学有效的手段。传统的基于脑电的犯罪心理测试分析方法主要以人工目视为主,效率低下,主观依赖性强,早已不能满足实际应用中的需求。由于任务的复杂性以及涉案人的主观性,采集到的脑电信号背景噪声强,认知成分不明确,功能区域不明显,加大了对其中隐秘信息检测的难度。本文主要基于特定思维任务下的诱发脑电位,基于改进的双探测刺激复合反应实验范式,开展对隐秘信息检测领域中的脑电信号处理方法的研究,主要工作和贡献如下:
(1)   在对认知任务下的脑电信号进行盲源分离的基础上,引入了样本熵参数来寻找眼电伪迹对应的独立分量,减少了伪迹去除过程对先验知识和主观判断的依赖。在实现伪迹信号自动识别的同时,为满足应用中的实时要求,加入了样本熵快速计算方法,用以处理持续时间较长的信号段。
(2)   采用基于小波包分解的方法对脑电信号进行特征提取。利用小波包分解后的子带能量和反映信号复杂度的小波包熵构成特征向量,并事先对导联进行选择,避免了因使用导联数过多造成的特征向量维数过大的问题。
(3)   建立了基于相似性度量学习的个人隐秘信息脑电特征分类器模型,充分利用了不同类脑电数据间的区别性信息以及同类数据间的相似性信息,取得了高正确率的检测效果。为缩短训练时间,应对实际应用场景中可能出现的训练样本数量不足的情况,从理论上提出了在线的基于度量学习的分类器训练方法,以实现模型的在线更新,降低对侦查过程中相关刺激信息搜集数量的要求。
(4)   为满足实案侦查中对检测效率的要求和被试涉案程度检测的需求,提出了一种加权多尺度小波核极限学习机方法,借由模糊化因子的引入,实现了更好的分类能力。实验结果证明该方法能够有效减少模型训练时间且分类效果稳定。另外,采用组合核函数,提出了一种在线增量核极限学习机,以实现检测模型的在线调整,提高分类器的学习能力和泛化能力。
(5) 提出一种将深度置信网络用于犯罪信息检测中的脑电特征提取方法,无监督的学习到了脑电信号的深度特征,基于基础的分类模型,实现了对隐秘信息的有效检测,提高了特征表达的自适应性,降低了对分类器模型的要求,为将来实案侦测任务中,面对不同场景、不同被试群体下的应用环境打下基础。
英文摘要The public safety and the normal order of society are seriously threatened by criminal behaviors involving corruption, terrorist and mafia-like. Methods of these crimes are extremely complex. The crimes are difficult to combat and cause great losses for the citizens and the country. In actual investigation, the perpetrators of such crimes usually have a wealth of experience and strong psychological quality. Brain activities play a significant role in these crimes. Therefore, compared with the traditional criminal psychological investigation methods, the use of EEG which can reflect the brain function and mental state of physiological to detect relevant information the perpetrators tried to hide is a more scientific and effective means. The traditional method of criminal psychological analysis based on EEG is mainly based on artificial visualization. Low efficiency and subjective dependency of which cannot meet the needs of practical applications. Due to the complexity of the task and the subjectivity of suspects, the background noise of collected EEG signals is strong. Moreover, the cognitive component is not clear. And the functional area is not obvious. All of the above make the difficulty of the hidden information identification increased. Based on the improved DPCTP experimental paradigm, this paper mainly studies the EEG signal processing methods in the field of hidden information recognition. The main works and contributions of this paper include:
  1. Based on the blind source separation of the EEG signals under the cognitive task, the sample entropy is introduced to find the independent components corresponding to the artifacts, which reduces the dependence on the prior knowledge and the subjective judgment. The automatically identify of artifacts is realized, meanwhile, a fast calculation method of the sample entropy is introduced to deal with the signal segment of long duration in order to meet the real-time requirements in the application.
  2. The method of wavelet packet decomposition is adopted to extract features of EEG segments. The feature vectors are comprised of the sub-band energy getting through the wavelet packet decomposition and the wavelet packet entropy which reflects the complexity of signals. The leads used are selected in advance, which avoids the problem that the eigenvector dimension is too large.
  3. Taking advantages of the distinguishing information between samples belong to different classes and the similarity information among samples belong to the same class, this paper models a new EEG classifier based on similarity measure learning. To shorten the training time and deal with an insufficient number of samples in actual application scenarios that may arise, online training method based on metric learning is presented to achieve an online model update and to reduce the requirement of gathering relevant information during investigation process.
  4. To study on the detection of the degree of crime involvement and to meet the real-time requirements in the application, a weighted Multi-wavelets kernel Extreme Learning Machine method is proposed. By introducing the fuzzy factor, better classification ability is achieved. Experimental results show high efficiency and stable classification effect. In addition, the combined kernel function is adopted to obtain a new online incremental kernel Extreme Learning Machine. The learning and generalization ability of the classifier are improved.
  5. This paper presents a feature extracting method in the detection of the crime information based on EEG by using the deep confidence network. The features of EEG signals are unsupervised learned by deep learning. The effective detection of the hidden information is realized based on basic classifier model, the adaptability of the expression is improved, and the requirements of the classifier model are reduced, which lay the foundation for applications of practical cases detection tasks with different scenarios and different subjects in the future.
关键词隐秘信息检测 脑电信号 多解析度分析 相似性度量 极限学习 深度学习
文献类型学位论文
条目标识符http://ir.ia.ac.cn/handle/173211/14823
专题毕业生_博士学位论文
作者单位中国科学院自动化研究所
第一作者单位中国科学院自动化研究所
推荐引用方式
GB/T 7714
刘祺. 隐秘信息检测中脑电信号分析方法的研究[D]. 北京. 中国科学院研究生院,2017.
条目包含的文件
文件名称/大小 文献类型 版本类型 开放类型 使用许可
隐秘信息检测中脑电信号分析方法的研究_f(5070KB)学位论文 限制开放CC BY-NC-SA
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[刘祺]的文章
百度学术
百度学术中相似的文章
[刘祺]的文章
必应学术
必应学术中相似的文章
[刘祺]的文章
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。