Development of a Machine Learning based Radio source localisation algorithm for Tri-axial antenna configuration

TL;DR

This paper presents a machine learning approach using a tri-axial antenna setup and neural networks to accurately and rapidly localize radio sources, overcoming challenges faced by traditional methods at low frequencies.

Contribution

It introduces a novel ML-based radio source localization algorithm with a tri-axial antenna configuration, achieving high accuracy and real-time inference capabilities.

Findings

High localization accuracy with low loss (~0.02)

Rapid inference time (~5 ms)

Effective across a wide frequency range (0.3-30 MHz)

Abstract

Accurately determining the origin of radio emissions is essential for numerous scientific experiments, particularly in radio astronomy. Conventional techniques, such as the use of antenna arrays encounter significant challenges, specially at very low frequencies, due to factors like the substantial size of the antennas and ionospheric interference. To address these challenges, we employ a space-based single-telescope that utilizes co-located antennas, complemented by goniopolarimetric techniques for precise source localization. This study explores a novel and elementary machine learning (ML) technique as a way to improve and estimate Direction of Arrival (DoA), leveraging a tri-axial antenna arrangement for radio source localization. Employing a simplistic emission and receiving antenna model, our study involves training an artificial neural network (ANN) using synthetic radio signals. These synthetic signals can originate from any location in the sky and cover an incoherent frequency range of 0.3 to 30 MHz, with a signal-to-noise ratio (SNR) between 0 and 60 dB. Then, a large data set was generated to train the ANN model catering to the possible signal configurations and variations. After training, the developed ANN model demonstrated exceptional performance, achieving loss levels in the training ($\sim0.02$), validation ($\sim0.23\%$), and testing ($\sim0.21\%$) phases. The machine learning-based approach remarkably, exhibits substantially quicker inference times ($\sim5$ ms) in contrast to analytically derived Direction of Arrival (DoA) methods, which typically range from 100 ms to a few seconds. This underscores its practicality for real-time applications in radio source localization, particularly in scenarios with limited number of sensors.