Unveiling Cellular Antenna Orientations from Large Crowdsourced Datasets: A Deep Learning Approach

Abstract

The accurate and reliable localization of transmitter locations from crowdsourced measurements has enabled the large-scale analysis of the previously hidden network layout. Recent work has shown that signal-strength measurements from drive-test campaigns also unveil the antenna orientations - pro-viding an open-source network twin that can act as the backbone of coarse user-equipment positioning, operator benchmarking, or the generation of coverage and performance maps. In this work, we extent this drive-test based scheme to the regime of large-scale crowdsourced datasets and conduct an assessment of orientation inference on 4,950 sectors from a live LTE network. Fusing signal-strength and geometry-based features in a probabilistic Deep Learning image processing framework tackles the challenging characteristics of such noisy crowdsourced data collected in uncontrolled conditions. We further use transfer learning and weight sharing to extend our approach to allow for joint inference of sectors mounted onto the same base station. On the test dataset - representative of a complete network deployment - our selective predictors achieve median errors as low as 7.3° with 95 percentiles below 21°.