基于EMD的舒张期心杂音信号的分析与识别研究

发布时间：2018-01-01 08:18

本文关键词：基于EMD的舒张期心杂音信号的分析与识别研究　出处：《重庆大学》2016年硕士论文　论文类型：学位论文

【摘要】：随着经济的快速发展,人类饮食结构的不断改变,心血管疾病的发病率和死亡率均在迅速上升,严重威胁着人类的健康和幸福生活。心音是由心脏机械运动产生的振动信号,蕴含着与心血管疾病有关的大量诊断信息,因此,心音分析对于无创诊断心血管疾病具有重要的价值。心音的分类识别作为心音分析领域中的一个研究热点,其目的是利用分类器根据从不同心音中提取的特征参数判定所属疾病类型,而目前心音的特征提取与分类方法大多数是基于心音信号线性时变或时不变模型,而心杂音作为一种非线性非平稳的信号,线性的分析方法势必会忽视信号内部的一些重要信息。因此,本文提出基于经验模式分解(Empirical Mode Decomposition,EMD)的舒张期心杂音信号的特征提取与分类方法。首先,在分析心音和心杂音产生的生理机制和临床意义的基础上,筛选舒张期心杂音信号作为实验对象,从而有效的避免生理性杂音的干扰。针对适用于分析非线性、非平稳性信号的EMD在分解过程中产生的端点效应,提出比例延拓结合镜像延拓的方法给予抑制,数值信号和心音信号的测试表明该方法可以有效的减轻端点效应对EMD分解的影响。其次,在心音去噪方面,利用小波变换方法来实现,小波的三个重要参数小波基函数、分解层数和阈值通过3组实验来确定;心音定位方面,在由希尔伯特变换获取信号包络的基础上采用双阈值的方法对心音准确定位。在心音特征值提取方面,提出了基于EMD的特征提取方法:在EMD分解获取固有模态函数(Intrinsic Mode Function,IMF)的基础上,采用互相关系数准则筛选出主IMF分量(IMF1~IMF4),分别提取其Mel频率倒谱系数(Mel Frequency Cepstrum Coefficient,MFCC)、MFCC的一阶差分系数(△MFCC)及Delta特征,将其组合形成3个特征向量MFCC、MFCC+△MFCC及MFCC+Delta,简记为E+MFCC、E+M+△MFCC及E+M+Delta;针对EMD分解过程中的模态混叠问题而影响特征参数提取的准确性,采用总体平均经验模式分解(Ensemble Empirical Mode Decomposition,EEMD)代替EMD的方法来对心音信号进行分解获取IMF分量,对筛选出来的主IMF分别提取MFCC、△MFCC、Delta特征,经组合形成三个特征向量MFCC、MFCC+△MFCC及MFCC+Delta,简记为EE+MFCC、EE+M+△MFCC及EE+M+Delta。最后,选择具有通过较少样本就能训练出较为可靠模型的隐马尔科夫模型(hidden Markov model,HMM)作为分类器,选取临床采集到的正常心音和两类舒张期心杂音(也即主动脉关闭不全和二尖瓣狭窄)作为实验对象,训练样本与测试样本的比例为1:2,利用所提取到的特征向量来建立模型进行心音的分类识别。最终实验结果表明,所提出的2种特征参数提取方法的识别性能均优于传统的MFCC。同时为了进一步验证所提出的端点延拓方法的有效性,将其用于EMD分解提取特征参数,实验结果表明,该延拓方法所获得的识别率要高于未经端点处理EMD方法的识别率。
[Abstract]:With the rapid development of economy and the change of human diet, the morbidity and mortality of cardiovascular disease are increasing rapidly. Heart sound is a vibration signal produced by mechanical movement of the heart, which contains a lot of diagnostic information related to cardiovascular disease. The analysis of heart sounds is of great value in the non-invasive diagnosis of cardiovascular diseases. The classification and recognition of heart sounds is a hot topic in the field of heart sound analysis. The purpose of this method is to use the classifier to determine the disease type according to the characteristic parameters extracted from different heart sounds. At present, most of the feature extraction and classification methods of heart sounds are based on linear time-varying or time-invariant models of heart sounds. As a kind of nonlinear and non-stationary signal, the linear analysis method will inevitably ignore some important information inside the signal. This paper presents a method for feature extraction and classification of diastolic cardiac murmur signals based on empirical Mode decomposition (EMD). On the basis of analyzing the physiological mechanism and clinical significance of cardiac murmur and cardiac murmur, the diastolic cardiac murmur signal was selected as the experimental object, so as to avoid the interference of physiological murmur effectively. The endpoint effect of the EMD of non-stationary signal in the decomposition process is suppressed by the method of proportional continuation combined with mirror continuation. The test of numerical signal and heart sound signal shows that this method can effectively reduce the effect of endpoint effect on EMD decomposition. Secondly, in the aspect of heart sound denoising, wavelet transform is used to realize it. Wavelet basis function, decomposition layer number and threshold value of three important parameters of wavelet are determined by three groups of experiments. In the aspect of heart sound localization, based on the Hilbert transform to obtain the signal envelope, the method of double threshold is used to locate the heart sound accurately, and the characteristic value of heart sound is extracted. A method of feature extraction based on EMD is proposed: the intrinsic Mode function is obtained by EMD decomposition. The main IMF component (IMF1 / IMF4) was screened by using the correlation number criterion. The Mel cepstrum coefficients of Mel Frequency Cepstrum coefficients were extracted respectively. The first order difference coefficient (MFCC) and Delta characteristics of MFCC are combined into three characteristic vectors: MFCC MFCC and MFCC Delta. The results were summarized as E MFCC E M MFCC and E M DeltaA. The accuracy of feature parameter extraction is affected by the modal aliasing problem in the process of EMD decomposition. Ensemble Empirical Mode Decomposition was decomposed by the overall average empirical mode. EEMD) replaces the EMD method to decompose the heart sound signal to obtain the IMF component, and extracts the MFCC, MFCC / Delta features from the selected master IMF. Three characteristic vectors, MFCC MFCC and MFCC Delta. are formed by combination, which can be abbreviated as EE MFCC. EE M MFCC and EE M Delta. finally. The hidden Markov model (HMMM), which can train a more reliable model with fewer samples, is selected as the classifier. The normal heart sounds and two kinds of diastolic murmur (that is aortic insufficiency and mitral stenosis) were selected as experimental subjects. The ratio of training sample to test sample was 1: 2. The extracted eigenvector is used to establish the model for the classification and recognition of heart sounds. Finally, the experimental results show that. In order to further verify the effectiveness of the proposed endpoint continuation method, the proposed method is used to extract feature parameters by EMD decomposition. The experimental results show that the recognition rate of the extension method is higher than that of the untreated EMD method.
【学位授予单位】：重庆大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：R540.4;TN911.6

【相似文献】