Research On Key Methods Of Physiological Signal Processing In Wearable Health Monitoring

In recent years,with the rapid development of wearable technology,wearable health monitoring has attracted more and more attention.Although this field has achieved great progress,there are still some new challenges in the collection,processing and analysis of physiological signals collected by wearable devices.Firstly,heart rate monitoring using wrist-type photoplethysmographic(PPG)signals during subjects' intensive exercise is challenging,since the strong motion artifacts(MA)mixed in PPG signals are difficult to remove because the frequency ranges of MA and PPG signals are highly overlapped.Secondly,in the wearable health monitoring based on compressed sensing,the high computational complexity of existing reconstruction algorithms is difficult to meet the requirements of real-time monitoring.Finally,existing heart disease automatic diagnosis methods are prone to missed diagnosis or misdiagnosis when performing real-time analysis and diagnosis using ECG segments after compressed sensing sampling.In order to solve these problems,based on the wearable remote health monitoring system developed by our project team,this dissertation focuses on the elimination of MA,the reconstruction of ECG signals,and the classification of compressed ECG,where a series of algorithms are proposed.The main contents are as follows:(1)To eliminate strong MA in the PPG signals,we propose MA removal algorithms for two different application scenarios:?For the remote server,we propose a hybrid denoising method based on Pearson correlation binary decision,which takes advantages of nonlinear adaptive filtering and singular spectrum analysis denoising.Experimental results show that the proposed algorithm can effectively eliminate MA from the PPG signals,thereby improving the accuracy of heart rate estimation.The proposed algorithm has good generalization ability,and is suitable for processing and analyzing different users' data uniformly on the server with sufficient computing resources.? For wearable smart devices(such as smartwatches),a MA removal algorithm based on conditional generative adversarial networks(CGAN)is proposed.Experimental results show that,the speed of the proposed algorithm is increased by about an order of magnitude compared with existing MA removal algorithms.Therefore,the proposed algorithm can be easily embedded in wearable smart devices with limited computing resources due to the low computational complexity of the trained denoising model.(2)For compressed sensing reconstruction,existing non-sparse ECG signal reconstruction algorithms have high computational complexity.To address this issue,we propose non-sparse ECG signal reconstruction algorithms for two different application scenarios: ? For the remote server,the alternating direction multiplier method(ADMM)is proposed to optimize the iterative process of the block sparse Bayesian learning(BSBL)framework.Experimental results show that the proposed algorithm effectively accelerates the convergence speed of the reconstruction algorithm,and improves the real-time performance of signal reconstruction while ensuring the accuracy of signal reconstruction.The proposed algorithm has good generalization ability,and is suitable for processing and analyzing different users' data uniformly on the server with sufficient computing resources.? For wearable smart devices,the traditional iterative compressed sensing reconstruction algorithm is difficult to meet its real-time requirements.So we propose an ECG signal reconstruction algorithm based on dilated residual network.Experimental results show that the signal reconstruction model trained on a specific data set has not only better reconstruction accuracy than traditional iterative compressed sensing reconstruction algorithms but also a huge improvement in real-time performance.The reconstruction speed can be improved by about 2 to 3 orders of magnitude,ensuring that the proposed model can be easily embedded in wearable smart devices with limited computing resources.(3)For the automatic diagnosis of heart disease based on compressed ECG segments,existing classification algorithms are prone to missed diagnosis or misdiagnosis.To address this issue,two automatic heart disease diagnosis algorithms are proposed as follows:? For arrhythmia classification,unlike traditional arrhythmia diagnosis algorithm that focuses on single label classification based on heartbeat,we propose a multi-label classification algorithm to classify multiple arrhythmias that may exist in a same ECG segment.Experimental results show that the proposed algorithm can accurately classify multiple arrhythmias in a same ECG segment.? For atrial fibrillation detection,the prior information of the measurement matrix is exploited to initialize and fine-tune a deep network model for improving the classification performance.Experimental results show that the algorithm can effectively improve the performance of atrial fibrillation detection based on compressed ECG,especially at higher compression ratios.