Study On Compressive Sensing And Arrhythmia Classification Of Electrocardiosignal In Wearable System

With the development of society and the improvement of living standards,cardiovascular disease has become one of the major chronic diseases that threat people’s lives,arrhythmia is a common cardiovascular disease,and ECG signal is the key indicator of detection and diagnosis of arrhythmia and other heart disease.The wearable health monitoring system,as one of the most effective and practical new medical care way,enables continuous real-time monitoring of ECG signals and automatic classification of arrhythmia.Nevertheless,a large number of ECG data will be a huge burden on the storage and transmission of resource-constrained wearable systems,and compressive sensing is a good solution to the problem as a signal compression techniques.However,there are still many problems existing in the current methods for compressive sensing and automatic classification of ECG signal,and it starves for improvement and innovation.Therefore,it is of great significance to study the compressive sensing and arrhythmia classification of ECG signals in wearable health monitoring system.Based on the characteristics of limited resources,this thesis carries on the study in four main aspects,i.e.preproeessing,characteristic waveform detection,compression and reconstruction,feature extraction and arrhythmia classification of ECG signals.(1)Adaptive ECG preprocessing algorithm based on extended recursive least squares method was studied.The ECG preprocessing aims to eliminate the various noise interference in the ECG signal,as far as possible,to preserve the important characteristics of the original signal.This thesis presented an adaptive denoising method based on extended recursive least squares.This method can better remove the power frequency interference,baseline drift and electromyogrphy interference,and provide clean ECG signal for the follow-up ECG signal processing.It also has a low computational complexity.(2)QRS complex detection algorithm based on Pan-Tompkins was studied.QRS complex detection is the basis of automatic classification of ECG signals.This thesis proposed a QRS detecting method for ECG signals based on the Pan-Tompkins algorithm.The algorithm not only has a higher detection accuracy,but also has a faster detection speed,so that the accuracy and realism of the unity.(3)ECG compression and reconstruction algorithm based on compressive sensing was studied.The sparse representation of the signal is the core of the ECG signal compressive sensing.This process requires not only sparse representation of the original ECG signal,but also the time it takes.In order to reduce the time of dictionary learning,this thesis presented a "divide-merge" dictionary learning method,which utilizes internal clustering structure of ECG signal to train.The algorithm has a low time complexity,and can construct a dictionary,which more in line with the characteristics of ECG signal and better sparse represent it.In addition,for the practical application of multi-channel ECG signal,this thesis also proposed a multichannel ECG signal compressive sensing method based on spatio-temporal sparse model.It realizes the compression and reconstruction of multichannel ECG signal by exploiting both the temporal correlation of each channel signal and the spatial correlation among different channel signals.Under high signal reconstruction quality condition,this algorithm achieved superiority in compression ratio with low time complexity.(4)ECG feature extraction algorithm based on improved principal component analysis was studied.ECG feature extraction is the key to automatic classification of ECG signals,and the results directly affect the accuracy of arrhythmia classification.This thesis improved the principal component analysis method,which normalizes the difference between the mean values of the principal components of different categories by the variance of the special principal components,and selects some top principal components together with the QRS width and RR interval to compose the eigenvectors of the ECG signals according to the standardized results.This algorithm can effectively extract the characteristics of ECG signal.(5)Arrhythmia classification algorithm based on least squares support vector machine semi-supervised learning was studied.This thesis proposed a semisupervised learning classification method to enhance the classification effect by adding unlabeled sample training.On the basis of this,a three-layer multi-classification model was proposed and applied to the automatic classification of ECG signals of normal heart beat,right bundle branch block,atrial premature beats and ventricular premature beats.Furthermore,a direct feature extraction and automatic classification method for ECG signal is presented.It performs feature extraction and classification task directly from the compressed ECG data in the compressive domain,skipping the reconstruction step.The quantity of ECG data is greatly reduced and the efficiency of automatic classification is effectively improved under the premise that the classification performance of this algorithm is basically unchanged or very little declined.It also laid a good foundation for the future ECG signal processing research work in the compressed domain.(6)This thesis both used ECG data from foreign standard ECG database and the measured ECG data of nearly 1,000 patients in a domestic hospital to carry out experimental research on preprocessing,QRS detection,signal compression and reconstruction,ECG feature extraction,and automatic classification for N,A,V,and R ECG signals.It also analyzed the experimental results in depth,so that the effectiveness and practicality of the method were verified,and the robustness of the method was evaluated.