Acta Electronica Malaysia (AEM)

The deployment of Unmanned Aerial Vehicles (UAVs) as aerial user equipments (UEs) or base stations is a key enabler for 5G-Advanced and 6G networks, offering potential for enhanced coverage and agility. However, maintaining a reliable, high-throughput wireless backhaul link for UAVs is challenging due to their dynamic mobility and the resulting rapid channel variations, especially at millimeter-wave (mmWave) frequencies. Conventional beam management protocols, reliant on exhaustive beam sweeping, introduce significant latency and overhead, rendering them unsuitable for highly mobile scenarios. This paper presents a holistic solution: a 28 GHz dual-polarized hybrid beamforming transceiver integrated with a proactive deep learning￾based beam tracking algorithm. The RF front-end features a 16-element phased array with dual-polarized patch antennas and a BiCMOS integrated circuit (IC) beamformer, supporting both azimuth and elevation beam steering. The digital baseband implements a deep recurrent neural network (RNN) that leverages real￾time UAV kinematics (position, velocity, attitude) and historical channel data to predict the optimal beam pair between the UAV and the ground station, bypassing the need for traditional sweeping. A 2 Gbps OFDM waveform was used for over-the-air (OTA) testing. Experimental results demonstrate that the proposed system achieves a sustained throughput of >1.8 Gbps at a distance of 300 meters. Compared to a standard IEEE 802.11ay beam sweeping approach, the deep learning-based beam management reduces beam alignment latency by 94% (from 16.8 ms to <1 ms) and signaling overhead by 98%, enabling seamless handover even under aggressive UAV flight maneuvers with angular velocities up to 150°/s. This work successfully bridges advanced RF hardware with machine intelligence, providing a robust framework for high-speed, low-latency aerial links in next-generation cellular networks.