A Self-Supervised Framework for Space Object Behaviour Characterisation

This paper presents a self-supervised framework using a Perceiver-Variational Autoencoder (VAE) pre-trained on large-scale light curve data to enable automated space object behavior characterization, including anomaly detection, motion mode prediction, and synthetic data generation.

The Gist

The "Cosmic Detective" Framework: Watching the Stars to Protect Our Skies

Imagine you are standing in a massive, dark stadium filled with thousands of people. Most people are sitting quietly, but every now and then, someone starts dancing wildly, someone else stands up and walks in a strange pattern, and a few people start throwing glow sticks.

If you were trying to watch everyone at once, you’d quickly get overwhelmed. You wouldn't know if a person dancing is just having fun or if they’ve actually tripped and are in trouble.

This paper is about building an AI "Super-Observer" that can watch the "stadium" of space and instantly tell the difference between a satellite doing its job and a piece of space junk behaving dangerously.

---

1. The Problem: A Crowded Sky

Right now, we are launching more satellites than ever before. Space is getting crowded. To keep our GPS, weather reports, and internet working, we need to make sure these satellites don't crash into each other or turn into unpredictable "space debris."

Currently, humans have to look at data (called light curves, which are basically just measurements of how bright an object looks as it spins) to figure out what a satellite is doing. But there is too much data for humans to handle. We need an automated detective.

2. The Solution: The "Self-Taught" Detective (Self-Supervised Learning)

Most AI is like a student who needs a teacher to grade every single homework assignment. This is called "supervised learning." But in space, we don't have "answer keys" for everything—we don't know exactly what every weird flicker in a light curve means.

The researchers created a Foundation Model. Think of this like a student who spends years just watching the stadium. They don't have a teacher telling them "that's a dancer" or "that's a walker," but by watching millions of patterns, they learn the "rhythm" of the crowd.

They used a special architecture called a Perceiver-VAE.

• The Perceiver is like a high-speed camera that can focus on the most important movements without getting distracted by the background.
• The VAE is like a mental sketchpad. The AI tries to draw a picture of what it sees, compares it to reality, and learns from its mistakes.

3. The Three Superpowers of the AI

Once the AI "watched" 227,000 patterns from a real observatory, it developed three main skills:

• Skill 1: The Anomaly Detector (The "Something's Wrong" Alarm):
  Because the AI knows what "normal" looks like, it can spot something weird. If a satellite suddenly starts tumbling uncontrollably instead of spinning smoothly, the AI sees a "glitch" in its mental sketchpad and raises a red flag. It was 85% accurate at spotting these oddities.

• Skill 2: The Motion Predictor (The "Behavior Profiler"):
  The AI can look at a flicker and say, "Ah, that's a satellite pointing its solar panels at the sun," or "That's a piece of debris tumbling randomly." It was incredibly good at this, reaching a 95% accuracy score in identifying different types of movement.

• Skill 3: The Dreamer (Synthetic Data Generation):
  This is the coolest part. Because the AI understands the "rules" of how light curves look, it can actually imagine new ones. It can "dream up" fake but realistic data of a tumbling satellite. This is helpful because it allows us to train other AIs without needing to wait for a real satellite to crash to see what that looks like.

4. Why does this matter?

As we move toward a future where space is a vital part of our daily lives (for banking, navigation, and communication), we can't afford to have "blind spots."

This research is a first step toward a "Foundation Model for Space." Just as ChatGPT understands the patterns of human language, this model is learning the "language of light" in orbit. It’s building a digital eye that never sleeps, helping us keep the orbital highways safe, sustainable, and predictable.

Technical Summary

Technical Summary: A Self-Supervised Framework for Space Object Behaviour Characterisation

1. Problem Statement

The rapid proliferation of space objects (satellites, rocket bodies, and debris) poses significant risks to space safety, national security, and global economic infrastructure (e.g., GPS and telecommunications). Traditional methods for Space Object Behavioural Analysis (SOBA)—such as manual inspection or physics-based numerical modeling—are increasingly inadequate due to the massive volume of data generated by modern sensors and the high computational cost of complex simulations. There is a critical need for automated, scalable, and efficient methods to detect anomalies (unusual maneuvers or malfunctions) and characterize motion modes (e.g., tumbling vs. sun-pointing).

