Physics-guided GPS motion features for UAV mmWave beam prediction under high mobility

Tóm tắt

UAV mmWave systems can exploit multiple information sources to predict beams. Multimodal solutions often achieve higher accuracy, but they also increase sensing requirements, synchronization burden, and deployment complexity. In contrast, GPS-based approaches offer a simpler, lower-cost alternative, but their performance degrades noticeably under high UAV mobility when only position information is used. This paper addresses that limitation from the perspective of input representation design rather than predictor scaling. Specifically, we construct GPSderived motion features in the Earth-Centered, Earth-Fixed coordinate system and reorganize them according to the UE-BS geometry to form radial, direction-aware tangential, and velocity-angular feature families. Experiments on the public DeepSense dataset show that not all physically meaningful motion variables contribute equally to beam prediction. Among the investigated variants, the radial feature provides the most reliable overall improvement, increasing Top-1 accuracy from 73.74% to 74.60% while reducing the mean power loss to 0.1674 dB. The remaining variants exhibit different trade-offs across operating conditions, and the overall results consistently indicate that input feature design has a stronger impact on performance than predictor scaling in GPS-based UAV mmWave beam prediction.