【作者】 李耀光；

【导师】 崔巍；

【作者基本信息】 华南理工大学 ， 系统工程， 2019， 硕士

【摘要】 心电信号是人体心脏电生理活动的反映,自从上世纪初心电图机问世以来,人们一直致力于通过心电图进行心脏疾病的研究。随着信息技术的发展,利用计算机结合人工智能技术进行心电信号的自动分析已经得到广泛应用。然而由于心电数据采集环境的限制以及医生诊断可能存在的错误,心电信号自动分类领域存在训练样本错误标记的问题,会使得心电分类成为弱监督学习,分类器会被严重破坏,分类准确率明显降低。本文针对这一问题提出了基于交叉验证的识别方法,对训练集中可能存在的错误标记样本进行识别,提升了分类性能,取得了满意的效果。另一方面,由于心电信号相对稳定且防伪造和窃取能力强,其应用也不再仅限于心脏疾病诊断,而扩展到身份识别等领域。尽管已有不少相关工作,但是运动心电身份识别还有待进一步研究,该鲁棒性问题限制了心电身份识别的实际应用。针对该问题,本文建立了运动心电数据库,对运动心电身份识别的可行性进行了探讨。第一章,介绍本文的研究背景和研究意义,回顾了心电自动分析和心电身份识别的发展现状。第二章首先详细介绍了心电信号的产生机理和波形特点,然后介绍了心电图机的导联系统和心电研究领域广泛使用的数据库。在第三章中梳理了心电信号处理的主要流程。在心电预处理中介绍了小波变换法和中值滤波法。在心电波形检测中介绍了动态阈值法,在心电特征提取中介绍了时域特征、形态特征和频域特征。最后在分类算法中介绍了支持向量机、决策树、线性判别分析、朴素贝叶斯和k近邻算法。第四章则讨论在心电自动分类中存在的错误标记训练样本问题。针对该问题,本文针对前人提出的基于遗传算法的错误标记样本识别方法存在的问题,提出了一种基于交叉验证的多种机器学习算法结合的方法,并在经过人工添加错误标记噪声的MIT-BIH心律失常数据库上进行实验。结果表明经过基于交叉验证的识别方法去除错误标记训练样本后,分类性能指标得到明显改善。其中,以总体分类准确率为代表,如果训练集中错误标记样本的比例在20%以下,那么总体分类准确率可以回升至与没有人工添加标记噪声时相当的水平。如果错误标记样本比例达到30%,则总体分类准确率可以回升到比无人工添加标记噪声时略低的水平。而如果错误标记样本比例达40%,则经过本文方法处理之后的总体分类准确率仍然远高于不进行错误标记样本去除的情况。在第五章中,针对运动状态改变对于心电身份识别造成的困难,利用本课题组采集的运动心电身份识别数据库探讨运动心电身份识别的可行性。本章首先介绍了运动心电身份识别数据库的详细信息,并分析得到运动心电身份识别的关键在于提取出运动前后保持相对稳定的特征。然后将当前心电身份识别领域比较成熟的特征提取和选择方法在运动心电身份识别数据库上进行实验。实验结果表明现有文献提出的心电身份识别方法在静止状态下取得了不错的效果,但是运动状态改变的情况还有待进一步研究。更多还原

【Abstract】 Electrocardiogram(ECG)signals are the reflection of the electrophysiological activity of the human heart.Since the invention of electrocardiograph in the early 20 th century,people have been working on the research of heart disease through ECG.With the development of information technology,automatic analysis of ECG signals by computer and artificial intelligence has been widely used.However,due to the limitations of the ECG data acquisition environment and possible errors in the diagnosis by doctors,there are some wrongly labelled samples in the training set,which makes the classification of ECG become a weak supervised learning problem,the classifier be seriously damaged,and the classification accuracy be significantly reduced.Aiming at this problem,this paper proposes a cross validation based method to identify the mislabelled samples in the training set,thus improving the classification accuracy and achieving satisfactory results.Besides,because the ECG signals are relatively stable and very hard to steal and forge,their application is no longer limited to the diagnosis of heart disease,but extends to the field of identity recognition,too.Although a lot of related works have been done,the identification of ECG with exercise is still to be further studied.The robustness of exercise limits the practical application of ECG identification.In order to solve this problem,this paper establishes an exercise ECG database and discusses the feasibility of ECG identification with exercise.The first chapter introduces the background and significance of this study,and reviews the development of ECG automatic analysis and ECG identification.In the second chapter,the generation mechanism and waveform characteristics of ECG signals are introduced in detail.Then the lead system of electrocardiograph and the database widely used in ECG researches are introduced.In the third chapter,the main flow of ECG signals processing is analysed.Wavelet transform,median filter and notch filter are introduced in ECG preprocessing.Differential threshold method and Hilbert transform method are introduced in ECG waveform detection.Time domain features,morphological features and frequency domain features are introduced in ECG feature extraction.Finally,support vector machine,decision tree,linear discriminant analysis,naive bayes and knearest neighbour algorithm are introduced in the classification algorithm.The fourth chapter discusses the problem of mislabelled training samples in ECG automatic classification.To solve this problem,this paper proposes a method combining multiple machine learning algorithms based on cross validation to identify the mislabelled training samples.The proposed method is verified experimentally on the MIT-BIH arrhythmia database that has been artificially added with mislabelled noise.The results show that most of the classification performance indicators are significantly improved after removing the mislabelled training samples by the proposed method.Among them,as the overall classification accuracy for representation,if the mislabelled samples existed in the training set is less than 20%,the classification accuracy can be increased to the same level as there is no mislabelled sample in the training set with the help of the proposed method.If the proportion of mislabelled samples is 30%,the classification accuracy is slightly lower than the case of no mislabelled sample in the training set.If the proportion meets 40%,the classification accuracy is still much higher than the circumstance without filtering.In chapter 5,in view of the difficulties caused by the change of motion state in ECG identification,this paper discusses the feasibility of ECG identification with exercise by using the ECG identification database collected by our research group.This chapter first introduces the detailed information of the exercise ECG database and analyses the decisive factor of ECG identification with exercise is to extract the features that keep relatively stable before and after exercise.Then the feature extraction and selection methods,which are relatively mature in the current ECG identification field,are verified on the exercise ECG database.The experimental results show that the ECG identification methods proposed in the existing literatures have achieved good results in the static state,but the situation of the change of the motion state needs to be further studied.更多还原