2. Methodology

The authors propose a first step toward a Foundation Model for SOBA using a self-supervised learning (SSL) approach. The framework is built on a Perceiver-Variational Autoencoder (VAE) architecture, chosen for its ability to handle long-range dependencies in sequential data and its inherent extensibility to multimodal inputs (e.g., radar, hyperspectral data).

• Architecture: The model uses a Perceiver backbone, which employs a cross-attention mechanism to map high-dimensional input sequences into a compressed latent bottleneck. This is integrated with a VAE to ensure a continuous and structured latent space.
• Pre-training (Self-Supervised): The model was pre-trained on a large, unlabelled dataset of 227,000 real light curves from the MMT-9 observatory. It utilized a multi-task SSL objective:
  1. Reconstruction: Learning to reconstruct the full input sequence.
  2. Masking: Predicting randomly masked segments to learn contextual relationships.
  3. Forecasting: Predicting future states of the light curve to capture temporal dynamics.
• Fine-tuning (Downstream Tasks): The pre-trained encoder was frozen, and lightweight MLP heads were trained on two distinct simulated datasets:
  • Anomaly Detection: Using the CASSANDRA simulator (800 simulations of various satellite platforms like Starlink and Sentinel-3) to identify debris collisions or signal changes.
  • Motion Prediction: Using the GRIAL simulator (22,009 simulated light curves) to classify six motion modes (e.g., Sun-oriented, Earth-oriented, Tumbling, Spin, etc.).
• Generative AI: The model was used to generate de novo synthetic light curves via "neighbourhood sampling" in the latent space, guided by reference curves to ensure physical plausibility.

3. Key Contributions

• First SOBA Foundation Model Framework: Demonstrates that self-supervised pre-training on unlabelled light curves creates rich representations applicable to multiple downstream tasks.
• Multimodal-Ready Architecture: By adopting the Perceiver architecture, the framework is designed to eventually integrate diverse sensor modalities beyond visible light curves.
• Label Efficiency: Proved that pre-training significantly improves performance in "low-data" regimes, which is critical in space domains where labelled anomalous data is rare.
• Synthetic Data Generation: Developed a method to generate statistically consistent synthetic light curves to augment training datasets for Space Domain Awareness (SDA).

4. Results

• Pre-training Performance: The model achieved a reconstruction MSE of 0.009. High reconstruction errors successfully flagged potentially anomalous real-world patterns (e.g., rapid tumbling).
• Anomaly Detection: After fine-tuning, the model achieved 85% accuracy and a 0.92 ROC AUC. It outperformed supervised baselines (CNN, LSTM) and showed superior performance in $k$-shot learning (e.g., outperforming CNNs even with only 2 labelled examples per class).
• Motion Prediction: The model achieved 82.7% accuracy and a 0.95 ROC AUC. It was particularly effective at distinguishing complex motions like "Tumbling" and "Spin."
• Generative Quality: Synthetic curves were validated using Gaussian Processes (GP); five out of six motion classes showed $>94\%$ of time steps falling within the 95% confidence interval of real data distributions.
• Efficiency: Magnitude pruning showed the Perceiver layers are highly redundant (tolerating 70% sparsity), and INT8 quantization reduced model size by 75% with negligible accuracy loss, facilitating edge deployment.

5. Significance

This work represents a paradigm shift from manual, task-specific analysis to an automated, scalable AI framework for space safety. By demonstrating that a single model can simultaneously detect anomalies, predict motion, and generate synthetic data, the authors provide a blueprint for Space Situational Awareness (SSA) systems. The ability to perform high-accuracy characterization with minimal labelled data is particularly vital for national security and commercial satellite operators who must monitor vast, rapidly growing orbital populations.