A CNN Approach for 5G mmWave Positioning Using Beamformed CSI Measurements

TL;DR

This paper introduces a CNN-based method for 5G mmWave positioning using beamformed CSI data, demonstrating high accuracy in urban environments with minimal mean error of under 1 meter.

Contribution

It is the first to utilize AI with CNNs trained on beamformed CSI for 5G positioning, showing promising results in urban scenarios.

Findings

CNN achieves a mean error of 0.98 meters in urban environments.

The trained model is robust for measurements at known reference points.

Position estimation remains accurate for points outside the training set.

Abstract

The advent of Artificial Intelligence (AI) has impacted all aspects of human life. One of the concrete examples of AI impact is visible in radio positioning. In this article, for the first time we utilize the power of AI by training a Convolutional Neural Network (CNN) using 5G New Radio (NR) fingerprints consisting of beamformed Channel State Information (CSI). By observing CSI, it is possible to characterize the multipath channel between the transmitter and the receiver, and thus provide a good source of spatiotemporal data to find the position of a User Equipment (UE). We collect ray-tracing-based 5G NR CSI from an urban area. The CSI data of the signals from one Base Station (BS) is collected at the reference points with known positions to train a CNN. We evaluate our work by testing: a) the robustness of the trained network for estimating the positions for the new measurements on the same reference points and b) the accuracy of the CNN-based position estimation while the UE is on points other than the reference points. The results prove that our trained network for a specific urban environment can estimate the UE position with a minimum mean error of 0.98 m.