Deep Time Series Feature Extraction And Heart Rate Estimation From BCG Signals

Ballistocardiograms(BCG)is a non-invasive cardiac monitoring technology that records the weak reaction force of the heart’s pumping blood through the body.On the premise of being able to obtain BCG signals containing rich vital sign information,its indirect contact and ease of operation make BCG technology more suitable for daily monitoring of the heart.Nowadays,more and more people are paying attention to their own heart health issues with the continuous improvement of living standards.More comfortable,convenient and accurate heart rate monitoring technology has become an increasingly urgent demand.Therefore,this thesis mainly studies the deep learning methods of heart rate estimation of BCG signals from the aspects of methods selection and features extraction based on the previous research on BCG signals.In terms of methods selection,supervised regression and classification methods can guide the construction of the model with more information,which can effectively improve the heart rate estimation performance of BCG signals.In terms of features extraction,deep learning methods can fully mine the information in the data to obtain more representative BCG signals features.In addition,the introduction of appropriate regularization items in the objective function can effectively improve the model’s ability to extract specified information in the data.The main content of this thesis are summarized as follows:(1)A BCG signals heart rate estimation method based on bidirectional Long Short Term Memory(Bi-LSTM)regression is proposed to improve the heart rate estimation performance of BCG signals.This method innovatively introduces the idea of deep regression into the heart rate estimation of BCG signals,and directly establishes an end-to-end mapping model from BCG signals to heart rate through a bidirectional LSTM network and a fully connected network to reduce intermediate steps and labor costs.In addition,we select better parameters of the model by introducing priori knowledge in the construction of BCG data samples to further improve the performance of the heart rate estimation.The experimental results show that this method can reduce the heart rate estimation error effectively and has good stability.(2)A BCG signals heart rate estimation method based on Gated Recurrent Unit(GRU)and sparse posterior probability regularization(GRU-SPP)is proposed in this thesis.By making full use of prior information,this method can improve the heart rate estimation performance of BCG signals in single-person training,and solve the problem of poor generalization performance of previous methods in multi-person training.This method converts the problems of heartbeat detection and heart rate estimation into a supervised classification problem.Firstly,the construction of BCG samples is used to solve the problem of accurate heartbeat labeling.Secondly,the timing and waveform characteristics of the heartbeat signals are fully mined by the multi-layer GRU network and fully connected network.Finally,the heartbeat classification and heart rate estimation results can be got by the hybrid features that combine temporal and spatial information.In addition,we introduce sparse posterior probability regularization item in the objective function to improve the generalization performance and speed up the training process of the model.It can be seen from the experimental results that this method not only improves the accuracy of heart rate estimation in single-person training,but also achieves better heart rate estimation results in multi-person training;In addition,this method also shows strong robustness to disturbances in the data.(3)A one-dimensional Convolutional Neural Network method(1D-CNN)for cross-modal BCG heart rate estimation is proposed to reduce the difficulty of heartbeat detection,thereby further improving the performance of heart rate estimation.This method,inspired by the idea of signal enhancement,innovatively introduces a deep regression method into heartbeat detection.Specifically,we build a cross-modal mapping model from BCG signals to heartbeat pulse signals by a deep one-dimensional convolutional neural network,which not only greatly reduces the difficulty of heartbeat detection and heart rate estimation,but also reduces the labor cost of data labeling largely.The experimental results show that it further improves the heart rate estimation performance of BCG,with the advantages of easy optimization and strong robustness.