CSI-based Positioning in Massive MIMO systems using Convolutional Neural Networks

TL;DR

This paper demonstrates that a CNN trained on CSI data from Massive MIMO systems can accurately estimate indoor user positions with less than half a wavelength error, and transfer learning enables adaptation to different antenna configurations.

Contribution

The study introduces a CNN-based positioning method using CSI data and shows effective transfer learning across different antenna topologies in indoor MaMIMO systems.

Findings

CNN achieves less than half wavelength mean error in indoor positioning

Transfer learning reduces retraining effort for new antenna configurations

Dataset includes diverse indoor MaMIMO CSI measurements

Abstract

This paper studies the performance of a user positioning system using Channel State Information (CSI) of a Massive MIMO (MaMIMO) system. To infer the position of the user from the CSI, a Convolutional Neural Network is designed and evaluated through a novel dataset. This dataset contains indoor MaMIMO CSI measurements using three different antenna topologies, covering a 2.5 m by 2.5 m indoor area. We show that we can train a Convolutional Neural Network (CNN) model to estimate the position of a user inside this area with a mean error of less than half a wavelength. Moreover, once the model is trained on a given scenario and antenna topology, Transfer Learning is used to repurpose the acquired knowledge towards another scenario with significantly different antenna topology and configuration. Our results show that it is possible to further train the CNN using only a small amount of extra labelled samples for the new topology. This transfer learning approach is able to reach accurate results, paving the road to a practical CSI-based positioning system powered by CNNs.