Loading...
Enhanced Position-Aided Beam Prediction Using Real-World Data and Enhanced-Convolutional Neural Networks
K. I. HAMAD, EHAB MABROOK, M. MOURAD DONKOL, Ahmed
K. I. HAMAD, EHAB
MABROOK, M. MOURAD
DONKOL, Ahmed
Type
Supervisor
Date
2025-04-23
Collections
Research Projects
Organizational Units
Journal Issue
Abstract
Millimeter-wave (mmWave) communication systems utilize narrow beamforming to ensure adequate signal power. However, beam alignment requires significant training overhead, especially in highmobility scenarios. Previous research has utilized synthetic data for position-aided beam prediction, which does not fully capture real-world complexities. In this work, an Enhanced Convolutional Neural Network
model (E-CNN) is proposed for optimal prediction of beam indices with the aid of real-world GPS position data. The proposed E-CNN model has been investigated across nine different scenarios from the DeepSense 6G dataset and compared against the conventional algorithms. For 64-beams Scenario 1, the E-CNN model showed an increase in average top-1 accuracy from 55.57% to 63.92%, and in case of 32-beams, the accuracy increased from 71.34 % to 82.06%. For 16-beams, the accuracy increased from 86.17% to 94.64 %, while for 8-beams, the accuracy increased from 90.24% to 97.11%. In addition, besides showing significant power loss reduction in various scenarios, the proposed E-CNN model has demonstrated robustness regarding real-word conditions and adaptability for various beam setups. The model realized as high as a 50% power loss reduction in arguably the most challenging graphs, which is an exercise in reliability. This research fills the existing gap between the simulated aid beam alignment and real-world position beam aided alignment, which can be useful in improving beamforming in the upcoming wireless networks.