Research And System Implementation Of Identification Algorithm Based On PPG Signal

Nowadays,with the gradual improvement of social informatization,the problems of personal information security and social security are becoming more and more prominent,which puts forward higher requirements for identity recognition technology.The identification technology based on photoplethysmography(PPG)has become a research hotspot in the field of identification because of its advantages of safe and sustainable identification.However,it is still difficult to extract stable biometrics from PPG signals.In this thesis,the research of this technology is divided into three layers from bottom to top: signal acquisition and processing,feature extraction and representation and security application.According to the problems existing in each layer,corresponding algorithm is proposed,and a PPG signal identification system under the background of medical care service is displayed.The main work of this thesis is as follows:1.In the signal acquisition and processing layer,an adaptive method using dynamic time warping and statistical strategy(SQA-DNN)is proposed to to evaluate the signal quality,so as to solve the problem that the accuracy of the model is reduced due to the interference of PPG signal collected by wearable devices.Combined with deep neural network,end-to-end identity authentication is realized.The experimental results in three different noise scenarios show that: in a noisy environment,compared with the existing methods,SQA-DNN improves the overall accuracy of the authentication model by up to4%,and the individual accuracy by up to 12%.2.In the feature extraction and representation layer,a method based on unsupervised domain adaptation(UDA-DNN)is proposed to align the features in different periods,so as to solve the problem that the model performance is degraded due to the distribution difference of user PPG signal biometrics in cross-period scenes.Combined with deep learning scheme,more stable domain invariant features are extracted in multiple periods.The results of identity authentication experiments on cross-period datasets show that: UDADNN alleviates the problem of performance degradation of the authentication model in cross-period scenarios,and the decline is reduced by 2.5% compared to the non-domain adaptation method.3.In the security application layer,a semi supervised method combining deep learning,clustering algorithm and distance similarity(CBTM-DNN)is proposed to solve the problem that the trained recognition model in the downstream task of identity recognition can only identify known individuals based on prior knowledge,and cannot discover and identify unknown individuals without retraining.The method can accurately identify known users,efficiently discover and identify unknown users without retraining the model.The results of identity recognition experiments on public datasets show that CBTM-DNN has a recognition accuracy of about 99% for known users,and has a accuracy of above 93% and 86% respectively for discovering and re-identifying unknown users.In addition,this thesis introduces a remote physiological signal monitoring platform based on PPG signal identification under the background of intelligent medical care and the project supported by the Sichuan Provincial People's Hospital,which can provide information-based intelligent medical care services for the elderly.