Zachary Pollmiller, CFC; Christopher Irrgang, CFC; William Jantscher, CFC; Ryan Do, CFC
https://doi.org/10.64861/SWYV5238
Keywords: Satellite Characterization, AI for SDA, unsupervised machine learning, maneuver prediction, big data
Abstract:
In an increasingly congested and contested space domain, the ability to transition from reactive monitoring to proactive analysis is paramount for robust Space Domain Awareness (SDA). This research presents a novel machine learning framework designed to analyze a satellites pattern of life, predict future actions, and automatically detect anomalous behavior. The core goal is to furnish SDA analysts with predictive intelligence, enabling them to anticipate routine maneuvers and focus critical attention on unforeseen or potentially hostile activities.
The methodology employs a sophisticated two-stage approach for classifying and forecasting satellite maneuvers from telemetry data. Initially, an unsupervised Isolation Forest (IF) algorithm is leveraged to create a baseline of normalcy. This algorithm is particularly adept at parsing multi-modal data distributions, assessing both the maneuver’s delta-v magnitude and the temporal cadence between burns to distinguish between nominal station-keeping adjustments and true anomalies. The unsupervised nature of the IF model is a key advantage, as it negates the need for large, pre-labeled datasets. Following this anomaly detection phase, a Gaussian Mixture Model (GMM) is trained exclusively on the corpus of identified nominal maneuvers. This model generates a probabilistic forecast, predicting the timing of the next likely maneuver. Initial modeling has demonstrated significant success, forecasting maneuvers into 12-hour and 8-hour temporal bins with a notable degree of accuracy, providing analysts with a tiered window of predictive insight.
The expected outcome is a robust, automated system that significantly enhances the monitoring of high-priority Resident Space Objects (RSOs). By providing timely alerts for anomalous events and a reliable predictive look-ahead for routine operations, this framework promises to reduce manual assessment workloads and mitigate the cognitive burden on analysts. This capability directly contributes to a more transparent and predictable space environment. Ongoing and future work is focused on further enhancing the model’s predictive power. This includes a deep investigation into the Mamba state-space model to improve time-series forecasting, the application of Vandermonde smoothing to bolster input feature stability, and the integration of additional kinematic features to provide a richer, more descriptive dataset for the models.
Date of Conference: September 15-18, 2026
Track: AI for SDA Applications