Motion Heart Rate Anomaly Detection Based on Variational Autoencoder in Multiple Wearable Device Scenarios

Abstract

Deep learning and wearable devices for heart rate detection have been widely applied in sports for real-time body monitoring. However, existing deep networks such as convolutional networks (CNNs) and recurrent neural networks (RNNs) are unable to model the spatiotemporal features of time series signals. Moreover, these models are unable to model the uncertainty in complex motion scenes. To this end, this article constructs an effective abnormal heart rate detection system based on a variant variational autoencoder. First, the photoplethysmography (PPG) signals from different user terminals are collected and transmitted to the server through the wireless sensor network. Then, on the server side, we deployed a novel variant variational autoencoder (VAE) by exploiting the 1D convolution operation and the temporal convolutional network (TCN) module for spatiotemporal feature extraction of time series. Moreover, the VAE can effectively alleviate uncertainty in motion scenes. Finally, we conducted comparative experiments on our self-built dataset of abnormal heart rate during exercise, and the experimental results showed that the proposed model achieves the highest anomaly detection performance.