基于LabVIEW的心音信号分类识别系统设计

发布时间：2018-02-27 14:32

本文关键词： 心音 LabVIEW 预处理模块特征提取模块模型训练与识别模块　出处：《重庆大学》2014年硕士论文　论文类型：学位论文

【摘要】：心血管疾病已给人类带来了新的严峻挑战，并已经成为全球性公共卫生问题。心音是心脏和心血管系统机械运动状况的反映，因此，心音检测逐步成为临床辅助诊断心血管疾病的有效方法之一。对心音信号的临床检测常用方法是心音听诊和心音图，但是心音听诊技术容易受到人耳听力灵敏度和临床医师主观经验的影响，心音图能够弥补心音听诊的一些不足，但是它也存在着一些缺点，在一定程度上限制了它的应用。近年来，随着计算机和信号处理应用的逐步推广，心音分析仪的设计与开发已成为心音分析领域的发展趋势。目前出现的心音分析仪多数虽然可以对心音进行简单的时频分析，但是基本上都不具有对心音的分类识别功能，因此，心音分析仪的功能有待于进一步完善。鉴于此，本文在LabVIEW平台上设计了一款心音分类识别系统。本文首先给出了系统设计的整体方案，该系统一共分为三个子系统：预处理模块、特征提取模块、模型训练与识别模块。预处理模块，又分为以下四个模块：去噪模块、预加重模块、分帧加窗模块和端点检测模块。其中，，去噪模块是利用小波变换去噪算法来设计的，通过实验确定了小波母函数、分解尺度和阈值的合适选取；预加重模块是通过一个一阶数字滤波器来实现的；利用哈明窗设计了分帧加窗模块；端点检测模块是基于短时能量和短时过零率的原理来实现的。其次，在特征参数提取模块，本文介绍Mel频率倒谱系数及对其改进的其它四个参数：Mel频率倒谱系数结合其一阶差分系数、Mel频率倒谱系数结合其Delta特征、小波包变换改进的Mel频率倒谱系数、小波包变换改进的Mel频率倒谱系数结合其一阶差分系数(DWPTMFCC+ΔDWPTMFCC)的提取原理，同时详细分析了它们在LabVIEW平台上提取的难点并给出了解决方案。最后，在模型训练与识别模块，本文选取常见的识别模型——高斯混合模型(GMM)用于心音信号的分类识别。介绍了GMM的原理，分析了传统GMM的参数初始化算法K-means算法存在的缺点和不足，针对此，提出了三种算法：近似模糊C均值聚类算法、加权模糊C均值聚类算法和加权可选择模糊C均值聚类算法(WOFCM)对传统的GMM加以改进。此模块的关键点是这四个识别模型的训练过程与识别过程在LabVIEW平台上的实现与设计。将以上各个子模块集成在一起，就可以完成最终整体系统的设计。从特征参数和识别模型这两个角度对设计的系统进行测试，即对本文选取的临床上采集的正常心音信号和二尖瓣狭窄、主动脉瓣狭窄、主动脉瓣关闭不全、室间隔缺损、肺动脉瓣狭窄、心律不齐、二尖瓣关闭不全、S1分裂、S2分裂9类病理信号进行分类识别，综合考虑识别率和识别时间，当DWPTMFCC+ΔDWPTMFCC作为特征参数和WOFCM改进得到的GMM作为识别模型时表现出最为优越的识别性能，尤其是对异常心音信号而言，提高程度更为显著。因此，利用该参数和模型对最初设计的整个系统进行了简化处理，得到了最终的心音分类识别系统，这对目前的心音分析仪的功能进行进一步的补充完善，在研究心脏活动和早期的临床心血管疾病的诊断中有潜在的应用价值。
[Abstract]:Cardiovascular disease has brought new challenges, and has become a global public health problem. The heart sound is reflected in the heart and cardiovascular system of mechanical motion so that the heart sound detection has gradually become one of the effective methods for clinical diagnosis of cardiovascular disease. The common clinical detection method of heart sound signal is auscultation and phonocardiogram, influence but auscultation technique is vulnerable to human hearing sensitivity and the clinician subjective experience, PCG can remedy some shortcomings of auscultation, but it also has some shortcomings, which limits its application to some extent. In recent years, with the popularization of computer and signal processing applications, design and development of heart sound analyzer has a trend in the field of heart sound analysis. Although the majority of the heart sound analysis can be carried out at the frequency of heart sound simple However, the function of heart sound analyzer needs to be further improved. In view of this, a heart sound classification and recognition system is designed on the LabVIEW platform.
This paper first gives the overall scheme of system design, the system is divided into three subsystems: pre-processing module, feature extraction module, model training and recognition module. The pretreatment module, is divided into the following four modules: denoising module, preprocessing module, window module and endpoint detection module. The denoising module is to design the denoising algorithm using wavelet transform, wavelet function was determined by the experiment, the appropriate selection of scale and threshold decomposition; preprocessing module is realized by a digital filter; design of window module using Hamming window; endpoint detection module is the principle of short time energy and short-time zero crossing rate based.
Secondly, in the feature extraction module parameters, this paper introduces the Mel frequency cepstrum coefficient and the improvement of the other four parameters: the first order differential coefficient with Mel frequency cepstral coefficients, Mel frequency cepstral coefficients with the Delta feature and Mel frequency cepstrum coefficients of wavelet packet transform is improved, the first-order differential coefficient with frequency of Mel cepstrum coefficients of wavelet packet transform improved (DWPTMFCC+ DWPTMFCC) extraction principle, we analyze the problems and give them the extraction on the LabVIEW platform solutions.
Finally, in the model training and recognition module, this paper selects the recognition of Gauss mixture model (GMM) model commonly used for classification and recognition of heart sound signals. This paper introduces the principle of GMM, analyzes the existing parameter initialization algorithm of traditional K-means algorithm GMM the shortcomings and deficiencies, in view of this, we proposed three kinds of algorithms: approximate fuzzy C means clustering algorithm, weighted fuzzy C means clustering algorithm and weighted fuzzy C mean clustering algorithm (WOFCM) for the GMM to be improved. The key point of this module is the design and realization process of training and recognition of the four identification model on the LabVIEW platform.
Integration of the above sub modules together, we can achieve the design of the whole system. The final design of the testing system from the two aspects of feature parameters and recognition model, namely to collect clinical selected on the normal heart sound and mitral stenosis, aortic stenosis, aortic insufficiency, ventricular septal defect, arrhythmia, pulmonary valve stenosis, mitral regurgitation, S1 division, S2 division of 9 kinds of pathological signal classification, considering the recognition rate and the time when the DWPTMFCC+ Delta DWPTMFCC as characteristic parameters and WOFCM improved GMM as the recognition model showing the most superior recognition performance, especially in terms of abnormal heart sound signal, improving the degree is more significant. Therefore, the model parameters and the initial design of the whole system is simplified, the final classification of heart sound recognition system It will further improve the function of the current heart sound analyzer, and has potential application value in the study of cardiac activity and early clinical cardiovascular disease.

【学位授予单位】：重庆大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN911.7

【参考文献】