Machine Learning Based NLOS Radio Positioning in Beamforming Networks

In this paper, we address the challenging problem of radio positioning in non-line-of-sight (NLoS) conditions. To this end, we utilize measurements in the form of time-of-flight and gNodeB angular information in the context of 5G New Radio (NR) networks. Such measurements are processed by artificial neural networks with different snapshot and sequence-processing architectures to track the positions of the terminals. For model training, we consider a crowdsensing data acquisition scheme to effortlessly gather the desired measurements with the synchronized location tags. Realistic ray-tracing based evaluations on the so-called Madrid map at 28 GHz millimeter-wave band are provided, to assess the achievable performance while also varying the amount of uncertainties within the data. The obtained results show that radio positioning is feasible with accuracy in the order of 1 meter, or even below, also in challenging NLOS scenarios if the data and measurement uncertainties are small. The results also show that the sequence processing approach offers superior performance under practical measurement uncertainties.