远程心电张量特征抽取与分析

发布时间：2018-01-18 03:03

本文关键词：远程心电张量特征抽取与分析　出处：《上海交通大学》2015年博士论文　论文类型：学位论文

【摘要】：随着物联网技术,智能移动设备的不断普及,大规模远程心电的诊断平台不断成熟并且被广泛使用。由于远程心电诊断平台的规模巨大,自动的辅助诊断就显得尤为的重要,尤其是危重和急症病例的识别。本文的重点就是从去噪,预处理,特征抽取,特征降维,分类的整个流程中所涉及的基础理论和关键技术方面的问题。本文主要从辅助心电分析的角度出发,研究如何结合心电的特点,解决心电去噪、预处理、特征抽取、特征降维、特征分类和模式识别等问题,并且结合心电的不同特点提出相应的实现算法。并最终将这些方法应用到心电的辅助诊断和分类问题上。本文的主要工作和创新点包括以下几个方面:1.对于心电预处理过程来说,去噪是至关重要的环节。传统的心电去噪方法往往带来心电波形的走样和关键特征的丢失。另外,心电12导联信号本身存在着很大的冗余。这里我们提出了一种新的方法充分利用了心电信号本身的冗余性特点。我们重构出心电的2d和3d心电向量信号,然后通过投影重新获取多组原始心电信号,我们充分利用了波形中重要波形所在位置的信息,利用基于先验知识的加权主成分分析方法来去除噪声并且从重建心电信号中抽取出有用波形。2.从病理学角度出发,使用人工诊断过程中使用的特征进行分析有很大的困难。不同的心电疾病有着不同的心电波形,而且对于心电来说,有着很多种类的疾病,所以要从心电中准确无误地抽取出心电的病理学特征非常困难。因此我们尝试从机器学习的角度出发,从数据中学习出机器可以理解和容易处理的特征来进行分析,最终得到好的分析效果。除此之外,心电信号在频率上存在着一些对分类非常有用的特征,所以我们尝试在时频空的复合域上进行分析。所以我们提出了一些基于张量和多线性分析的方法直接在张量空间对数据进行分析,来尝试克服张量空间里特征稀疏和其它相关问题。3.张量算法的通病是目标函数非凸,容易落入局部最小解等问题。为了能够以大概率得到全局最优解。这里我们提出了一种计算框架,对于带约束的和不带约束的张量问题来进行求解。使用我们的方法,和其它张量学习算法所遇到的收敛困难等问题可以不同程度得到改善,而且往往可以得到更优的解。4.对于张量特征抽取算法,分为基于T2V映射并且以向量为输出的还有基于T2T映射并且以张量为输出的。本文提出的张量特征抽取算法基本上都是基于(T2V)的。虽然对于以张量为输出的张量特征抽取算法,我们可以对它的张量输出进行向量化然后使用现有向量空间分类器进行分类,但是这样也会面临结构信息丢失、参数过多过拟合、小样本问题等等。所以我们提出了一些可以直接以张量为输入的特征分类方法,最后和其它的方法进行比较。以上几项工作都是基于心电的特点提出的去噪、特征抽取和分类算法,本研究在给出模型框架的同时还给出了具体实现算法,并针对各种应用问题进行了实验分析。
[Abstract]:With the networking technology, the popularity of smart mobile devices, large-scale remote ECG diagnosis platform continues to mature and has been widely used. Because the remote ECG diagnosis platform of large scale, automatic diagnosis is particularly important, especially the identification of critical cases. This paper focuses on denoising preprocessing, feature extraction, feature reduction, basic theory and the key technical problems involved in the whole process of classification. This paper mainly from the auxiliary ECG analysis point of view, study how to combine the characteristics of ECG, solve the ECG denoising, preprocessing, feature extraction, feature reduction, feature classification and pattern recognition etc. and, according to the different characteristics of the proposed ECG algorithms. And finally apply these methods to diagnosis and ECG classification problems. The main work of this paper and the innovation package Including the following aspects: 1. for ECG preprocessing, denoising is very important. The loss of the traditional ECG denoising method often leads ECG waveform aliasing and key characteristics. In addition, 12 lead ECG signal itself has great redundancy. Here we proposed a new method to make full use of the the redundancy characteristic of ECG signal itself. We reconstructed the 2D ECG and VCG 3D signal, and then through the projection to re acquire multiple sets of the ECG signal, we make full use of the important position of the waveform waveform information, using weighted principal component analysis method based on prior knowledge to remove noise and extract useful waveform.2. starting from the perspective of pathology from reconstruction of ECG signals, the characteristics of using artificial diagnosis in the process of analysis is very difficult. The ECG of different diseases ECG waveform, and the ECG, there are many kinds of diseases, so from ECG accurate to extract ECG pathological features very difficult. So we try from the view of machine learning, data from the learning machine can understand and easy processing characteristics to analyze, finally get the analysis good effect. In addition, there are some features of ECG signal classification is very useful in frequency, so we try to analyze the time-frequency domain composite empty. So we put forward some analysis methods based on multilinear tensor and tensor space directly in the analysis of the data, to try to overcome the defects of tensor in the space of sparse features and other related problems of.3. tensor algorithm is non convex objective function, easy to fall into the local minimum problem. In order to get the probability of global optimal solution. Here we propose a computational framework for constrained and unconstrained tensor problem to solve the problem. We use the methods, problems encountered in the learning algorithm convergence difficulties and other tensor can improved to a certain extent, and can often get a better solution for the.4. tensor feature extraction algorithm, divided into based on the T2V mapping and vector output are based on T2T mapping and output to the tensor. This paper proposes a tensor feature extraction algorithm is basically based on (T2V). Although for the output of the Zhang Liangte tensor feature extraction algorithm, we can output the tensor vector and then use the existing vector space the classifier, but it will also face the loss of information structure, parameter over fitting, the small sample problem and so on. So we can put forward some directly to the tensor Feature classification method of input, and finally compared with other methods. The above work is presented the characteristics of ECG based denoising, feature extraction and classification algorithm, this study is also given in the model framework and gives the realization algorithm, and the experimental analysis for various applications.

【学位授予单位】：上海交通大学
【学位级别】：博士
【学位授予年份】：2015
【分类号】：R540.41

【参考文献】