Classification Of ECG Signal Based On Feature Processing And Capsule Network

In recent years,with the development of computer technology and artificial intelligence technology,the automatic classification method of ECG(Electrocardiogram)is becoming more and more mature.It mainly includes the following four steps: signal acquisition,signal preprocessing,signal feature processing,signal classification.Signal acquisition mainly refers to using hardware equipment to obtain the original ECG signal,signal preprocessing mainly refers to the operation of denoising the original signal and removing the baseline drift,signal feature processing refers to the operation of extracting,selecting,fusing the features of the original signal to highlight the differences between different kinds of signals to obtain better classification performance.Signal classification refers to the use of classifier to classify the signals after feature processing.The traditional feature extraction method has the disadvantages that feature extraction is not good enough,as only single dimension features are often extracted,and multiple dimension features can not be effectively combined together.The traditional feature selection method is divided into four types: Filter method,Wrapper method,Embedded method and Hybrid method.The Hybrid method is widely used because it combines the advantages of Filter method with that of Wrapper method,but there are few kinds of Hybrid method and there is still room for improvement in feature selection and feature dimensionality reduction.Furthermore,with the usage of deep learning methods,remarkable results have been achieved in the classification of ECG signals,in which Capsule Network(Capsnet),as a new neural network structure,has achieved remarkable results in the field of handwritten digital classification.This thesis mainly focuses on the two aspects of ECG classification: Multi-dimensional feature extraction and selection method and Capsnet based ECG signal classification.The main tasks of this paper are as follows:(1)A Multi-interval Symmetrized Dot Pattern(MSDP)is proposed to improve the shortcomings of the Symmetrized Dot Pattern method(SDP),such as the lack of feature extraction caused by single interval parameter,the low accuracy and the inefficiency caused by the determination of parameters by experimental selection method.By using MSDP,one-dimensional signal features can be represented by visual graphics.After MSDP transformation,one-dimensional signals can be converted into two-dimensional symmetrical petal patterns.The difference between signals can be characterized by the difference of petal shape.According to this characteristic,ECG signal can be converted by MSDP,and then the features reflected by the shape of petals can be extracted to classify different types of ECG signals.(2)A WRelief F-GA-SVM based multidimensional feature selection method is proposed to select and reduce the dimensions of the extracted original features.Firstly,an Average-Weight-Relief F method(WRelief F)is proposed to overcome the shortcomings of unbalanced sampling and unaccurate calculation of feature weights of Relief F,which is a traditional Filter method.Then by taking advantages of WRelief F for fast feature selection and GA-SVM for accurate feature selection,a new WRelief F-GA-SVM method is proposed.The classification result of Support Vector Machine(SVM)is used as the population optimization fitness of Genetic Slgorithms(GA).Finally,the "one-agingst-all" SVM classifier is used to complete the ECG signal classification.(3)A new Capsnet based ECG signal classification method is proposed.First,a new Capsnet model is proposed,which improves the problem of insufficient feature extraction in traditional Capsnet.Then the MSDP algorithm is used to transform the original ECG signals into two-dimensional images and these images are used as the input of Capsnet model.And the data enhancement method is used to enhance the images of the training set,which is used to train the Capsnet model,then the test set is used to test the Capsnet model,where the classification performance and results are obtained.Finally,the above methods are verified experimentally on the MIT-BIH open data set,Experiments show that the multi-dimensional feature extraction and classification method and the Capsnet based ECG signal classification method can obtain higher ECG signal classification accuracy that other methods.