CASIA OpenIR  > 毕业生  > 硕士学位论文
基于EMD与增强Poincaré散点图的生理信号多尺度分析与识别
梁嘉琪1,2
学位类型工学硕士
导师郭大蕾
2018-05-24
学位授予单位中国科学院研究生院
学位授予地点北京
学位专业控制理论与控制工程
关键词经验模态分解 增强庞加莱散点图 特征提取 生理信号 多尺度分析与识别
摘要    多尺度现象是客观世界的普遍现象,但多尺度科学是近几年才提出的。多尺度科学是一门研究各种不同空间尺度或时间尺度相互耦合现象的科学,是目前科学研究的热点和难点之一。在信号处理领域,传统的分析法只考虑信号在一个尺度上的复杂性,不利于全面了解系统规律,同时忽略了不同尺度上包含的信息与关联。本文将经验模态分解(Empirical Mode Decomposition, EMD)作为信号处理中多尺度研究的基础方法,通过将原始信号分解为多个时间尺度分量展开研究。
    由此,针对富含多尺度信息的生理信号,例如心电图(Electrocardiogram, ECG)、脑电图(Electroencephalogram, EEG)等,本文展开多尺度复杂性方法的研究,通过对生理数据的多尺度分解得到不同尺度上的信息,从而探究多尺度下不同生理状态的特征及其差异,最终从定量与定性两个角度完成不同生理信号不同生理状态下的多尺度分析,提高分析结果的直观性、客观性和准确性,具体工作如下:
    对于生理信号,研究了其从时域到频域再到非线性分析方法的发展过程以及多尺度分析方法,最终以EMD与庞加莱散点图两类方法为基础对信号进一步展开研究。首先,通过EMD将生理信号分解为多个尺度,在各尺度分量上提取包括能量、熵等在内的特征参量,用于探究不同生理状态下的非线性特征差异。
    其次,针对庞加莱散点图忽略了数据点的分布信息的缺陷,提出增强庞加莱散点图,在反映数据相关性的同时增强描述其分布信息,提升直观性,揭示生理信号的非线性波动规律。在此基础上与多尺度思想结合,创建生理信号的多尺度增强庞加莱散点图,将不同尺度信息以相空间图形化分析方法展现,从定性的角度进行不同生理状态下的分析,直观展现不同尺度下的差异特征。定量方面,在提取传统庞加莱特征参数的同时,结合实际图形的展现,定义并提取四向峰度与四向偏度用以描述信号的分布信息,更全面的展现生理状态差异。
进一步结合多尺度思想和增强分布表达,构造生理信号的多尺度增强差分庞加莱散点图,在直观展现数据差分特征的同时,提取本文定义的多尺度差分庞加莱分布熵,改善以往缺乏尺度信息以及细节分布情况的计算形式。除此之外,还构造了多尺度增强延迟庞加莱散点图并提取针对延迟性特征的参数,将它们应用于具体数据集上进行对比分析。
    最后,对上文所提取的全部多尺度特征和其多尺度关联特征进行排序,通过从优选择不同个数的特征组成特征子集并用于支持向量机(Support Vector Machine, SVM)分类器的训练,从而获得最优的特征子集和最佳的分类器模型。在肌电信号、ECG等数据集上进行相应实验,获得针对不同生理信号的最优特征子集与分类器,再在测试集上测试其性能,说明多尺度分析给生理信号研究带来的重要意义。
本文搭建了具有普适性的生理信号多尺度分析与识别框架,利用最优分类模型与庞加莱快速视觉判断进行疾病或者生理状态的识别,从定性与定量两个角度完成生理信号的多尺度分析与识别。
其他摘要The Multiscale phenomenon is common in the objective world, but multiscale science is an emerging scientific field proposed in recent years. Multiscale science is the study of phenomena which couple distinct length or time scales and it is one of the hotspots and difficulties in current scientific research. In the field of signal processing, the traditional analysis methods only consider the complexity of physiological signal on the single scale, which is not conducive to a comprehensive understanding of the dynamics of the system and ignores the information and relevance in different scales. This paper employed Empirical Mode Decomposition (EMD) as a basic scaling method to decompose the original signal into multiple time-scale components for further research.
Thus, focusing on physiological signals which are rich in multiscale information, such as electrocardiogram (ECG), electroencephalogram (EEG), and etc., this topic aims to develop a multiscale complexity analysis method. Through multiscale analysis of physiological data, multiscale features on different scales can be extracted. Based on these multiscale features, the differences between different physiological conditions can be explored. This multiscale analysis method can be applied to various physiological signals to recognize different physiological conditions and improves the intuitiveness, objectivity, and accuracy of the analysis results. The specific work is shown as follows:
For physiological signals, the time domain analysis, frequency domain analysis, nonlinear analysis, and multiscale analysis methods were studied. Finally, EMD and Poincaré Plot are mainly researched and employed as the basic methods in this paper. Firstly, the physiological signal is decomposed into several scales. For each time scale, features including energy and entropy were extracted to explore the non-linear characteristics under different physiological conditions.
Secondly, taking into account of the drawbacks for Poincaré Plot, the enhanced Poincaré Plot is proposed. This method can display the probability distribution of each point while maintaining the data correlation, which is intuitive for revealing the fluctuations of data. Moreover, combined with the multiscale idea, Enhanced EMD Multiscale Poincaré Plot (EEMP) was introduced in which the characteristics of each scale can be presented in the way of phase space graphical. It provided a quick visual identification of different physiological conditions with a qualitative perspective. Meanwhile, a quantitative multiscale analysis is also indispensable. In addition to the traditional Poincaré Plot features, new parameters called four-direction kurtosis and four-direction skewness were defined to describe the data distribution, with which the physiological state can be described more comprehensively.
Thirdly, Enhanced EMD Multiscale Modified Poincaré Plot (EEMMP) is created for exposing differential characteristics of the physiological signal and EMD Multiscale Modified Poincaré Plot Distribution Entropy(EMPDE) was defined to improve the previous limitation and describe the details of the data distribution. In addition, Enhanced EMD Multiscale Lagged Poincaré Plot (EEMLP) is also proposed and the parameters for EEMLP are extracted to show the delay characteristics of the physiological signal. EEMMP and EEMLP both are applied to the specific database for comparative analysis.
Ultimately, all the scale features and their relevance features are ranked and a different number of the optimal features are selected to form various features subsets. After the support vector machine (SVM) classifier training, the optimal feature subset and the best SVM classifier model can be obtained. The corresponding experiments are performed on the datasets of electromyography and ECG. According to the trained model, the performance is then tested on the test dataset, which indicates the importance of multiscale analysis in the physiological signal research.
This paper builds a universal multiscale analysis and recognition framework of physiological signals, which uses the optimal classification model and Poincaré Plot rapid visual judgment to identify diseases or physiological conditions from both qualitative and quantitative perspectives.
语种中文
文献类型学位论文
条目标识符http://ir.ia.ac.cn/handle/173211/20931
专题毕业生_硕士学位论文
作者单位1.中国科学院自动化研究所复杂系统管理与控制国家重点实验室
2.中国科学院大学
推荐引用方式
GB/T 7714
梁嘉琪. 基于EMD与增强Poincaré散点图的生理信号多尺度分析与识别[D]. 北京. 中国科学院研究生院,2018.
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
梁嘉琪-硕士毕业论文-签字版.pdf(8267KB)学位论文 暂不开放CC BY-NC-SA请求全文
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
[梁嘉琪]的文章
百度学术
百度学术中相似的文章
[梁嘉琪]的文章
必应学术
必应学术中相似的文章
[梁嘉琪]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。