Remote Human Vital Sign Monitoring Using Multiple-Input Multiple-Output Radar at Millimeter-Wave Frequencies

Non-contact respiration rate (RR) and heart rate (HR) monitoring using millimeter-wave (mmWave) radars has gained lots of attention for medical, civilian, and military applications. These mmWave radars are small, light, and portable which can be deployed to various places. To increase the accuracy of RR and HR detection, distributed multi-input multi-output (MIMO) radar can be used to acquire non-redundant information of vital sign signals from different perspectives because each MIMO channel has different fields of view with respect to the subject under test (SUT). This dissertation investigates the use of a Frequency Modulated Continuous Wave (FMCW) radar operating at 77-81 GHz for this application. Vital sign signal is first reconstructed with Arctangent Demodulation (AD) method using phase change’s information collected by the radar due to chest wall displacement from respiration and heartbeat activities. Since the heartbeat signals can be corrupted and concealed by the third/fourth harmonics of the respiratory signals as well as random body motion (RBM) from the SUT, we have developed an automatic Heartbeat Template (HBT) extraction method based on Constellation Diagrams of the received signals. The extraction method will automatically spot and extract signals’ portions that carry good amount of heartbeat signals which are not corrupted by the RBM. The extracted HBT is then used as an adapted wavelet for Continuous Wavelet Transform (CWT) to reduce interferences from respiratory harmonics and RBM, as well as magnify the heartbeat signals. As the nature of RBM is unpredictable, the extracted HBT may not completely cancel the interferences from RBM. Therefore, to provide better HR detection’s accuracy, we have also developed a spectral-based HR selection method to gather frequency spectra of heartbeat signals from different MIMO channels. Based on this gathered spectral information, we can determine an accurate HR even if the heartbeat signals are significantly concealed by the RBM. To further improve the detection’s accuracy of RR and HR, two deep learning (DL) frameworks are also investigated. First, a Convolutional Neural Network (CNN) has been proposed to optimally select clean MIMO channels and eliminate MIMO channels with low SNR of heartbeat signals. After that, a Multi-layer Perceptron (MLP) neural network (NN) is utilized to reconstruct the heartbeat signals that will be used to assess and select the final HR with high confidence.