Deep Learning for Seamless Location-Awareness in 5G and Beyond Wireless Networks

Abstract

USER localization and tracking play a critical role in future wireless communication systems, where beamforming-based radio links at millimeter-wave (mmWave) frequencies enable fine-grained spatial multiplexing and access to large bandwidths. Nevertheless, the operation of such deployments becomes increasingly complex due to the large number of parameters and different radio channel characteristics. Highly accurate and reliable location information enables seamless operation while preserving resources. Integrating Artificial Intelligence (AI) models within the communication systems is one of the key enabling innovations when moving beyond 5th Generation Mobile Networks - New Radio (5G NR).

This Thesis presents a number of contributions in the field of AI-based solutions for localization, tracking, and location awareness within wireless networks. The first contribution of this Thesis targets the proactive functionality of handover (HO) management within mmWave networks, where sustaining strong radio links without unnecessary HOs is a critical challenge to ensure high throughput without interruptions, as well as preserving radio resources. To this end, Artificial Neural Network (ANN) models tackling HO management are proposed, utilizing the estimated user location as an additional input boosting mobility-related decision-making. Furthermore, a separate and hierarchal solution for cell-level and beam-level mobility is proposed, as beam-level mobility can be operated locally and without additional overhead, while cell-level mobility generally requires additional signaling and temporary connection interruptions. A numerical evaluation of the proposed approaches shows consistent localization capabilities reliably reaching meter-level accuracies and up to 95% reduction of unnecessary HOs caused by signal uncertainties.

The second topic addressed in this Thesis concerns signal uncertainties and nonline- of-sight (NLoS) condition as undesired, yet unavoidable aspects in communication systems, and proposes techniques and ANN approaches to mitigate their effects on positioning capabilities. While considering realistic channels and uncertainties, the proposed feature pre-processing techniques and model architectures are capable of achieving meter-level positioning accuracies in both line-of-sight (LoS) and NLoS channel conditions. The contributions include an analysis of network measurements, as well as refinement of novel features from channel frequency response capable of stable and reliable channel representation. Additionally, the numerical analysis shows that considering temporal dependencies within the data further improves tracking capabilities on both LoS and NLoS conditions, while estimating user velocity and heading as additional outputs with almost perfect accuracy.

Changing building layout or altering base station (BS) deployment affects signal propagation patterns throughout the environment. The third contribution in this Thesis concerns adapting the machine learning (ML) model to data heterogeneity and changes within mmWave network deployments, as one of the key challenges of implementing ANN into real-world systems. The contributions consider applying a transfer learning strategy, adjusting the model to the changed deployment with limited data, developing a novel hypernetwork (HN) model for efficient factory pre-training in numerous scenarios, and combining location information from different sources using a data fusion model to improve localization performance. The numerical results of Transfer Learning (TL) approaches indicate faster model adaptation and improved localization accuracies, especially when paired with the proposed HN model. The fusion model is capable of reducing the localization errors by more than 50% compared to the stand-alone model.

The final contribution of this Thesis addresses data scarcity in indoor positioning systems by augmenting ANN training strategies and architectures to alleviate their requirements of voluminous training datasets. A strategy to combine numerous small heterogeneous datasets to a unified representation via ANN encoder is shown to improve positioning performance across deployments by up to 25%. Similarly, a novel architecture adapting the classification-based model to a robust regressor shows significant and consistent performance improvements over traditional model architectures and other baselines.

The overall findings show that implementing ANN models within the communication networks reduces the number of HOs by up to 95% compared to the current algorithm, and enables reaching sub-meter-level localization accuracies indoors, as well as in densely built areas, despite the received signal being subject to significant uncertainties and blockage. Furthermore, the Thesis proposes several key strategies and advancements in the field of ML in the form of novel layers, architectures, or algorithms.

The Thesis presents clear advancements and novelty in the fields of ANNs and wireless communications, with contributions beyond the current state-of-the-art.