Study On Compression,reconstruction And Classification Of ECG Signals In Wearable ECG Devices

It has become a development trend to use wearable devices as a carrier to collect various physiological signals and physiological indicators of the human body,and to monitor and evaluate the health status of the human body based on corresponding algorithms.In order to improve the endurance of wearable devices,the collected signals need to be compressed in wearable ECG devices,and the compressed ECG signals need to be reconstructed at the receiving end.At the same time,in order to timely treat and intervene the sick people,the reconstructed ECG signals need to be detected and diagnosed.Based on the above considerations,this paper studies the compressed sensing reconstruction method and arrhythmia classification method of wearable ECG devices.The main research contents are as follows:Firstly,research on non-negative constraint dictionary learning algorithm based on CS.In the process of ECG reconstruction,the selection of sparse base has great influence on the reconstruction accuracy.In this dissertation,a non-negative constraint dictionary learning algorithm is proposed to construct an overcomplete dictionary that is positively correlated with the original signal as a sparse basis for signal reconstruction.Under the ₀ or ₁norm constraints,dictionary atoms that are negatively correlated with the original signal are inevitably generated,therefore resulting the suboptimal dictionary with atoms to"cancel each other"by addition and subtraction to approximate training samples.In this paper,by adding non-negative constraints in the coding stage,the negative coding coefficients are removed.During the dictionary update process,the dictionary atoms are updated by the block coordinate descent method,and finally a dictionary that is positively correlated with the original signal is obtained.At the same time,in order to ensure the sparsity of the coding coefficients,we add a penalty term into the objective function to remove small coding coefficients and achieve the result of sparse coding.Experimental results show that the proposed algorithm can obtain higher reconstruction accuracy under the same compression rate compared to other similar algorithms.Secondly,research on the automatic classification algorithm of arrhythmia based on the label consistent non-negative representation（LCNR）.Ecg signal waveform has great similarity in geometric structure and amplitude,etc.With the increase of ECG signal data,the difference between classes becomes smaller,which will lead to the reduction of the generalization ability of the classification model and adversely affect the classification results of ECG signal.In this dissertation,the non-negative representation classification algorithm is improved,and the"discriminant sparse coding error term"and"classification error term"are integrated into the objective function,so that the coding coefficient can maximize the inter-class difference and minimize the intra-class difference.At the same time,an over-complete dictionary is obtained through the dictionary learning algorithm,which solves the problem that the coding process of the non-negative representation classification algorithm is very slow when the training sample size is large.The algorithm can be divided into three stages.In the coding stage,alternate direction multiplier method is used to solve sparse solutions under non-negative constraints.In the dictionary learning stage,block coordinate descent algorithm is used to simultaneously learn a compact discriminant dictionary and a multi-class linear classifier.In the final classification stage,the non-negative representation model is used to solve sparse coefficients and the multi-class linear classifier is used for classification.Since the coding coefficient has the property of maximizing the difference between classes,a simple linear classifier can obtain a higher classification accuracy in the classification stage.The experimental results show that the label consistent non-negative representation classification algorithm reaches the classification accuracy of 99.83%in the scheme proposed by the American Medical Instrument Promotion Association,and 99.28%in the scheme based on category.The sparse decomposition efficiency can be improved by learning a discriminant dictionary,and the operation efficiency is increased by 5 times compared with the non-negative representation classification algorithm.Thirdly,research on automatic arrhythmia diagnosis algorithm based on discriminant convolution sparse coding（DCSC）.Ecg abnormalities are manifested as local abnormalities in electrocardiogram.How to extract these local abnormal features is very important to improve the classification accuracy.The convolutional sparse coding algorithm is trained to obtain a set of convolution dictionaries that can match the inherent structure of ECG signals and can be used to extract local features of ECG signals.Therefore,a discriminant convolutional sparse coding model is proposed to extract local features of ECG signals through convolution operation and the sparse coefficient is taken as the eigenvalue.By adding"discriminant sparse coding error term",the trained convolution dictionary can be discriminant to improve the discrimination of sparse coefficients.Finally,the classification task is realized by linear SVM classifier.In order to reduce the computational cost,the DCSC model is converted to the frequency domain,and the ADMM method is used to optimize the DCSC model.At the same time,in order to reduce the dimension of eigenvalues,we use the maximum pooling operation for the sparsity coefficient.Finally,these pooled coefficients are used as features and fed to linear SVM classifier for ECG classification task.The experimental results show that the convolution dictionary is superior to the over-complete dictionary in feature extraction.At the same time,in the training stage,the convolution dictionary is less affected by the number of trained samples,so the classification accuracy can be higher when the number of samples is small.