Research On Removal Of Motion Artifact And Heart Rate Estimation Method Based On PPG

With the improvement of material living standards,more and more people are actively participating in various sports.Wearable devices are widely used to help people understand their physical condition in life and exercise.Photoplethysmography(PPG)is considered as the mainstream method of current wearable devices because of its low sampling rate and easy portability.However,PPG signals are easily contaminated by motion artifacts(MA),especially during strenuous exercise,so it is challenging to accurately estimate heart rate in real time during intensive activity of the subject.A series of denoising approaches have been proven to work well on elimination of MA in some specific activities.However,they may not be work well on different kinds of fitness activities.This thesis presented a hybrid denoising method for accurately estimating HR under different fitness activities data set,consisting of two stages.In stage 1,constrained volterra-based RLS(CVRLS)is firstly used to remove the MA of two PPG signals using the concatenation of three acceleration signals at each time instant as the reference signals.The new constrained volterra-based RLS based denoising is a nonlinear system,which is more accurate to describe the relation between MA and acceleration reference signals.Beside,a constraint parameter added to the update equation of filtering can overcome the problem that the MA is over estimated in volterra-based RLS.To Further remove the MA,two PPG signals and three acceleration signals are decomposed by the CEEMD respectively.It can solve the mode-mixing problem of EMD and achieve smaller residual noise than the EEMD.Stage 2 propose a XGboost-based speak tracking algorithm to select the spectral peak corresponding to HR,formulating the problem of spectral peak tracking into a pattern classification problem,then calculate the HR according to classification results.In order to verify the performance of the proposed algorithm,in the experimental simulation stage,the proposed algorithm is first verified on two sets of public data sets,and then further simulated on the data sets of four other different motions collected in this thesis Compared with the methods published in the recently published literature,the proposed algorithm shows better accuracy and robustness in experiments with six different fitness activities.The heart rate mean absolute error error of the six data setswas 2.04 ? 1.37 BPM(heartbeats per minute).The results show that the proposed method has a good application in wearable sensors and can accurately monitor heart rate in different types of activities.