基于非负矩阵分解的盲源分离算法在心电信号消噪中的研究

发布时间：2018-03-03 17:47

本文选题：盲源分离　切入点：非负矩阵分解　出处：《太原理工大学》2014年硕士论文　论文类型：学位论文

【摘要】：心电信号(Electrocardiogram,ECG)是人体的重要生理信号之一,其中包含着大量关于心脏的病变、病理状况的相关信息,也反映了心脏与心血管的结构及其生理和病理的特性。心电信号的分析诊断对心血管等系统疾病的诊断具有很重要的意义,其精确度和可靠度会直接地影响心脏疾病患者的临床医疗诊断和愈后的效果评价。然而,传统的心电识别方法是医生临床听诊,很显然此过程具有一定的主观判断和不稳定性,准确性比较差。现阶段,从人体体表获取的心电信号,或多或少都会受到工频干扰、肌电干扰以及基线漂移等多种噪声的干扰。因此消除心电信号的噪声,对之后临床研究的特征波检测以及病理诊断等需求都具有重要的意义。非负矩阵分解(Non-negative Matrix Factorization, NMF),作为一种新兴的特征分离方法,由Lee和Seung等人在盲源分离的应用背景下于1999年提出,并发表在Nature杂志上,并且慢慢发展成为了信号处理和数据分析的有效方法。通过在矩阵分解过程中加入非负的矩阵元素,非负矩阵分解使得分解结果呈现出的完全不同,完成了降维的非线性目标。随着盲源信号处理研究的逐渐加深,非负矩阵分解已经逐步成为信号处理、生物医学和图像处理等多个研究领域中最受学者青睐的数据处理工具之一。本文将非负矩阵分解应用于对心电信号的消噪,具有收敛速度快、稀疏性、非负性、降维等特性。在对基本NMF算法的学习中,NMF加人了非负的约束。这样,通过分解得到的基信号数据以及用于重构的权重系数都是非负的。在这种模式下,只允许线性叠加运算,这就保证了“局部构成整体”模式。因此,NMF被认为是提取局部特征的一种方法。但是,NMF算法得到的“部分”有时候并不是像我们预期的那样局部化,而且基本NMF方法在某些时候的识别率不是很高。出于对NMF原算法的深入学习,本人在研究局部信号数据时建立PNMF算法,其目的是通过引入稀疏性限制获得编码矢量(矩阵H)真正的局部分解对象,并使基本组件(矩阵W)局部稀疏化,加强基成分的局部化特征,使算法适用于局部特征非常重要的应用。本文结合NNF算法特点及心电信号特征,首次提出了一种新的NMF算法——PNMF对心电信号盲源分离。结合MIT/BIH国际标准数据库中ECG数据和模拟基线漂移、工频干扰以及肌电干扰噪声合成含噪声心电信号,并应用新提出的PNMF算法进行盲源分离实验研究,对分离结果采用信噪比(Signal to Noise Ratio,SNR)评价参数进行量化评价,与3种不同的NMF算法进行了对比,同时将PNMF算法与FastICA算法的的分离结果做比较,从分离精度的角度来看,本文的算法取得了最佳的效果。实验结果表明PNMF算法可有效分离心电源信号,为实际心电后期准确诊断提供了一定的参考依据。
[Abstract]:Electrocardiogramme (ECG) is one of the important physiological signals in human body, which contains a lot of information about the pathological changes and pathological conditions of the heart. It also reflects the structure of the heart and the cardiovascular system and its physiological and pathological characteristics. The analysis and diagnosis of ECG signals are of great significance for the diagnosis of cardiovascular diseases and other systemic diseases. The accuracy and reliability of ECG can directly affect the clinical diagnosis and evaluation of the curative effect of patients with heart disease. However, the traditional ECG recognition method is clinical auscultation, so it is obvious that this process has some subjective judgment and instability. The accuracy is poor. At this stage, ECG signals obtained from the body surface of the human body are more or less interfered by various noises such as power frequency interference, myoelectric interference and baseline drift. Therefore, the noise of ECG signals is eliminated. It is of great significance for the requirement of characteristic wave detection and pathological diagnosis in later clinical research. Non-negative Matrix factorization, as a new feature separation method, was proposed by Lee and Seung in 1999 under the background of blind source separation and published in Nature magazine. By adding non-negative matrix elements into the matrix decomposition process, the non-negative matrix decomposition makes the decomposition result completely different. With the development of blind source signal processing, the nonnegative matrix decomposition has gradually become signal processing. One of the most popular data processing tools in biomedicine and image processing. In this paper, non-negative matrix decomposition is applied to de-noising ECG signals, which has the characteristics of fast convergence, sparsity, non-negativity and dimensionality reduction. In the learning of the basic NMF algorithm, the nonnegative constraints are added. Thus, the base signal data obtained by decomposition and the weight coefficients used for reconstruction are non-negative. In this mode, only linear superposition operations are allowed. So NMF is considered to be a way to extract local features. But sometimes the "part" of the NMF algorithm is not as localized as we might expect. Moreover, the recognition rate of the basic NMF method is not very high at some times. In order to study the original algorithm of NMF, I establish the PNMF algorithm when studying the local signal data. The purpose of this algorithm is to obtain the real local decomposition object of the encoding vector (matrix H) by introducing the sparsity restriction. It also makes the basic component (matrix W) local sparse, strengthens the localization feature of the base component, and makes the algorithm suitable for the application of local feature. Based on the characteristics of NNF algorithm and ECG signal, this paper presents a new NMF algorithm for blind source separation of ECG signals for the first time, combined with ECG data and analog baseline drift in MIT/BIH international standard database. Power frequency interference and myoelectric interference noise are used to synthesize noise-containing ECG signals, and the new PNMF algorithm is used to study the blind source separation experiment. The signal to noise ratio (SNR) signal to Noise (SNR) evaluation parameters are quantitatively evaluated. Compared with three different NMF algorithms, the separation results of PNMF algorithm and FastICA algorithm are compared from the point of view of separation accuracy. The experimental results show that the PNMF algorithm can effectively separate the signal of cardiac power supply and provide a certain reference for the accurate diagnosis in the later period of ECG.
【学位授予单位】：太原理工大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN911.7

【参考文献】