Wireless Localization via Learned Channel Features in Massive MIMO Systems

Future wireless networks will evolve to integrate communication, localization, and sensing capabilities. This evolution is driven by emerging application platforms such as digital twins, on the one hand, and advancements in wireless technologies, on the other, characterized by increased bandwidths, more antennas, and enhanced computational power. Crucial to this development is the application of artificial intelligence (AI which is set to harness the vast amounts of available data in the sixth-generation (6G) of mobile networks and beyond. Integrating AI and machine learning (ML) algorithms, in particular, with wireless localization offers substantial opportunities to refine communication systems, improve the ability of wireless networks to locate the users precisely, enable context-aware transmission, and utilize processing and energy resources more efficiently. In this dissertation, advanced ML algorithms for enhanced wireless localization are proposed. Motivated by the capabilities of deep neural networks (DNNs) and the advancements in massive multiple-input-multiple-output (MIMO) systems, the dissertation aims to address some of the fundamental limitations of ML-based localization approaches related to dependability, generalization, and data scarcity. The dissertation has three main parts, each dedicated to addressing specific challenges and introducing new algorithms. The first part of this dissertation focuses on improving the scalability and reliability of supervised learning techniques for wireless localization. Here, two variational DNN approaches are presented, designed to overcome the limitations in measuring the uncertainty of DNN-based position estimates in co-located massive MIMO systems. Further, this part extends the investigation to assess the localization accuracy in a distributed antenna system (DAS). It also introduces a strategy for selecting the most relevant subset of remote radio heads (RRHs) to alleviate the fronthaul overhead associated with dense wireless network deployments. The second part of the dissertation is a transition from supervised to unsupervised learning. This part is concerned with subspace and metric-learning approaches that can learn low-dimensional channel features. Addressing the challenges related to data scarcity, this part introduces a contrastive task and a Siamese-based DNN to learn a four-dimensional channel representation that is useful for wireless localization. Compared to a base DNN classifier, the proposed method significantly improves localization performance, particularly in small data and non-line-of-sight (NLOS) conditions. Finally, the third part of the dissertation reconsiders the foundational components and optimization strategies generally used in DNN-based localization methods. It first introduces a transformer-based model for more robust channel feature learning. It then builds upon the transformer-based method and proposes a non-contrastive self-supervised learning (SSL) approach. This part of the dissertation shows how to exploit the macroscopic and microscopic characteristics of the channel to achieve better transfer learning across various configuration settings, propagation environments, and wireless downstream tasks. Moreover, it investigates multiple variants of transform-based models and uses multiple evaluation approaches and datasets to assess the localization accuracy in both co-located massive MIMO systems and a DAS.