Deep learning with synthetic data for wireless NLOS positioning with a single base station

Abstract

Traditional wireless positioning methods exhibit limitations in the face of signal distortions prevalent in non-line-of-sight (NLOS) conditions, especially in the case of a single base station (BS). Moreover, the adoption of deep learning (DL) methodologies has lagged behind, largely due to the challenges associated with generating real-world datasets. In this paper, we present a comprehensive approach leveraging DL over large-scale synthetic wireless datasets (the recent WAIR-D in this case, which was co-produced by Huawei) to overcome such challenges and address the case of single-BS NLOS positioning. The aim of the paper is to practically explore the extent to which synthetic wireless datasets can help to achieve the positioning objectives. Towards this direction, we develop a map-based representation of a radio link, demonstrating its synergistic effect with feature-based representations in MLPs. Furthermore, we introduce a UNet-based neural model which incorporates input maps and radio link representations and generates as output a heatmap of potential device positions. This model achieves an 11.3-meter RMSE and 76.5% prediction accuracy on NLOS examples (1.5-meter, 99.9% for LOS) assuming perfect information, surpassing the MLP baseline by 47%. Finally, we provide further insights into the model’s ability to predict top device positions, the characteristics of predicted heatmaps as indicators of confidence, and the crucial role of map availability and radio path angles in model performance, thus revealing an unconventional perspective on incorrect predictions.