Game Theoretic Training Enabled Deep Learning Solutions for Rapid Discovery of Satellite Behaviors

Dan Shen, Carolyn Sheaff, Genshe Chen, Jingyang Lu, Mengqing Guo, Erik Blasch, Khanh PhamIntechOpen 2020 · DOI · OpenAlex

Abstract

This research presents a game theoretic training model enabling a deep learning solution for the rapid discovery of satellite behaviors from collected sensor data. The solution, termed GTEL (Game Theoretic Training Enabled Deep Learning), consists of two primary components: a Pursuit-Evasion (PE) Markov game model for data augmentation and a convolutional neural network (CNN) for satellite behavior classification. Under the PE game theoretic framework, various satellite behaviors are simulated to generate synthetic datasets with labels for training. Simulation results demonstrate that the trained model can efficiently and correctly classify satellite behaviors with up to 99.8% accuracy.

Introduction

As space becomes increasingly crowded and contested, continuous Space Situational Awareness (SSA) is vital. Traditional tracking algorithms treat satellite movement as an optimal control problem, assuming objects follow predictable orbital mechanics or exhibit random, noisy deviations. However, this approach fails when dealing with uncooperative or intelligent Resident Space Objects (RSOs) that deliberately maneuver to evade tracking, confuse sensors, or obscure their intentions.

The core challenge addressed is the "zero-shot" problem: training a machine learning model to rapidly detect and classify adversarial satellite maneuvers when real-world training data for such rare or classified events is practically non-existent. This work bridges the gap by using game theory to generate high-fidelity synthetic training data.

Methodology

The proposed GTEL framework integrates several advanced computational techniques into a cohesive pipeline for behavior discovery.

System architecture of the GTEL framework, illustrating the flow from surveillance assets to behavior classification.

1. Markov Game Engine (Data Generation)

Instead of standard orbital propagation, the interaction between an observer satellite and a target satellite is modeled as a two-player Pursuit-Evasion (PE) Markov game. The target satellite intelligently maneuvers to maximize tracking uncertainty (entropy), while the observer manages its sensors to minimize it. This game acts as a sophisticated data augmentation engine, simulating realistic, adversarial evasion tactics.

State transition diagram for the Markov game, showing probabilities p1 and p2 for state changes.

2. CNN-Based Classification

The simulated sensor data is fed into a custom CNN. The network processes inputs shaped as three parameters (azimuth, elevation, range) over 15 time steps. The architecture utilizes multiple 2D convolutional layers, max pooling, and dropout layers (30%) to prevent overfitting. The final layer uses a softmax activation to classify maneuvers into one of 143 distinct directional labels, binned in 15-degree increments.

3. Dataset and Augmentation

Base orbital data was sourced from real Two-Line Elements (TLEs) via space-track.org. The Markov game engine propagated these orbits and injected specific maneuvers (e.g., increasing or decreasing orbital energy). The resulting dataset contains 57,332 tracks for training and 6,371 tracks for testing, each consisting of 15 sequential observation measurements relative to a ground site.

Key Results and Insights

  • High Classification Accuracy: The CNN model achieved 98% training accuracy and 96% validation/test accuracy across 143 distinct satellite behavior classes. The confusion matrix showed a near-perfect diagonal, indicating the model rarely confused adjacent maneuver trajectories.
  • Sensor Management Dynamics: The game model highlights a resource-management dynamic where the observer satellite must intelligently toggle its sensors (on/off control) to maximize information gain without wasting limited power or bandwidth.
  • Filter Interpretability: Visualization of weight variations in the CNN filters revealed that middle dense layers were primarily responsible for capturing the "anomaly" features of maneuvers, providing a degree of interpretability to the model.

Conclusion

The GTEL framework demonstrates that game theory is a highly effective tool for data augmentation in SSA, overcoming the lack of real-world adversarial data. By forcing the model to learn from intelligent evasion tactics rather than random perturbations, the system achieves high accuracy in classifying complex behaviors. Future work will explore multi-player game theory and the integration of blockchain technology to secure communications between distributed tracking assets.

Summary updated

Abstract

The chapter presents a game theoretic training model enabling a deep learning solution for rapid discovery of satellite behaviors from collected sensor data. The solution has two parts, namely, Part 1 and Part 2. Part 1 is a PE game model that enables data augmentation method, and Part 2 uses convolutional neural networks (CNNs) for satellite behavior classification. The sensor data are propagated with the various maneuver strategies from the proposed space game models. Under the PE game theoretic framework, various satellite behaviors are simulated to generate synthetic datasets with labels for the training to detect space object behaviors. To evaluate the performance of the proposed PE model, a CNN model is designed and implemented for satellite behavior classification. Python 3 and TensorFlow are used in this implementation. The simulation results show that the trained machine learning model can efficiently and correctly classify the satellite behaviors up to 99.8%.
BibTeX
@inproceedings{shen2020game,
  author = {Dan Shen and Carolyn Sheaff and Genshe Chen and Jingyang Lu and Mengqing Guo and Erik Blasch and Khanh Pham},
  title = {{Game Theoretic Training Enabled Deep Learning Solutions for Rapid Discovery of Satellite Behaviors}},
  booktitle = {IntechOpen},
  year = {2020},
  doi = {10.5772/intechopen.92636},
}
Figure