Continuous vital-sign monitoring enables early detection of abnormal physiological patterns, which is crucial for preventive healthcare. However, conventional monitoring devices face critical limitations: wearables have compliance issues, cameras raise privacy concerns, and clinical devices disrupt patient rest. This thesis addresses these constraints by developing robust Wi-Fi sensing infrastructure that provides a stable signal acquisition foundation upon which future diagnostic applications can be developed.
Wi-Fi systems continuously estimate the wireless channel for communication. This channel estimate inherently captures motion-induced variations through human reflections, enabling detection of physiological motions like respiration. However, detecting such movements is challenging due to environmental noise and interference like walking. Furthermore, commodity Wi-Fi devices limit access to raw in-phase and quadrature (IQ) samples, which restricts channel estimation to just a fraction of the Wi-Fi frame, the training field. Moreover, standard Wi-Fi operates in bistatic configurations with spatially separated transmitter (TX) and receiver (RX), causing phase errors from unsynchronized clocks. Additionally, limited bandwidth yields coarse range resolution of several meters, which is insufficient to separate vital-sign reflections from nearby interference.
These limitations motivate monostatic architectures, where co-located TX and RX chains share a common reference clock, eliminating synchronization errors while enabling easier raw IQ access. Such architectures support in-band full-duplex operation, enabling simultaneous transmission and reception. However, strong self-interference (SI) from TX-RX coupling overpowers weak target reflections. While SI cancellation (SIC) can mitigate SI, conventional analog SIC (AnSIC) systems that generate cancellation signals directly in the analog domain (A-AnSIC) require frequent tuning that inadvertently cancels physiological movements.
This thesis enables stable Wi-Fi-based vital sign monitoring through architectural and signal processing innovations. First, we introduce novel digital SI cancellation (DiSIC) techniques leveraging neural networks, which effectively capture RF non-linearities to enable high-quality SIC with lower computational complexity compared to polynomial models. Then, we develop an initial monostatic sensing system, which demonstrates accurate multi-target breathing rate estimation. To provide analog SIC, we develop an AnSIC system that digitally generates the cancellation signal (D-AnSIC), achieving stable 40 dB cancellation without continuous tuning. To further remove the SI, we introduce our SI distillation and removal (SIDAR) method, which combines phase calibration with low-rank decomposition to isolate SI from weak target reflections. SIDAR enables reliable 10-meter non-line-of-sight monitoring even without D-AnSIC, though combining both yields optimal performance. Finally, w
