Navigating the Future — Machine Learning for Wireless Sensing and Localization
Year: 2024
Type: dissertation
Abstract: Cellular-based localization and sensing pave the way for a variety ofapplications across various domains, ranging from autonomous driv-ing to emergency care and intelligent traffic management. Althoughtraditional methods have been effective, they still face challenges such as therequirement for highly accurate models and the inherent complexity of thealgorithms. This thesis explores the potential of integrating machine learning(ML) techniques to augment the performance of sensing and localizationsystems. Three main topics are covered by this thesis, namely, ML-aidedchannel estimation, sensing, and localization.The first topic focuses on the calibration of the RF chain and the esti-mation of the propagation channel, which serves as essential prerequisitesfor numerous subsequent applications including arriving angle estimation,radio localization, digital beamforming, and sensing. We introduce a novelRF chain calibration algorithm for massive multiple-input multiple-output(MIMO) systems, using uplink signals. We derive the maximum likelihoodestimator (MLE) and its corresponding Cramér-Rao Lower Bound (CRLB).Additionally, we propose a novel ML-powered channel estimation pipeline fororthogonal frequency division multiplex (OFDM) systems, which efficientlyestimates channel coefficients for all OFDM grids based on only using alimited number of pilot signals.The second topic addresses ML-based wireless sensing using massiveMIMO systems. We introduce a novel pipeline that first employs tensor-decomposition algorithms to extract channel characteristics. Subsequently, afully connected neural network is deployed to classify human movements. Our pipeline is evaluated by a measurement campaign using a massiveMIMO testbed in an indoor environment, providing empirical evidence ofthe system’s efficacy in wireless sensing applications. Our work shows thesensing capacity of massive MIMO in such scenarios.The third topic focuses on cellular localization, comprising three mainsections. The first section introduces classical radio localization algorithms,including time of arrival (ToA), angle of arrival (AoA) and time differenceof arrival (TDoA). The second main section addresses ML-based localizationwith a massive MIMO system. We introduce a novel pipeline composed ofseveral parallel processing chains. Each chain is trained on distinct channelfingerprints, including the channel impulse response (CIR) and covariancematrices. To improve localization accuracy, we model the position error ineach processing chain as a Gaussian distribution and combine the outputs tocompute localization uncertainty. Furthermore, we investigate the requiredtraining density using the Nyquist theorem. Our pipeline is evaluatedthrough indoor measurements, showing a centimeter-level localization ac-curacy. The third section investigates ML-aided cellular localization usinga 5G new radio (NR) system that operates in beam space. We presentan ML-based localization pipeline that integrates attention mechanisms andadvanced uncertainty estimation algorithms. Unlike in the previous section,this uncertainty estimation approach is not restricted to a Gaussian assump-tion. Validation of our pipeline is conducted through an extensive outdoormeasurement campaign encompassing both Line-of-Sight (LoS) and NonLine-of-Sight (NLoS) propagation scenarios. Measurement results indicatethe sub-meter accuracy levels with our pipeline.
Author: Tian, Guoda
Language: English
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7067421231
Cites:
Cited by:
Related to: