AGMANet: Adaptive-Grid Multi-Head Attention Network for Robust Large-Scale Outdoor LTE Localization
IEEE Access · 2026 · 14 · 97526-97545
DOI: 10.1109/access.2026.3707316 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Mohamed Magdy El-Sayed und 2 weiteren Autor:innen. Die Arbeit nennt 61 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Fingerprint-based localization using deep learning has recently emerged as a reliable solution to outdoor positioning in cellular networks. However, existing approaches typically rely on fixed-resolution uniform grids that do not adapt to the non-uniform spatial distribution of user measurements, and employ simplistic decoding strategies that limit positioning accuracy. In this paper, we propose AGMANet, an adaptive grid, sequence-aware localization framework based on a Bidirectional Long Short-Term Memory (BiLSTM) network with Multi-Head Self-Attention (MHSA), leveraging mobile network radio features to achieve high localization accuracy. AGMANet introduces three key contributions: (1) a newly developed signal-aware weighted centroid technique that assigns greater weights to measurements with stronger reference signals, improving grid-cell centroid estimation; (2) a novel adaptive quadtree grid technique that employs a principled set of splitting rules, including density, spatial spread, child viability, and error enhancement gain gates, to concentrate resolution where it is most needed; and (3) an innovative top- $k$ temperature-scaled posterior-weighted centroid decoder that aggregates predictions from the most confident grid cells to mitigate boundary misclassification effects. The framework was evaluated on a large-scale commercial Long-Term Evolution (LTE) dataset covering a 45 km2 outdoor environment. AGMANet consistently outperformed state-of-the-art sequence-based models, achieving a median absolute error of 6.5 m, representing a 31% improvement over the strongest baseline. It also reduced the interquartile range by approximately 11%, indicating significantly improved robustness to localization outliers, while employing approximately 10% fewer trainable parameters than state-of-the-art sequence-based models, demonstrating superior efficiency–accuracy trade-offs.
Zitieren
APA
El-Sayed, M. M., Fahmy, Y. A. H., & Kafafy, M. (2026). AGMANet: Adaptive-Grid Multi-Head Attention Network for Robust Large-Scale Outdoor LTE Localization. IEEE Access, 14, 97526-97545. https://doi.org/10.1109/access.2026.3707316
Vancouver
El-Sayed MM, Fahmy YAH, Kafafy M. AGMANet: Adaptive-Grid Multi-Head Attention Network for Robust Large-Scale Outdoor LTE Localization. IEEE Access. 2026;14:97526-97545. doi:10.1109/access.2026.3707316
BibTeX
@article{ElSayed2026,
author = {El-Sayed, Mohamed Magdy and Fahmy, Yasmine A. H. and Kafafy, Mai},
title = {AGMANet: Adaptive-Grid Multi-Head Attention Network for Robust Large-Scale Outdoor LTE Localization},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {97526--97545},
doi = {10.1109/access.2026.3707316}
}
EUVIMED