Research And Implementation Of Vital Signs Monitoring And Cardiac Motion Modeling Based On Few-shot Learning And Radar

The acquisition of information pertaining to vital signs and cardiac activity holds paramount significance within the domains of intelligent healthcare and personal health management.Radar-based solutions,characterized by their non-invasive nature,excel in the retrieval of vital sign and cardiac motion-related data.They additionally offer advantages such as safeguarding user privacy and resilience to variable lighting conditions.This thesis is aimed at a comprehensive exploration of vital sign and cardiac motion-related information latent within radar signals.In scenarios where radar data samples are limited,the study endeavors to develop and implement a radar signal-based framework for vital sign monitoring and cardiac motion information modeling.This thesis proposes a vital sign monitoring scheme founded on autocorrelation and variational mode decomposition.It successfully addresses the intricate challenge posed by high-order harmonics within respiratory signals,which can severely disrupt the accurate extraction of heart rate signals.The proposed scheme enables the extraction of vital signrelated information from radar signals.It predominantly employs an algorithm rooted in autocorrelation functions to estimate the respiratory rate from radar echo signals emitted by the subject under examination.Additionally,it employs an adaptive notch filter to eliminate frequency components corresponding to high-order harmonics within the respiratory signal.Subsequently,the variational mode decomposition algorithm is employed to estimate the heart rate from radar echo signals of the subject.The thesis conducts meticulous experiments to validate the efficacy of the proposed vital sign monitoring framework,and it conducts a comparative analysis against pertinent research in the field.The comparative experimental results demonstrate that the proposed vital sign monitoring framework presented herein surpasses other algorithmic approaches,achieving an average heart rate accuracy of 89.7%which is an average increase of 5.4%compared to multiple comparative experiments.And achieving an average absolute error of 1.7 breaths per minute in respiratory rate extraction,which is an average reduction of 1.4 breaths per minute compared to multiple comparative experiments.Moreover,this thesis introduces a composite neural network integrating convolutional neural network and bidirectional long short-term memory network.This integrated approach aims to resolve the intricate challenge of extracting cardiac motion-related information from radar signals,a task traditionally deemed arduous.The thesis primarily leverages a 3D micro-Doppler focusing algorithm to process radar echo signals collected from human subjects.It introduces a data augmentation-driven solution for few-shot learning and proposes a composite neural network combining convolutional neural network and bidirectional long short-term memory network.This composite neural network is designed to effectuate a cross-domain nonlinear mapping,transitioning from radar echo signals to electrocardiogram signals.In so doing,it establishes a unified linkage between the mechanical dynamics of cardiac motion and the electrical signal conduction facets thereof.Experimental validation showcases the remarkable superiority of this solution,with an average correlation coefficient of 0.853 between the presented electrocardiogram signals,reflecting a noteworthy 0.071 enhancement compared to the comparative experiment,while maintaining consistently outstanding temporal accuracy in electrocardiogram waveform-associated events.