| Author: | Zhao, Xiaopeng |
| Title: | Neural radio-frequency radiance fields : development and applications |
| Advisors: | Yang, Lei (COMP) |
| Degree: | Ph.D. |
| Year: | 2025 |
| Department: | Department of Computing |
| Pages: | xx, 204 pages : color illustrations |
| Language: | English |
| Abstract: | Recently, the growth of the global Internet of Things (IoT) market has driven the development of numerous wireless systems. To meet the demands of various applications, these systems operate across diverse frequencies and employ different modulation schemes. A standardized wireless channel model is commonly used to describe radio-frequency (RF) signal propagation of these systems. Accurate channel prediction is crucial for optimizing the performance of communication and sensing technologies. However, complex interactions such as reflection and diffraction between RF signals and environmental entities pose challenges. Conventional methods of wireless channel prediction struggle with the complex nature of real-world environments. To address these limitations, this thesis first introduces the design of Neural Radio-Frequency Radiance Fields (NeRF2), a model that provides precise predictions of wireless channels at any location in the environment. NeRF2 begins by measuring a sparse set of signals from the scene of interest and then employs these measurements to train a neural radiance field. This field represents the scene as a continuous volumetric function, detailing the electromagnetic properties of each voxel. Once trained, NeRF2 can predict the wireless channel at previously unmeasured locations by tracing new paths through the voxel representation. As a model at the physical layer, NeRF2 has extensive applications, including enhanced channel prediction in Frequency Division Duplex (FDD) systems and improved accuracy in wireless localization. Moreover, random phase noise from sources like oscillator jitters in real-world wireless systems impacts channel estimations, especially in low-power backscatter systems. To address this issue, we introduce consistent phase estimation protocols that effectively overcome the challenges of achieving accurate phase estimation across long distances. These protocols resolve the π-ambiguity commonly encountered in commercial RFID readers. Additionally, we further refined these protocols to eliminate flicker noise and neutral white noise, and to correct spatial and temporal imbalances, thereby enhancing the robustness of the system. Owing to its precise channel prediction, NeRF2 can generate synthetic training datasets that improve deep-learning-based indoor localization algorithms. However, conventional methods still face challenges from their reliance on high-quality training data and limited adaptability in diverse environments. To overcome these limitations, we introduce the Transformer-based Localization (TBL) model, improving the localization accuracy by its historical position understanding. We also propose a semi-supervised training approach to reduce the need for extensive label collection. To further enhance transferability, we developed LocGPT, a pre-trained version of the TBL model, using 1.3 million data samples. In new environments, this model requires only a minimal dataset for fine-tuning. Another application of NeRF2 is in environmental sensing. We introduce Satellite Radiance Fields (SaRF), an approach that uses crowdsourced smartphone GPS data to construct accurate 3D urban maps. This methodology capitalizes on the observation that materials with higher densities, such as concrete and metal, typically cause greater attenuation of RF signals. By training a neural radiance field, SaRF accurately learns the attenuation properties of each voxel with respect to satellite signals, enabling the detailed reconstruction of 3D voxel maps. In conclusion, this thesis sheds light on the open challenges in modeling the propagation of RF signals. To address the challenges, we introduce the NeRF2 and demonstrate its effectiveness in applications such as wireless indoor localization and 3D urban voxel map reconstruction. Looking ahead, we discuss potential future work, including adapting the neural channel model to dynamic environments. Additionally, we explore transferring methodologies from optical to electromagnetic neural radiance fields, potentially enhancing the implementation of wireless digital twin networks. |
| Rights: | All rights reserved |
| Access: | open access |
Copyright Undertaking
As a bona fide Library user, I declare that:
- I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
- I will use the Database for the purpose of my research or private study only and not for circulation or further reproduction or any other purpose.
- I agree to indemnify and hold the University harmless from and against any loss, damage, cost, liability or expenses arising from copyright infringement or unauthorized usage.
By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms.
Please use this identifier to cite or link to this item:
https://theses.lib.polyu.edu.hk/handle/200/13641