Research On ECG Feature Extraction And Human Identification Algorithm

The technology of extracting ECG signal has a history of one hundred years,and has been very mature.It can be extracted directly through the thumb electrode,which is very convenient to operate.As a biological feature,ECG also has good stability,universality and uniqueness,and belongs to human internal characteristics,which is not easy to be forged and has high security.In recent years,the technology of using ECG signal for identity recognition has been developed,and many technical breakthroughs have been made.However,the existing ECG feature extraction methods often extract a single category of features in time domain or frequency domain,and lack a feature extraction method that makes rational use of the combination of time and frequency domain.Aiming at such problems,this thesis integrates time-domain features and frequency-domain features to make them complement each other,and proposes a new feature extraction method of ECG signal.This thesis uses ECG data from MIT-BIH,ECG-ID and PTB databases.They do not limit the state of the collector when collecting signals,which is in line with the actual application scenario.The sampling rates of different database signals are different to verify the universality and robustness of the algorithm in this thesis.According to the frequency distribution characteristics of ECG signal,this thesis uses 4-order Butterworth filter and wavelet decomposition two-step denoising method to obtain pure ECG signal.The pan Tompkins algorithm is used to detect the R-wave peak in the ECG signal.Based on this,the ECG signal is divided into multiple beats,and the QRS complex with the most concentrated energy is intercepted.Each QRS complex is subjected to Fourier transform after framing and windowing to obtain their power spectrum.According to the frequency characteristics of ECG signal,Mel filter bank is built to extract the characteristics of low-frequency part.Finally,the characteristics of frequency domain cepstrum coefficient are obtained by calculating logarithmic energy and discrete cosine transform.Through the comparative experiments of different classifiers,it is found that the feature classification using SVM can obtain the best recognition performance.In order to further improve the accuracy of identity recognition,based on the above research,using deep learning technology,this paper proposes an identity recognition algorithm combining time-frequency analysis,which integrates the characteristics of time domain and frequency domain.In the time domain part,a multi-scale time-domain feature extraction network for processing one-dimensional time-domain signals is designed by referring to the idea of hole convolution in time-delay neural network and two-dimensional convolution neural network;In the frequency domain,the cepstrum coefficients are input into BP neural network to obtain the abstract frequency domain features;After the two parts of features are spliced,the final softmax full connection layer is input for classification and recognition.This method takes into account various characteristics of ECG in time and frequency domain.The experimental results show that this method has obvious advantages in recognition performance compared with other ECG identification methods.