Covariance Difference of Arrival based Fingerprinting Localization

Abstract

We define covariance difference of arrival (CDOA) features derived from channel state information that can be used for machine learning based fingerprinting localization in nonline of sight (NLoS) conditions, with minimal communication overhead. Taking advantage of the uniqueness of the multipath channel between the base station (BS) and user equipment (UE) at different locations in the geographical region of interest. UEs compute CDOA features, consisting of pair-wise distances between covariance matrices of received signals from multiple BSs. Measured features are fed back to the network, where fingerprinting localization is performed. We consider both k-nearest neighbour and neural network localization, and investigate the trade-off between localization performance and communication overhead. In simulations of a NLoS 5G NR factory scenario with eightantenna BSs, CDOA features provide a localization error less than 0.91 m in 80% of the cases, as compared to 0.78 m for a benchmark method where UEs feed back complete measured covariance matrices to the network, and 1.36 m for power difference of arrival features. Comparing to complete covariance feedback, CDOA features reduce communication overhead by 98%.