Event

ANNOUNCEMENT: Ph.D. Research Proposal Exam

Name: Sanjaya Herath

Committee:

Professor Christopher Metzler (Chair)

Professor Behtash Babadi

Professor Zhambyl Shaikhanov

Date/time: 05/18/2026 from 10:30am – 12:30pm

Location: AVW 2460

Title: Learning Based Methods to Improve RF Sensing

Abstract:  Radio-frequency (RF) sensing systems with large-scale arrays comprising 1000+ elements promise breakthrough capabilities, including ultra-high angular resolution, precise target localization, robust interference suppression, and high-fidelity imaging of complex scenes in challenging environments. These advances enable next-generation applications in radar, wireless communications, and autonomous systems. However, realizing this potential is hindered by limited hardware resources, scalability constraints, and complex propagation effects, which limit the efficiency and practicality of conventional signal processing methods. This proposal addresses these challenges by developing learning-based computational approaches that integrate model-based signal processing with data-driven techniques to enhance sensing performance and efficiency without increasing hardware complexity.

First, we introduce a self-supervised learning-based framework to expand the field of view (FoV) of passive millimeter-wave (mmWave) imaging systems. Unlike traditional approaches that treat aliasing artifacts as a nuisance, we reinterpret them as a source of useful information. Specifically, we cast aliasing removal as an underdetermined deconvolution problem and solve it by enforcing temporal regularity using time-varying Implicit Neural Representations. This approach enables wide FoV reconstruction from limited observations and is validated through both simulations and real-world experiments on a state-of-the-art passive mmWave imaging system.

Second, we address the computational bottleneck of large-scale beamforming. The widely used Minimum Variance Distortionless Response (MVDR) beamformer provides strong interference suppression but incurs cubic computational complexity with respect to the number of antenna elements, limiting its applicability to large arrays. To overcome this limitation, we introduce a scalable learning-based method tailored for large arrays. Through extensive simulations, we demonstrate that the proposed method achieves near-MVDR accuracy with substantially improved efficiency, making a step towards real-time deployment in large-scale array systems.