Research On Algorithms Of Physiological Signal Analysis And Processing For Wearable Devices

Physiological signal analysis and processing algorithms for wearable devices have developed rapidly due to their potential applications in many fields,such as heart rate monitoring and emotion recognition.Yet,due to the characteristics of physiological signals from wearable devices,there still exist some issues.On one hand,physiological signals,such as photoplethysmography(PPG)and electrocardiograph(ECG),are easily contaminated by motion artifacts(MA)due to human motions,which may result in the unreliability of physiological signals(PPG and ECG)obtained.On the other hand,there is individual signal difference for physiological signals,including a signal difference(intersubject signal difference)between subjects and a signal difference(intra-subject signal difference)within a same subject,which may lead to a poor generalization of an emotion recognition model.To address these issues,this dissertation proposes some physiological signal analysis and processing algorithms for wearable devices.Specifically,the main contents of this dissertation are as follows:(1)A MA removal algorithm based on sparse representation using PPG signals is proposed,used to overcome different strong MA(in PPG signals)caused by different types of quasi-periodic motions(quasi-regular motions,such as running and elliptical machine training).The proposed algorithm can learn the characteristics of different types of strong MA via dictionary learning and sparse representation,which can effectively remove different types of strong MA caused by various quasi-periodic motions.Based on six physical exercise datasets(including six typical quasi-periodic motions),experiments on heart rate monitoring using the proposed algorithm show that the algorithm can effectively remove different strong MA caused by quasi-periodic motions from PPG signals,and show that our heart rate monitoring method using the algorithm has lower error compared with some state of the art methods.(2)A MA removal algorithm based on Volterra Recursive Least Square(VRLS)and Variational Mode Decomposition(VMD)using PPG signals is proposed,used to remove strong MA(in PPG signals)caused by irregular motions(such as jumping and boxing).The proposed algorithm exploits random forest-based binary decision to combine VRLS and VMD,which can effectively remove strong MA caused by irregular motions while reducing the computational complexity.The algorithm does not rely on motion state(that is,it does not require that motions are regular),which is suitable for irregular motions.Based on a public dataset(including various irregular motions),experiments on heart rate monitoring using the proposed algorithm show that the algorithm can effectively remove strong MA caused by irregular motions from PPG signals,and show that our heart rate monitoring method based on the algorithm achieves better performance compared with some other methods.(3)A MA removal algorithm based on Singular Spectrum Analysis(SSA)and feature fusion using ECG is proposed,used to eliminate MA(in ECG signals)caused by electrode stretching due to human motions.The proposed algorithm first eliminates MA by SSA.Then,it extracts both statistical features(based on prior information)and depth features(based on hidden information obtained by deep learning),and integrates the two kinds of features,improving the robustness of extracted features to MA.Based on two datasets(including ECG with MA),experiments on emotion recognition using the proposed algorithm show that the algorithm can effectively remove MA in ECG signals,and show that our emotion recognition method based on the algorithm has better performance compared with some other methods.(4)An individual signal difference elimination algorithm based on balanced distribution adaptation(BDA)using ECG is proposed,used to overcome individual signal difference in ECG signals.The proposed algorithm not only can reduce inter-subject signal difference via BDA,but also can reduce intra-subject signal difference by online data adaptation(ODA),ensuring the online emotion recognition performance.Based on two public datasets,experiments on emotion recognition using the proposed algorithm verify the effectiveness of the proposed algorithm on eliminating individual signal difference in ECG signals,and show that our emotion recognition method based on the algorithm has better performance compared with some other methods.