Radio-frequency (RF) sensing for deep awareness of human physical status

• Main Author: Mohammad Al-Muhazerin Mohammad Yatim
• Other Authors: Luo Jun
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2022
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
• Online Access: https://hdl.handle.net/10356/156820
• Institution: Nanyang Technological University

Respiration rate is defined as the number of breaths taken in a minute. It is one of the vital signs used in assessing a person’s state of health. It can affect the heart rate, blood pressure, and oxygen saturation and it varies among individuals and age groups. Contact type sensors are primarily used in the medical sector to determine the respiration rate. However, these sensors must be thoroughly sterilized before they can be used on another patient to reduce the spread of disease. The alternative is to use a contactless sensor, such as an Ultra-wideband (UWB) radar. It uses radio frequency to detect movements in front of the radar. However, due to the typical radio frequency challenges such as multipath fading and interference, using UWB radar to determine respiration rate is still experimental. A solution to this is to leverage on neural network. In recent years, there is an emerging trend to use a neural network to solve difficult problems. In this project, two models using 1-dimensional convolution neural networks were proposed. The two models proposed are Lit Review Model and Lit Review with Pooling Model. The results of both models are compared. The purpose of comparing the results from both models is to evaluate the scientific claim that adding a pooling layer degrades the accuracy of the prediction. Additionally, the results of both models are combined to create a new result. The purpose of this is to evaluate whether combining the results from the best of both models will improve the prediction. This dual model is called the Ensemble Model. The scientific claim holds when the models were compared against their respective batch sizes. However, it was seen that it is not necessarily the case that adding a pooling layer produces the best neural network. The best Lit Review with Pooling Model was slightly better than the best Lit Review Model. The Ensemble Model shows an improvement of 1.79% in training Root Mean Square Error (RMSE) and 7.25% in validation RMSE when compared to the Lit Review with Pooling Model. On the other hand, when compared to the Lit Review Model, the Ensemble Model performed 29.85% better in training RMSE and 6.04% better in validation RMSE.