基于CNN-seq2seq的运动心率检测系统

A study of heart rate measurement systembased on CNN-seq2seq under the movement condition

针对当前可穿戴式心率检测设备在运动条件下的心率测量准确度不高的问题,提出了一种卷积神经网络结合序列到序列网络(CNN-seq2seq)的深度学习算法,提取在运动状态下的光电容积脉搏波(photoplethysmograph,PPG)中的心率值的方法.结合卷积神经网络在特征提取方面的特点,并利用长短期记忆网络在时序数据处理上的优势,建立了卷积神经网络结合序列到序列的+注意力机制的网络模型.方法采集了30名身体健康的受试者在静止、行走、慢跑和快跑四个状态下的PPG信号,并通过有抗干扰能力的心电设备同步采集他们的心电(electrocardiogram,ECG)信号,将PPG信号作为神经网络输入信号,将ECG信号简化后保留心率特征,作为网络标签,然后对CNN-seq2seq网络进行训练,网络输出得到具有准确心率特征的类PPG信号,从而实现对运动条件下的心率测量.将CNN-seq2seq网络输出结果与对应的ECG信号计算每分钟心率值,心率估计的平均误差和均方误差为0.25±1.31.实验结果证明:CNN-seq2seq网络模型对于运动心率预测能得到比较理想的结果.这为实现运动心率的便携式测量提供了一种可行方案.

To address the problem of low accuracy of heart rate measurement under the movement condition of the wearable device,a deep learning algorithm based on CNN-seq2seq is proposed to extract the accurate heart rate value in photo plethysmograph(PPG).Combining the characteristics of CNN in feature extraction and taking the advantages of LSTM in time series data processing,CNN-seq2seq+Attention Mechanism network model is applied.30 healthy subjects are collected and their PPG signals at rest,walking,jogging and running states are recorded,and acquire their ECG signal synchronously through an ECG(electrocardiogram,ECG)device with strong anti-jamming ability.The PPG and the ECG signal are worked as the neural network input and input signals,respectively.These signals are used for CNN-seq2seq network training to get accurate heart rate of similar PPG signals,realizing the heart rate measurement under the movement condition.Comparing the output result of CNN-seq2seq network with the corresponding ECG signal calculate the heart rate per minute,the error of heart rate estimation is 0.25±1.31.The experimental results show that CNN-seq2seq network model can predict relatively ideal results for the extraction of heart rate in movement condition.The algorithm provides a feasible method for the practical measurement of heart rate under the movement condition.