Beamformed Channel Matrix Positioning using 5G Testbench CSI data with a Deep-Learning Pipeline

Abstract: Within the telecommunications industry, a positioning system for estimating user equipment (UE) location using purely information available for the basestation has an enormous number of potential uses. The link between physical position and the network channel state enables potential positioning systems to function by under- standing the network channel state dependency on location, using a model-based, data-based, or a combined approach. A key exploitable phenomenon linked to position is that of multi-path propaga- tion, wherein transmissions can arrive from multiple directions to the UE, with a unique propagation pattern corresponding to a unique environment. In fifth gen- eration wireless technology (5G), multi-path components are already exploited for beamforming with massive multiple-input multiple-output (MIMO) technol- ogy. Basestations therefore have a preexisting pipeline for obtaining beamformed channel matrices from channel state information (CSI) transmitted by the UE. A data-driven approach using multi-path propagation phenomena for positioning is possible through utilizing the already available beamformed channel matrix in the basestation. In this thesis the practical data-driven deep-learning approach for UE position- ing in 5G using beamformed channel matrices is examined. Real-world data is utilized to judge the applicability of the approach, with measurements done on a commercial-grade Ericsson 5G testbench in both non-line-of-sight (nLoS) and line- of-sight (LoS) scenarios. Using a similar approach to other papers in the field, a su- pervised deep-learning approach is used for instantaneous position estimation. For improving positioning accuracy through trajectory estimation, a novel approach of using particle filtering with network ensemble outputs for kernel density estima- tion of an observation probability density function is proposed. The results show that using the outlined methods position is possible to estimate in real-world pedes- trian tests with a mean accuracy of 2-5 meters, even with nLoS conditions and poor underlying GNSS training data quality