@Article{cmc.2026.080559,
AUTHOR = {Zheng Xu, Zihao Pan, Ning Yang, Daoxing Guo},
TITLE = {LRT-BF: A Lightweight and Robust Blind Beamforming Method for High-Dynamic UAV Communications},
JOURNAL = {Computers, Materials \& Continua},
VOLUME = {},
YEAR = {},
NUMBER = {},
PAGES = {{pages}},
URL = {http://www.techscience.com/cmc/online/detail/26846},
ISSN = {1546-2226},
ABSTRACT = {Unmanned Aerial Vehicle (UAV) communications in complex electromagnetic environments face challenges such as strong interference, high dynamic Doppler shifts, and limited onboard computing power. In these scenarios, traditional blind beamforming algorithms suffer from slow convergence and difficulty in handling Gaussian-like signals (e.g., Orthogonal Frequency Division Multiplexing (OFDM)). To address these issues, this paper proposes a Lightweight Robust Transfer learning-based Blind Beam Forming method (LRT-BF). This method constructs a self-supervised optimization framework centered on a pre-trained signal classifier and innovatively introduces a joint loss function combining classification confidence guidance with output power minimization, achieving fully blind interference suppression without requiring Direction of Arrival (DOA) priors. To address the high dynamic characteristics of UAVs, a Frequency Domain Randomization (FDR) augmentation strategy is introduced, endowing the feature extractor with Doppler-invariant perception capabilities under frequency offsets of ±5 kHz. By reconstructing the network backbone using Depthwise Separable Convolutions (DSC), a computational reduction of 8.3× and parameter reduction of 7.0× are achieved with negligible accuracy loss (retaining up to 99.8% accuracy). Furthermore, by incorporating Temperature Scaling mechanisms and signal subspace initialization, the problems of gradient saturation and convergence stagnation in few-snapshot scenarios are effectively resolved. Simulation results demonstrate that under conditions of extremely few snapshots (L<64) and strong interference, the average interference suppression depth of LRT-BF reaches −41.2 dB, an improvement of over 48 dB compared to traditional Fast Independent Component Analysis (ICA) and Constant Modulus Algorithm (CMA) algorithms. Its Central Processing Unit (CPU) inference latency is only 1.64 ms, achieving a 4.6× real-time acceleration. Beyond these theoretical metrics, these hardware-efficient characteristics confirm the immense potential of LRT-BF for practical implementation, providing a highly feasible, low-latency anti-jamming solution for SWaP-constrained UAV swarms and edge nodes.},
DOI = {10.32604/cmc.2026.080559}
}