Satellite Maneuver Detection Using Unsupervised Anomaly Models

The patent replaces rule-based mechanical detection methods with a machine learning-based automated system. The ML model automatically analyzes TLE data to detect maneuvers, eliminating the need for manual rule-based analysis and significantly improving detection speed while maintaining accuracy.

The system enables self-service detection where the ML model autonomously processes TLE data and identifies maneuvers without requiring continuous human intervention or manual rule updates. The model learns from historical data and automatically adapts to detect new maneuver patterns.

Instead of requiring labeled maneuver data to train the model, the patent inverts the approach by training the ML model on unlabeled TLE data using unsupervised learning techniques. The model learns normal orbital patterns and automatically identifies deviations as potential maneuvers, eliminating the need for scarce labeled training data.

The ML model performs self-training by automatically learning from the structure and patterns in unlabeled TLE data. Through unsupervised learning, the model develops its own understanding of normal orbital behavior and can independently identify anomalies without requiring external labeled datasets.

The patent extracts and prioritizes only the most relevant conjunctions for analysis by applying filtering criteria based on orbital parameters, satellite types, and proximity metrics. This extraction process separates critical conjunctions requiring detailed analysis from routine cases, enabling timely detection of threatening actions while maintaining reliability.

Instead of examining all conjunctions equally, the patent applies partial action by focusing computational resources on a subset of high-priority conjunctions identified through automated filtering. This approach processes the most critical cases in detail while using automated triage for the remainder, achieving timely detection without exhaustive analysis.