Source Range Estimation Using Bayesian Deep Learning For Single Antenna Software Defined Radio Receivers

Abstract

The current physical model for radio frequency source-range estimation using single antenna receivers has substantial prediction error in real-world settings and does not provide uncertainty estimates for these predictions. The problem is amplified when the source is unknown and detailed information about its environment is not available. This research addresses this issue by developing deterministic and Bayesian deep neural network models that outperform the physical model on, and provide high-quality uncertainty estimates for, synthetic data, which were generated using a newly developed framework. Moreover, it presents the epistemic uncertainty as a rough indication of trust for the models prediction, while acknowledging the challenges of establishing an absolute reliance threshold. Additionally, this thesis proposes using the physical model as a prior for the variational inference methods, resulting in models that outperform those using standard zero-mean Gaussian priors. This study paves the way for future development of new radio frequency source-localization techniques for simple single antenna Software Defined Radios, with applications in signals intelligence organizations, Maritime Search and Rescue Operations, and other domains where the use of complex receivers is not practical.

Open PDF

Document Details

Document Type
Technical Report
Publication Date
Sep 01, 2023
Accession Number
AD1224537

Entities

People

  • Tomas A. Richards

Organizations

  • Naval Postgraduate School

Tags

Fields of Study

  • Computer science

Readers

  • Computational Modeling and Simulation
  • Positioning, Navigation, and Timing (PNT) Technology.
  • Statistical inference.

Technology Areas

  • AI & ML
  • AI & ML - Bayesian Inference
  • AI & ML - Neural Networks