Human Identification System Research Using Electrocardiograms Based On Neural Network

Human identification become an essential problem for people's more and more requirements on the reliability of information security. The traditional methods using documents, keys, and passwords to identity have the defects to be easily copied, stolen, forgotten or lost, which is far from satisfying the requirements of the people. Therefore, more reliable biometrics technology emerged. Electrocardiogram (ECG), because of its unique advantages, is used as biometric identification, which is gaining attention of researchers home and abroad. The mechanism of ECG features is complex and not easily imitated. And it has a high safety factor, so ECG is more suitable for medical and insurance fields than other methods. For the current biometric recognition technology (signature, voice or other behavioral characteristics, fingerprints, retina, iris, face, DNA et.al.), ECG identification is an effective complement, and which can be integrated with other biological characteristics to form multi-modal identification systems so as to further improve the reliability of biometric identification.Based on existing researches at home and abroad, an intensive study of ECG signal preprocessing, feature extraction, feature analysis and feature weight analysis was made in the paper. A neural network classifier was designed to realize the ECG identification and it was optimized by GA algorithm and DNA algorithm. The main research was concluded as follows:(1) Preprocessing and feature extraction of ECG signal. We have analyzed the frequency of ECG signal and the noise included by using wavelet and wavelet threshold methods filter the low and high frequency noise in ECG signal. According to the singularity analysis of ECG by wavelet detection principle, the ECG signal is decomposed by quadratic spline wavelet atrous algorithm. First, according to the correspondence between R wave and the maxima of wavelet decomposition, we extract the position of the peak point. Then, by using linear fitting minimum error method, we determine the starting point and the end point of QRS wave. And then we take the peak point of R wave as the reference point, set the search window to extract the peak points of P wave and T wave using the same strategy and again use the linear fitting minimum error method to extract the start and end points.(2) Weight analysis of ECG feature and selection of the optimal feature subset. The presence of redundant features exist in ECG signal will increase the computational and decrease the accuracy of identification. Therefore, linear discrimination analysis is used to take weight analysis to the ECG features extracted so as to determine the contribution of each feature, and then sort the features according to the weight. Features are selected according to the principle of forward sequential. Evaluated by the accuracy rate of BP neural network classification, the optimal characteristics for identification subset is determined then.(3) Design and optimization the neural network classifier. A BP neural network with three layers is used as the ECG identification classifier. Take the optimal feature subset as the input of the classifier, the output layer is up to the number of samples. As the BP neural network has the problems of easily falling into local minimum and being not convergence, GA and DNA algorithm are used to optimize it. Finally, clinical ECG data is used to test the performance of the classifier designed.