A Self-Supervised Framework for Space Object Behaviour Characterisation

TL;DR

Space object behaviour analysis is essential as orbital populations grow. The authors present a Space Safety and Sustainability Foundation Model built on a self-supervised Perceiver-VAE trained on 227k light curves to learn rich latent representations, enabling anomaly detection, motion prediction, and synthetic data generation. The approach achieves 88% anomaly-detection accuracy (ROC AUC ~0.90) and 82.7% motion-classification accuracy (ROC AUC ~0.95), with effective forecasting and scalable data synthesis via reference-based latent sampling. This work demonstrates the potential of foundation-model-like frameworks for SDA and outlines directions for benchmarks, physics-informed validation, privacy considerations, and multimodal data integration.

Abstract

Foundation Models, pre-trained on large unlabelled datasets before task-specific fine-tuning, are increasingly being applied to specialised domains. Recent examples include ClimaX for climate and Clay for satellite Earth observation, but a Foundation Model for Space Object Behavioural Analysis has not yet been developed. As orbital populations grow, automated methods for characterising space object behaviour are crucial for space safety. We present a Space Safety and Sustainability Foundation Model focusing on space object behavioural analysis using light curves (LCs). We implemented a Perceiver-Variational Autoencoder (VAE) architecture, pre-trained with self-supervised reconstruction and masked reconstruction on 227,000 LCs from the MMT-9 observatory. The VAE enables anomaly detection, motion prediction, and LC generation. We fine-tuned the model for anomaly detection & motion prediction using two independent LC simulators (CASSANDRA and GRIAL respectively), using CAD models of boxwing, Sentinel-3, SMOS, and Starlink platforms. Our pre-trained model achieved a reconstruction error of 0.01%, identifying potentially anomalous light curves through reconstruction difficulty. After fine-tuning, the model scored 88% and 82% accuracy, with 0.90 and 0.95 ROC AUC scores respectively in both anomaly detection and motion mode prediction (sun-pointing, spin, etc.). Analysis of high-confidence anomaly predictions on real data revealed distinct patterns including characteristic object profiles and satellite glinting. Here, we demonstrate how self-supervised learning can simultaneously enable anomaly detection, motion prediction, and synthetic data generation from rich representations learned in pre-training. Our work therefore supports space safety and sustainability through automated monitoring and simulation capabilities.

Figures (10)

• Figure 1: Graphical structure of the paper. First, in Section \ref{['results:pre-training-training']} we describe pre-training our Foundation Model with a large unlabelled dataset of real light curves. We encode these into a rich latent representation with a Perceiver-VAE architecture, updating these representations based on three self supervised learning (SSL) tasks: Reconstruction, Forecasting, and Masking. We next analyse the results of pre-training, which includes pre-flagging of anomalies based on reconstruction difficulty (Sections \ref{['sec:pretraining_preflagging']}-\ref{['sec:pre-flagged_anomalies_analysis']}, and forecasting quality of the model (\ref{['sec:forecasting']}. Following this, we descibe fine-tuning our rich representations for two downstream tasks: a) anomaly prediction (Section \ref{['sec:anomaly_detection_finetuning']}, and b) motion prediction (Section \ref{['sec:fine-tuning_motion']}). Finally, we demonstrate further utility of our representations by generating de novo datasets according to a particular motion type (Section \ref{['sec:synthetic_data_generation']}).
• Figure 2: Pre-training approach. Input array(s) (M) provide keys (K) which index the data (e.g., timestep in a timecourse) and values (V) which represent the information at each K. The model includes a learned latent array (N) which provides queries (Q) for the Cross Attention mechanism. Q interacts with K and V to extract relevant information from the input. The latent representation (z) is computed from this mechanism using the mean and log variance, capturing compressed features of the input. In this way, z acts as a latent bottleneck. Self Attention layers then learn meaningful relations within this latent space. This architecture incorporates elements of Variational Autoencoders (VAE) in its training process, whereby the loss is calculated by sampling probabilistically from the latent space. Figure adapted from jaegle_perceiver_2021.
• Figure 3: Training curves for a VAE-Perceiver model trained on the MMT-9 light curve dataset. Training and Validation losses decrease from >approximately 0.015 to 0.0011 and 0.0012 respectively. The KL Divergence of the latent distribution initially increases, before plateauing at approximately 0.85. Validation loss decreases to a similar plateau, without the initial first epoch decrease.
• Figure 4: The ten highest reconstruction error test light curves, where A is the highest, and J the tenth highest. These curves exhibit periodic variation over time in the measured standardised magnitude (K-T). As in A-K but for the ten lowest error test light curves, which exhibit mostly linear variation in the standardised magnitude over time. Some curves exhibit troughs in standardised magnitude (K, P, T).
• Figure 5: Test light curves with the latter 25% of the timecourse masked at inference. Our Perceiver-VAE was trained with several self-supervised tasks — including a forecasting loss (see Section \ref{['methods:pre-training_strategy']} for details). This means that at inference, the future state of a light curve can be predicted by the pre-trained model. Top row: the lowest three forecasting mean squared error test samples. Bottom row: the highest three forecasting mean squared error test samples. Qualitatively, the original signal (turquoise) and known values (cyan) are aligned. In the lowest error samples, the forecasted values also align well with the original signal, capturing the dynamics well. In the highest error samples, the general trends of the masked regions are successfully forecasted, but the magnitude of the peaks/troughs in the signal is not fully captured.