HUMAN BIOMETRIC SIGNALS MONITORING BASED ON WIFI CHANNEL STATE INFORMATION USING DEEP LEARNING

Abstract

As civilisation has progressed, the public’s attention to health monitoring has increased. Health monitoring equipment has become ubiquitous. Most of the health monitoring instruments mainly detect human biometric signals, such as heart rate and respiratory rate. These indicators can reflect human primary health conditions, both mental and physical and classify sleep stages. The classification of sleep stages, as well as the monitoring of breathing rate and heart rate, can aid in the evaluation of health and the diagnosis of disorders by doctors. Some health problems, such as depression, insomnia, obesity and other diseases, may benefit from it.

A large number of methods have been proposed to detect breathing and heartbeat, for example, by using wearable devices and non-contact devices. Traditional methods (e.g. polysomnography) are mainly used in clinical treatment, which requires patients to stay in hospital and wear a lot of sensors to monitor sleep. Most current monitoring solutions for heart rate and breathing rate use wearable devices attached to the human body. This thesis presents a WiFi-based method combined with convolutional neural networks (CNNs) to predict heart rate and respiratory rate.

We first proposed a single input and two outputs of the CNNs to simultaneously estimate both heart rate and breathing rate. This network utilises the inner relationship between heart rate and respiration rate of the channel state information (CSI). Therefore, heart rate and respiration rate should be jointly estimated because they are related to each other. We leverage the amplitude and phase information of CSI collected by a pair of WiFi devices as an input for the neural network. We then train the network to predict the heart rate and breathing rate without the traditional complex feature-selection algorithms. For algorithms using sub-carrier selection strategy, the network may not be able to achieve its optimal performance, as it uses CSI partially. Therefore, we utilise the CNN model to improve accuracy and reduce computational complexity. Our WiFi-based approach solves the problems of privacy issues and the environmental factors. Using the generalisation ability of the CNN, one can easily adapt to different environments and avoid the unreliability of the analysis method due to environmental changes. For this system, in the real environment, the estimation error of breathing rate is 0.2 beats per minute, and heart rate is 0.6042 beats per minute. The overall accuracy of this system can achieve 99.109% and 98.581%, respectively.

In addition, we design and compare two neural networks based on deep learning for categorising four types of sleep stages, including wake, rapid eye movement (REM) sleep, non-rapid eye movement (NREM) light sleep and NREM deep sleep. The first neural network uses the data calculated by the cardiopulmonary coupling (CPC) algorithm as an input. For the second network, the input data is the CSI matrix without further processing. By comparing two neural-network classification approaches, we find that the WiFi sleep-stage neural network performs better because the CSI is influenced not just by large-scale movement (body movement) but also by tiny-scale movement (chest movement). Human body-movement data, respiration data and heart rate during sleep were already included in the CSI data. This also explains why we do not use the motion-sensing module in our system. We directly use the CSI matrix without further processing as input for classification. The overall accuracy of the method based on WiFi can achieve 95.925%. The accuracy for using the CPC algorithm is 90.15%.