Research On BCG Signal Processing And Disease Diagnosis Of Contactless Health Monitoring System

The blood pumped by the beating of the heart will generate an impact force in the blood vessels,and the weak vibration signal caused by this force spreading to the body surface is the ballistocardiogram(BCG).Using flexible sensors,BCG signals can be collected without contact,realizing non-invasive and long-term monitoring of physical health status.It has great application potential in home care,telemedicine and auxiliary disease diagnosis.However,due to the high sensitivity of the sensor that collects the BCG signal,the collected BCG signal is susceptible to noise interference.The internal causes of interference include human organs,muscles or unconscious weak movements,and general human movements,while external causes include environmental interference,equipment noise,etc.These interferences negatively affect the shape of the BGC signal,making it difficult to accurately identify the physiological information contained in the signal itself,leading to insufficient estimation accuracy of physiological information such as heart rate and respiration rate.At the same time,the boundaries between the heartbeats of BCG itself are not very clear,and there are too many interference factors for BCG,which reduces the accuracy of disease diagnosis based on BCG.Therefore,this dissertation studies the problem of noise suppression of BCG signals to obtain better quality heartbeat signals,and then detects the health status of the subjects.By deeply mining the diverse characteristics of BCG,BCG-based disease identification and diagnosis are realized.In this dissertation,a mattress-based BGC health monitoring system is designed based on piezoelectric film sensors,and a BCG signal processing framework is constructed to recover the heartbeat signal from the BCG signal with low SNR.In view of the large-scale interference such as respiration and general motion in BCG,as well as some weak motion or friction causing local damage or loss of the signal,variational modal decomposition is used to decompose the BCG signal into multiple modal components in this dissertation and based on the difference of frequency band of heartbeat and respiration,the components associated with respiration are selected to reconstruct and then separate the respiration.Then,a log-likelihood ratio discriminant model is built to distinguish motion and calm signals based on their difference in probability density distribution.Finally,a denoising method based on Multichannel Singular Spectrum Analysis(MSSA)is designed to recover the heartbeat signal,which can reduce weak motion or friction interference in BCG.In the MSSA,the one-dimensional BCG is reshaped into a two-dimensional signal matrix,so that each column of the matrix contains the complete waveform of the hearbeat signal.Then the multichannel singular spectrum decomposition is performed on the matrix to obtain multiple signal components and then achieves denoised signals by reducing the rank.By taking the number of BCG lateral textures as the rank,the proposed MSSA based denoising method can reconstruct the signal matrix.The proposed method takes full use of the characteristic of the similar structures in two-dimensional reshaped matrix to solve the problem of reconstruction component selection at low SNR The results of BCG signal recovery on the simulated and measured data show that the denoising method proposed in this dissertation can better restore the signal structure,and thus improve the SNR of the BCG signal.Especially,the severely distorted part of the BCG signal is properly restored.In order to realize the assisted diagnosis of diseases at home,an improved Res2Net(IRes2Net)was proposed,which uses BCG signal to detect the health status of subjects.IRes2 Net uses the multi-scale convolution to extract the long-term global features and the detailed features between wavelets of the signal.At the same time,channel attention module is embedded to aggregate and learn the features of each channel and redistribute the importance weights for each channels,so that the network can fully mine the physiological features of BCG.By fusing the BCG signal features extracted by multi-scale convolution and channel attention module,IRes2 Net solves the problem of inconspicuous disease characteristics caused by the unclear BCG signal boundaries,and thus improves the identification accuracy of five disease states such as tachycardia,hypertension.Cross-entropy is the loss function,and the SGD algorithm is used to update gradient,so that the result of network classification could approximate to the real sample distribution.The experiments on BCG signal from the open dataset for five disease diagnosis show that IRes2 Net makes full use of the multi-scale features of the BCG signals,and achieves high classification accuracy.