University of British Columbia
Learning to calibrate : graph neural network for online calibration of phase shifters in hybrid beamforming systems
Abstract
dc:descriptionThe rapid increase in cellular network traffic and user demands has boosted interest in millimeter-wave (mmWave) communications for fifth-generation (5G) and beyond, essential for addressing bandwidth and capacity limitations. However, mmWave communications face challenges, notably severe path loss due to short wavelengths. To address this, high-gain directional radiation patterns are essential. These are achieved through the use of large-scale antenna arrays employing hybrid analog and digital beamforming, for ensuring reliable, high data-rate transmissions. Nonetheless, these systems' phase shifter networks (PSNs), which steer the radiation pattern in the desired direction, are highly susceptible to phase deviations resulting from manufacturing defects, hardware degradation, and temperature shifts. These deviations significantly affect beam direction, necessitating periodic PSN calibration to maintain operational effectiveness. Online calibration approaches can overcome the limitations of conventional calibration methods. It is performed in real-time within the antenna’s operational environment based on over-the-air pilots measurements. This eliminates the necessity for expansive equipment, specialized circuitry, and a separate time-consuming calibration procedure. However, existing online calibration methods typically involve optimization-based approaches, that solve a high dimensional non-convex program. This results in computationally demanding algorithms, which pose an even greater computational challenge for large antenna systems. Addressing the limitations of traditional model-based iterative algorithms, this thesis introduces the first machine learning-based online calibration method dedicated for PSNs in mmWave hybrid beamforming systems. We implement a graph neural network (GNN) architecture that effectively parameterizes a mapping from the received pilots signals to optimized PSN phase deviation estimates. Our GNN design is highly suitable for the online calibration task, incorporating a heterogeneous graph representation of wireless communication systems that integrates prior knowledge into the neural network structure. The design leverages neural network parameter sharing and exploits permutation equivariance and invariance, enhancing scalability and generalization across diverse systems without retraining. To benchmark the performance of learning-based methods, we also derive the hybrid Cramer-Rao bound (HCRB) for the phase deviation estimates. Simulation results validate the effectiveness of our approach, demonstrating that the performance of our GNN closely aligns with the HCRB and exhibits robustness in dynamically changing environments.
Degree
thesis:*- Name thesis:degree_name
- Master of Applied Science - MASc
- Level thesis:degree_level
- master's
- Discipline thesis:degree_discipline
- Electrical and Computer Engineering
- Grantor dc:publisher
- University of British Columbia
- Year dc:date
- 2024
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Roth, Idan
- Contributors dc:contributor
-
- Lampe, Lutz
Rights
dc:rights- Statement dc:rights
-
- Attribution-NonCommercial-NoDerivatives 4.0 International
- Language dc:language
- eng
Identifiers
dc:identifier.*- Handle dc:identifier
- http://hdl.handle.net/2429/89584