Olivia Widjaja, University of New South Wales Canberra; Melrose Brown, University of New South Wales Canberra; Michael Lachut, EOS Space Systems
https://doi.org/10.64861/RFYE2043
Keywords: Manoeuvre detection, orbit determination, characterisation, machine learning
Abstract:
The capability to detect satellite manoeuvres in geosynchronous Earth orbit (GEO) is a critical component of Space Domain Awareness (SDA). Timely identification of anomalous behaviour enables operators to assess risk and intent, an increasingly important requirement as the population of Resident Space Objects (RSOs) continues to grow. Currently, manoeuvre detection methods relies on statistical thresholding of orbit determination residuals, where dynamical models are required to be simplified to maintain computational efficiency; this can significantly affect the robustness of the algorithm when observability is limited.
Recent advances in Scientific Machine Learning (SciML) have demonstrated that differential equation models can be augmented with learned components while preserving their physical structure. In particular, Universal Differential Equations (UDEs) embed trainable forcing terms directly within governing dynamical equations, enabling inference of unknown or unmodelled effects without discarding established physics. This paradigm builds upon physics-informed learning approaches but retains the numerical integration framework explicitly, treating inference as a modification of the equations of motion rather than a residual constraint. UDE-based methods have demonstrated strong performance in nonlinear system identification and parameter inference problems governed by differential equations.
This study investigates the development of a UDE-based architecture for detecting and characterising impulsive manoeuvres in GEO using angles-only observations. Rather than approximating orbital trajectories with a purely data-driven surrogate, the proposed framework retains a physics-based propagator and introduces a structured manoeuvre term within the equations of motion. The focus is on radial burns, which remain comparatively under-examined in the literature despite their distinctive dynamical signatures in GEO. The method uses a combination of simulated non-manoeuvring and manoeuvring cases observed from a representative ground-based optical site over arcs of up to 24 hours. A manoeuvre data set inspired from literature is extended to include radial burns to examine direction-dependent observability. All datasets will start with a simple two-body system, before increasing the fidelity to 40×40 order spherical harmonic gravity model with solar radiation pressure and lunisolar perturbations.
The primary contribution of this work is the design and analysis of a precursor to a physics-embedded UDE learning architecture. Its utilisation of high-fidelity numerical propagator alongside equations of motion acts as a physics-based bridge across data gaps. Moreover, this study aims to address the detection and characterisation radial manoeuvres, a commonly occurring yet under-studied type of manoeuvre. Currently, a prototype of an orbit estimation pipeline with physics integration has been developed, modelling a simple two-body equation alongside an impulsive manoeuvre. Preliminary results using simulated angles-only data already indicate promise in solving velocity with moderate accuracy. Further work in developing the neural network is required to solve for all unknown or unmodelled effects. Additionally, improving the physics model to account for perturbations such as solar radiation pressure is still required. This study evaluates a differentiable, physics-constrained precursor to a UDE for estimating a single impulsive GEO manoeuvre from angles-only data. The deliberately low-dimensional pulse model is used to test observability, optimisation and discrimination in preparation for future expansion to multi-sequence manoeuvres and continuous low-thrust scenarios.
Date of Conference: September 15-18, 2026
Track: AI for SDA Applications