Physics-informed generative neural networks for RF propagation prediction with application to indoor body perception

TL;DR

The paper tackles the challenge of real-time RF propagation prediction in the presence of human body motion, where traditional EM body models are too slow for strict real-time sensing. It introduces a physics-informed generative framework based on a Conditional Variational Auto-Encoder that learns a prior over EM field responses $E_theta$ conditioned on coarse body features $theta_k$, enabling fast generation of body-diffraction samples. The CVAE is trained to reproduce diffraction-based effects and is validated against both a diffraction model and FEKO full-wave simulations, demonstrating consistent array-level responses in dense MIMO configurations. The approach achieves generation rates on the order of tens of samples per second at $f_c=2.4$ GHz, offering a practical surrogate for real-time passive body localization and sensing with ambient RF signals.

Abstract

Electromagnetic (EM) body models designed to predict Radio-Frequency (RF) propagation are time-consuming methods which prevent their adoption in strict real-time computational imaging problems, such as human body localization and sensing. Physics-informed Generative Neural Network (GNN) models have been recently proposed to reproduce EM effects, namely to simulate or reconstruct missing data or samples by incorporating relevant EM principles and constraints. The paper discusses a Variational Auto-Encoder (VAE) model which is trained to reproduce the effects of human motions on the EM field and incorporate EM body diffraction principles. Proposed physics-informed generative neural network models are verified against both classical diffraction-based EM tools and full-wave EM body simulations.

Figures (6)

• Figure 1: Link geometry with body, sketched as a 2D EM perfectly absorbing sheet, and Multiple-Input-Multiple-Output (MIMO) antennas placed in the monitored area. The top view of this area is shown on the right as well.
• Figure 2: Conditional VAE system, encoder and decoder GNN structures.
• Figure 3: Array layout setups (top) and corresponding responses (bottom) $R_{\theta}(\gamma)$ for $\gamma=\pi/2$. Target has dimensions $h_{S}=1.65$ m, $w_{S,1}=0.55$ m, $w_{S,2}=0.25$ m. Two subject positions are considered, namely $x=2$ m, $y=0.25$ m (left) and $x=2$ m, $y=-0.25$ m (right). Response is obtained using VAE-generated field samples $\widehat{\mathbf{E}}_{\theta}^{\mathrm{VAE}}$ (green/red solid lines) and compared with the array response obtained with the EM body diffraction model (dashed lines).
• Figure 4: Body-shape model built in FEKO® with size $h_{S}=1.80$ m, $w_{S,1}=0.52$ m, and $w_{S,2}=0.32$ m. Simulation settings are shown as well.
• Figure 5: Excess attenuation values obtained by FEKO® software (solid) and compared with C-VAE generation (dashed) for varying antennas of the array (9 antennas) and target dimensions. Three nominal subject positions are considered, namely $x=2$ m, $y=0.25$ m (red), $x=2$ m, $y=0$ m (black) and $x=2$ m, $y=-0.25$ m (green). For all cases, attenuation responses account for small body movements around the nominal positions.