Research On The Non-contact Vital Signs State Recognition

Vital signs detection can be divided into two types,contact detection and non-contact detection.Contact vital signs detection requires direct contact between the device and the human body,which limits its potential applications,and it also affects the life quality of the detected users.The non-contact detection can realize the detection at a distance or even penetrating the obstacles.It has the advantages of strong penetration,long detection distance,and less impact by obstacles.Therefore,it has wide application in natural disaster rescue,medical monitoring and other fields.Vital signs state change refers to changes in the respiratory rate and heart rate of a vital object under detection,such as sudden cardiovascular and cerebrovascular diseases that lead to a rapid increase in heart rate.To detect the changes in heart rate or respiratory rate（surges or drops）,in this thesis,Doppler radar and Transformer network are used for non-contact vital signs state recognition.The main work of this thesis is as follows:This thesis first introduces the state-of-the-art of non-contact vital signs detection,focusing on Doppler radar and its detection principles.According to its principle,two sinusoidal signals with frequencies of heart rate and respiration rate are used to model the movement of the heart and chest respectively.The reflected wave of the human body received by the Doppler radar is modulated by these two superimposed sinusoidal signals.Then,in order to recover the abnormal points in the radar sampled data and improve the availability of the data,this thesis proposes a data reconstruction algorithm based on the compressed sensing theory.Firstly,the Isolation Forest algorithm is employed to detect the abnormal points of the collected data and record their location,regarding the data with removed outliers as randomly down-sampled low-dimension signals.Then,a sparse orthogonal basis is designed according to the time-domain smoothness of the sampled signal and an observation matrix is also designed according to the location of the outliers,so that the missing data can be reconstructed by solving the₁l norm optimization problem.For the vital signs signals in different states collected by Doppler radar,this thesis proposes a Transformer Network method to identify the state of the detected object,to address the problem that most of the previous neural network learning methods ignore the dependence of timing data at different moments.The proposed method uses 2D convolution to divide data into blocks and computes in parallel through multi-headed attention mechanism with chi-square biased to extract its features,and finally the MUSIC（MUltiple SIgnal Classification）algorithm in spectral estimation is used to estimate the heart rate and respiratory rate of the data.In order to prove the effectiveness of the proposed methods,this thesis performed a large number of simulation experiments and hardware experiments with MATLAB and PEM11002Doppler radar.It also compares the impact of different data lengths and SNR（Signal-to-Noise Ratios）on the recognition accuracy.This method performs well in the experiments.It successfully identifies the state of the detected object,and the heart rate and respiration rate of the data can be estimated.The results show that the method in this thesis has good real-time performance and accuracy.