Multi-Tier Labeling and Physics-Informed Learning for Orbital Anomaly Detection at Scale

Researchers developed a multi-tier labeling system combining physics-based rules, Kalman filtering, and machine learning to detect orbital anomalies across thousands of LEO satellites. The approach generated 8.6M labeled training sequences from 232M historical records, enabling a Transformer model to achieve 55.4% maneuver recall and 62.8% decay recall—addressing a critical gap in space situational awareness infrastructure.

Orbital anomaly detection has become operationally critical as low-Earth orbit grows increasingly congested, yet the field lacks scalable solutions beyond manual review or overly conservative rule-based systems. This research tackles the fundamental bottleneck: the absence of ground-truth labeled data at scale. Rather than waiting for perfect labels, the authors construct a cascading weak supervision pipeline where simpler detectors (physics rules, Kalman filters) feed increasingly refined signal to a neural network, ultimately creating 430M timesteps of training data across 11 orbital features.

The innovation lies in treating labeling as a multi-stage process where each tier trades speed for accuracy. The IMM-UKF bank discovered 42.6x more anomalies than pure rule-based detection on overlapping satellites, demonstrating that intermediate complexity captures behavioral nuance the simplest systems miss. The Transformer architecture then learns to synthesize these weak signals into a high-recall triage classifier—explicitly designed not to produce final determinations but to surface candidates for human or downstream filtering.

For space operators and satellite fleet managers, this represents a pathway toward automating the resource-intensive conjunction screening and decay forecasting that currently depends on domain expertise and manual triage. The 107% relative improvement from adding time-delta features suggests temporal patterns are underexploited in existing operational systems. The authors' vision of a Neural-ODE-based orbital world model hints at future systems that could simulate orbital dynamics directly, reducing reliance on Two-Line Element propagation. This work establishes practical scaffolding for deploying machine learning to space domain awareness without requiring perfect ground truth.

• →Multi-tier weak supervision generated 8.6M labeled orbital sequences from 232M historical TLE records, solving the data scarcity bottleneck in space anomaly detection.
• →The IMM-UKF Kalman filter stage discovered 42x more anomalies than pure rule-based detection, proving intermediate complexity captures behavioral patterns simple rules miss.
• →The trained Transformer achieved 62.8% decay recall and 55.4% maneuver recall, positioned as a high-recall triage classifier rather than final attribution system.
• →Time-delta feature engineering alone yielded 107% relative improvement in decay recall, indicating temporal patterns are underutilized in operational space systems.
• →This approach provides a scalable blueprint for deploying neural networks to space situational awareness without waiting for perfect labeled data from manual review.