Complex-Valued Neural Networks for Distributed Machine Learning Assisted RF Positioning

Despite recent progress in applying machine learning to RF positioning, the potential of complex-valued neural networks (CVNNs) in multi-antenna systems remains largely untapped. In this paper, we present a novel low-complexity CVNN model that jointly performs (1) line of sight classification and (2) first path time of arrival (ToA) estimation for radio link. In the 3GPP indoor factory test environment, the trained model achieves 95.3% LoS/NLoS classification accuracy which is much higher than conventional solutions. The model also produces highly accurate ToA estimates for the LoS links. The trained model is tested in environments substantially different than the environment data used for training. The model is found to be robust and generalized well to different operating environments. Using the outputs of the CVNN model, 2D positioning error of 4.7 cm at 90<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">th</sup> percentile can be achieved.