OrbitAI leverages machine learning to predict potential collisions between satellites and space debris, ensuring the safety and efficiency of space operations. By using real-time data, neural networks, and space simulators, this project aims to enhance satellite autonomy and improve space traffic management.
OrbitAI will rely on a variety of data sources to ensure accurate predictions:
- Satellite Orbital Data: Data on the position, velocity, and trajectory of satellites.
- Primary source for Satellite data: https://www.space-track.org/
- Space Debris Data: Information from space agencies and tracking stations about debris, including size, velocity, and orbital details.
- Primary source of Space Debris data: https://discosweb.esoc.esa.int/
- Environmental Factors: Data on solar radiation, atmospheric drag, and gravitational perturbations that affect satellite paths (May be implemented later).
To achieve accurate predictions, several types of neural networks will be employed:
- Recurrent Neural Networks (RNNs): For handling time-series data and predicting satellite trajectories over time based on past movement.
- Long Short Term Memory Model (LSTM): For handling the vanishing/exploding gradient problem that traditional RNNs face.
- Graph Neural Networks (GNNs): For modeling relationships between multiple satellites in a constellation, facilitating coordination (May be implemented later).
Training will focus on the following techniques:
- Supervised Learning: Using labeled data from past satellite movements and collision scenarios to train models on predicting future events.
- Reinforcement Learning: Applying RL techniques for autonomous satellite maneuvers, allowing satellites to make real-time decisions based on collision risk.
- Simulation-based Training: Creating simulated satellite orbits to augment the dataset, particularly for rare events like satellite collisions, to improve model robustness.
To validate the trained models and test real-time predictions, OrbitAI will integrate with space simulators like Orekit or GMAT (General Mission Analysis Tool). These simulators allow the virtual testing of satellite constellations and debris interactions, providing a controlled environment to evaluate the models' effectiveness.
- Simulations will also help refine maneuver strategies by allowing virtual trial-and-error of satellite responses to predicted collisions.
- CI/CD: OrbitAI Developers have automated many of the version control processes.
- When a developer pushes new code or makes a pull request, GitHub creates a fresh Ubuntu virtual machine instance, downloads all required dependencies onto the VM, and checks out the rest of the repository all to perform tests for safe integration.
