Research Of Deep Feature Extraction Algorithm For ECG Identification

In recent years,with the rapid development of Internet and information technology,personal identification system is indispensable in daily convenient living environment.Using the inherent biometrics of human body to realize identity recognition has higher reliability,so it has been widely studied by researchers.ECG signal is the process of recording the electrical activity of cardiac pacing,which contains rich personal information and meets the basic characteristics of identity recognition.Compared with existing biometrics,the main advantages of using ECG(Electrocardiograph)for identity recognition are as follows: ECG signal is measured in living body and difficult to forge;The signal will not be lost under the external influence and exists all one’s life;ECG signal is one-dimensional signal with small capacity;The signal acquisition is convenient and the cost is low.Therefore,the individual classification model based on ECG signal provides new thought for the research of secure and reliable identification.The common process of ECG identity recognition includes three steps: ECG signal preprocessing,feature extraction and processing,feature classification and recognition.Among them,the algorithm of feature extraction and processing is very important for the recognition results.The traditional feature extraction algorithm has some problems that are relying on the visual features of signals deeply,weak robustness of model for the changing data,insufficient generalization performance of model for multi-source data and so on.The specific contents are as follows:(1)The low-level feature extraction algorithm can not mine deep information of signal,which makes it difficult to improve the classification accuracy;(2)When people are in the state of movement or emotional fluctuation,the classification effect of model is easy to be affected;(3)When the signal sources come from different acquisition conditions,the generalization ability of model is insufficient.Aiming at the above problems in the existed identification models,based on the physiological characteristics of ECG signals,the related deep learning algorithms are utilized to mine and automatically learn features.This paper studies the deep feature extraction algorithm for ECG identification and the main contents are as follows:1.Long Short-Term Memory ECG identification algorithm based on scale features is proposed to realize the deep feature extraction of low dimensional scale signal representation.Firstly,GP(Gaussian Pyramid)technology is used to perform down sampling on the fixed intercepting ECG signal segments to realize the multi-scale analysis of signal.With the higher pyramid level changing,the signal size get small and so as the resolution.The signal representations in compatible scale spaces are selected and input into the stacked two-layer long-short-term memory neural network.In view of that LSTM network model has memory unit and it is good at processing long-term ECG signals,with the progress of deep learning,the nonlinear expression ability of deep features gradually increases.Finally,the Softmax layer of network is used as the classification layer to identify the individual identity.The proposed model which is applied in the databases of healthy people and individuals with arrhythmia diseases respectively shows the features obtained by neural network can accurately identify different types of individuals,and the feature classification accuracy of the model is significantly higher than that of the underlying morphological features.2.Fragment Pooling Stacked Autoencoder ECG identification algorithm based on different heart rate is proposed to solve the limitations of fixed feature interception methods on heart rate variability.During feature processing,the change problem of heart rate is existing.So traditional signal segmentation of choosing forward and behind R wave,which makes the total amounts of feature segments information are inconsistent and the recognition rate is difficult to improve.In order to eliminate the influence of heart rate variability caused by individual movement on signal recognition,the dynamically segmented beat data is combined with the spatial pyramid pool layer of the segment to obtain the FSPP(Fragment Spatial Pyramid Pooling)feature input autoencoder network with same dimension.Two autoencoders after cascade training are used as the hidden layer and Softmax classification output layer to form a complete depth recognition network.Error reverse transmission is used to fine tune to achieve accurate individual classification.By analyzing the simulation results of the database collected after motion it can be seen that the deep features with better discrimination will be obtained through deep learning FSPP features,and its performance is better than the classification model of shallow machine learning,which enhances the robustness of identification algorithm.3.The ECG identification model based on two-level fusion feature extraction algorithm is proposed to realize the efficient identification of mixed data sources in different sampling frequencies.Firstly,the features of Hilbert transform and power spectrum are extracted from the segmented heart beat data and combined into a set to obtain the primary fusion features.Secondly,PCANet(Principal Component Analysis Network)structure is as input to extract the deep features of signal.Considering that two-layer features of PCANet describe the primary features from different angles and the feature dimension extracted by PCANet is high.So MF(Max Fusion)algorithm is proposed to fuse and compress the two-layer automatic learning features.Finally,the linear kernel support vector machine is used to obtain the label of single feature classification to complete individual identification.The primary fusion feature of first level can overcome the impact of different sampling frequencies on recognition of mixed data sets.The PCANet fusion feature of second level can not only obtain the deep features with individual difference information,but also improve the efficiency of classification and recognition.Integrating two-level fusion feature algorithm which is realized the complementary feature information of different levels and the higher recognition accuracy is obtained in single data set and mixed data set that verifies the generalization performance of the model.