A Physics-Informed Edge-AI Classifier for 3-D WiGig Beam Sensing Problem

Abstract

This paper presents a novel approach to addressing Line-of-Sight blockage and deafness in WiGig device-to-device (D2D) communication links through a compact, physics-informed learning framework. We propose a 60 GHz electromagnetic field sensor based on a four-element patch antenna array and a lightweight Multi-Layer Perceptron classifier. A three-dimensional dataset is generated using a 3D implementation of the Friis transmission equation, incorporating polarization loss and realistic antenna gain patterns, and represented in spherical coordinates $(\rho, \phi, \theta)$. The trained MLP accurately classifies the spatial quadrant of an RF source from received power measurements. The model is deployed on an STM32N6 microcontroller using TensorFlow Lite, demonstrating efficient edge-AI inference. Results show minimal memory usage (as low as 177 Bytes Flash) and low computational cost (146 MACCs), with inference latency down to 0.049 ms, enabling practical, low-power mmWave D2D sensing solutions.