Dr Yang Yang

yang.yang16@unsw.edu.au (Dr Yang Yang) — Wed, 10 Dec 2025 00:00:00 +0000

Michael Ling is a Research Assistant at UNSW Sydney working on the NSW Space Research Network’s Pattern of Life Identification for Geosynchronous Satellites project. His research uses machine-learning methods to characterise GEO satellite behaviour and support satellite manoeuvre detection and space domain awareness.

His previous research applied multiple sets of orbital proper elements and machine learning to classify low-Earth-orbit debris families.

Satellite Manoeuvre Detection and Pattern of Life

yang.yang16@unsw.edu.au (Dr Yang Yang) — Fri, 01 Nov 2024 00:00:00 +0000

Problem

A geostationary satellite changes its behaviour and the people responsible for the catalogue find out late, or not at all. The hard part is not seeing that the residuals moved — it is deciding whether that movement was a real manoeuvre, a mismodelled solar radiation pressure, or simply a bad track. Get that call wrong in one direction and you lose custody of the object; get it wrong in the other and you generate alarms nobody trusts.

Doing this at the scale of the GEO belt makes it worse. Analyst-led assessment does not scale to hundreds of objects observed nightly, and detectors trained on assumed dynamics degrade quietly as the population and operating practices change around them.

Capability

We build manoeuvre detection and pattern-of-life identification for geostationary and geosynchronous satellites, grounded in observational data rather than simulated behaviour. Three parts:

A labelled optical dataset. The Optical Pattern of Life Analysis and Library (OPAL) dataset gives detectors something real to train and be evaluated against.
Detection methods. Image-based and time-series transformer models for identifying manoeuvres, benchmarked against conventional residual-threshold approaches.
A pattern-of-life framework. Integrating the above so that individual detections become a behavioural characterisation of an object over time, which is what an operator actually acts on.

Related characterisation work — satellite polarimetry through the UNSW Observatory, the ongoing PhD project of Asad Rizvi — supplies independent evidence about attitude and surface state to corroborate a detection.

The work is carried out by Michael Ling, research assistant on the NSW Space Research Network project, together with several students on supervised thesis projects covering manoeuvre modelling, track association and object characterisation.

Demonstrated result

Competitively funded as a NSW Space Research Network Pilot Research Project for 2025–2026, Pattern of Life Identification for Geosynchronous Satellites Using Transformer-Based AI Foundation Models, in collaboration with Macquarie University and the University of Adelaide.
A dedicated research position stood up to deliver the GEO pattern-of-life characterisation work.
Method development presented externally, including transformers and deep learning for satellite manoeuvre detection at the UNSW AI Symposium (slides).
Complementary manoeuvre-modelling work on factor graph optimisation for orbit determination with Gaussian approximation of impulsive manoeuvres, developed as a supervised thesis project.

External use

The project is co-delivered with Macquarie University under NSW Space Research Network funding, and it feeds the NSW Space Research Network white paper on Space Domain Awareness.

Next partnership opportunity

A manoeuvre and pattern-of-life assessment on your GEO or LEO time series: a labelled event list over the period you care about, and detection performance benchmarked against your current baseline, with false-alarm behaviour stated openly. See Partner with us.

Debris Family Classification

yang.yang16@unsw.edu.au (Dr Yang Yang) — Wed, 10 Dec 2025 00:00:00 +0000

Problem

When a satellite breaks up, the fragments enter the catalogue as hundreds of unattributed objects. Reconnecting them to their parent event matters for attribution, for understanding how the breakup happened, and for predicting where the cloud goes next.

Machine-learning methods using proper elements have made progress on this, but they inherit a quiet failure mode: a model trained on an outdated representation of the debris environment degrades as the population evolves around it. There is also a specific technical trap — normalising quaternion-set features for a neural network destroys the orbital size information the classifier needs, and the loss is invisible until accuracy is measured.

Capability

A computational pipeline that generates synthetic fragmentation data from explosive breakup events using a Standard Breakup Model, propagates it under a high-fidelity dynamical model, and extracts proper elements across three representations — modified equinoctial (MEE), Poincaré (PNC) and quaternion (QTN) sets. Neural networks are then trained on combinations of those element sets to decide whether a pair of fragments shares a parent.

Extending beyond the modified-equinoctial space used by previous approaches widens the dynamical fingerprint available to the classifier. The pipeline also includes an augmented quaternion representation, QTNp, which explicitly restores the semi-latus rectum lost during feature normalisation.

Demonstrated result

ROC-AUC improved from 0.789 to 0.858 in synthetic Starlink-like LEO experiments, comparing the joint MEE + PNC + QTN feature set against the MEE-only baseline, with corresponding gains in accuracy and F1.
The QTNp augmentation lifted quaternion-set accuracy from 0.31 to 0.60, by restoring the orbital size information that standard normalisation discards — identifying and closing a failure mode that had not previously been characterised.
Published as a preprint, arXiv:2512.08495, led by undergraduate researcher Michael Ling, and presented at the Australian Space Research Conference 2025 (slides).

External use

The method addresses a recognised gap in space sustainability and space domain awareness: maintaining attribution of fragmentation debris as the circumterrestrial environment evolves. The preprint is openly available, and the finding about quaternion-set normalisation applies to any classifier built on that representation, not only to this pipeline.

Next partnership opportunity

Applying the classifier to a real breakup event of interest to you, or extending it from synthetic training data to your catalogue. This fits most naturally as a sponsored thesis or co-funded project — see Partner with us.

Supervised Classification of LEO Debris Families Using Multi-Set Proper Elements

yang.yang16@unsw.edu.au (Dr Yang Yang) — Wed, 10 Dec 2025 00:00:00 +0000

We are excited to announce a new preprint led by our undergraduate student Michael Ling:

Title: Supervised Classification of LEO Debris Families Using Multi-Set Proper Elements Authors: Michael Ling and Yang Yang arXiv: https://arxiv.org/abs/2512.08495

Abstract: Machine learning techniques using proper elements to reconnect families of satellite fragmentation debris have recently advanced, becoming key to space sustainability and domain awareness. However, an evolving circumterrestrial environment may limit their applicability, particularly when models are trained on outdated debris representations. In this work, we devise a computational pipeline using synthetic fragmentation data from explosive breakup events, generated via a Standard Breakup Model and propagated under a high-fidelity dynamical model. Proper elements are extracted using adapted algorithms for modified equinoctial (MEE), Poincaré (PNC), and quaternion (QTN) sets. Extending beyond previous approaches limited to MEE space, we include PNC and QTN sets to broaden the dynamical fingerprints available to the classifier. Neural networks trained on various element combinations are used to determine if fragment pairs share a parent. Crucially, we identify a fundamental limitation when applying standard quaternion sets to neural networks: the loss of orbital size information during feature normalization. We introduce an augmented representation (QTNp) that explicitly restores the semi-latus rectum, improving accuracy from 0.31 to 0.60 compared to the standard set. In synthetic Starlink-like LEO experiments, expanding proper-element sets generally improves discrimination. The best model, using a joint feature set (MEE + PNC + QTN), achieves an ROC-AUC of 0.858 compared to 0.789 for the MEE-only baseline, alongside higher accuracy and F1 scores.

Congratulations to Michael and the team for this achievement!