Kriti Tripathi, Clearbox Systems Pty. Ltd.; Travis Bessell, Clearbox Systems Pty. Ltd.; Thomas Q. Wang, Clearbox Systems Pty. Ltd.; Tim Spitzer, Clearbox Systems Pty. Ltd.
Keywords: Passive RF, Machine Learning, Computer Vision
Abstract:
The recent increase in space traffic due to the rapid proliferation of satellites, both governmental and commercial, has necessitated advances in space surveillance technology for improved tracking and characterisation of space objects. Passive Radio Frequency (RF) is a relatively new technology compared with traditional sensors that perform space surveillance, including optical telescopes and radars. Passive RF sensors can assist in tracking and characterisation of active space objects through the RF transmissions that they emit. Passive RF data can be represented as Power Spectral Density (PSD) data which can be used to infer the characteristics and state of a satellite. If no or limited prior knowledge is known about the satellite, such as its approximate orbit and transmission characteristics, a wide bandwidth and long duration capture is required to increase the probability of intercept of a transmission from the satellite using passive RF sensors. Satellites can operate over a wide range of radio frequencies in the same or closely adjacent frequency ranges of terrestrial signals and other satellites, making the identification and attribution of individual satellite signals challenging in these types of captures.
Individual satellite signal identification in a wide bandwidth and/or long duration capture, in the presence of other satellite signals, terrestrial signals and noise, is a laborious task that requires human input. Automation of detection of satellite signals in passive RF data can boost time and resource efficiency as well as lower misclassification due to human error. This can aid in eliminating the dependency on prior knowledge about a satellite required to task a passive RF sensor to a smaller and less cluttered frequency and time window. Additionally, it enables the automated detection of satellites in captures from omni-directional antennas, which increases the number of simultaneous satellites observable by a passive RF system.
This work presents a two-staged approach to automating the Human-In-The-Loop (HITL) component of identifying satellite signals in both small and large bandwidth and time data captures, using: 1. Image Classification, and 2. Object Detection. A customised Convolutional Neural Network (CNN) is designed for the image classification stage that classifies a small bandwidth Power Spectral Density (PSD) data capture as containing a satellite signal or not. Several out-of-the-box object detectors such as Faster Region-based Convolutional Neural Network (RCNN), You Only Look Once (YOLO) and Single Shot Detector (SSD) are compared for the object detection problem whereby the detector determines the frequency and time window of one or more satellite signals in a large bandwidth and long duration PSD data capture. The results show that for the image classification problem, the custom CNN model presented in this paper outperforms an existing Energy Detection technique. Furthermore, for the Object Detection problem, SSD looks to be a promising technique for autonomously detecting satellite signals in large data captures.
Date of Conference: September 27-20, 2022
Track: Machine Learning for SSA Applications