Machine Learning-Based Fingerprint Positioning for Massive MIMO Systems

Abstract

In this paper, we investigate a user terminal (UT) fingerprint positioning problem in massive multiple-input multiple-output (MIMO) orthogonal frequency division multiplexing systems under non-line-of-sight scenario. Exploiting the advantages of a large-scale array and wide bandwidth, we introduce a spatial refined beam-based channel model and advocate a beam domain channel amplitude matrix as location-related fingerprint embedding abundant and stationary multi-path information, including amplitude, angle of arrival and delay of arrival, which are favorable to positioning. Taking the fingerprints as the input, we propose a novel machine learning enabled positioning method to localize the 2-dimension position of UTs. Specifically, we propose two fully-connected neural networks (NNs) to complete the task. The first NN is used to classify the position coordinates of the UTs, and the second NN functions as the regression module to estimate the UTs' position coordinates. The simulation results present that the first NN can classify the UTs efficiently and the second NN can reliablely regress the UTs' location. With the increasing of beams, the positioning performance of the proposed methods can achieve superior accuracy compared with conventional approaches.

A Selection of Starting Points for Iterative Position Estimation Algorithms Using Feedforward Neural Networks

“…The mobile terminal measured beam reference signal received power (BRSRP) data, which were processed by the neural network and random forest algorithm together with direction of departure (DoD) data from the beamformer to estimate a terminal position in an urban area, with additional data from a LOS-NLOS detector. A similar setup with one base station equipped with an antenna array is also considered in [ 29 ], but in this publication, a channel frequency response matrix was estimated as position-related data to be processed by the neural network.…”