Deep Learning Approaches to Star Tracker Attitude Determination
Investigation and development of deep learning methodologies for satellite attitude determination using star trackers to overcome traditional limitations in challenging orbital conditions.
Requirements
- M.Sc. in Machine Learning, Data Science, Computer Science, Mathematics, Telecommunications, or similar
- Good knowledge of Python
- Software development skills
- Basic concepts of image processing
- Basic concepts of data science, concerning data analysis, data processing and deep learning
Description
Star trackers provide critical attitude determination for satellites, serving as the gold standard for spacecraft orientation systems with accuracy far exceeding other sensors. However, traditional star tracking algorithms face significant limitations. These conventional systems rely on geometric feature extraction and pattern matching techniques that perform poorly in challenging conditions such as image noise, partial star field visibility, and rapid attitude changes. Additionally, these algorithms demand substantial computational resources, restricting their use on smaller satellites with limited processing capabilities.
This thesis addresses how to enhance the robustness, accuracy, and efficiency of star tracker systems through deep learning approaches. The research will investigate whether neural networks can overcome traditional limitations while maintaining or improving performance across diverse operational scenarios encountered in space.
The work begins with an examination of conventional star tracking methodologies, including lost-in-space algorithms and sequential tracking methods, to understand their fundamental principles and inherent constraints. Simultaneously, the research will analyze existing deep learning applications in aerospace navigation to identify transferable techniques.
A crucial research component involves utilizing astronomical catalogs like Hipparcos and Gaia, which contain precise stellar positional data essential for creating comprehensive training datasets. The project will use open source synthetic star field generator that creates realistic images simulating various satellite orientations, star magnitudes, and partial occlusions typical in operational conditions. This addresses the fundamental challenge of limited labeled real-world data for training effective neural networks.
The neural network design phase will explore architectures optimized for the unique challenges of star pattern recognition, focusing on CNNs suited to sparse point-source imagery. Attention mechanisms will be investigated to help networks identify key star patterns amidst background noise. The ultimate goal is developing end-to-end networks capable of directly estimating attitude quaternions from raw star field images, potentially advancing research in how satellites determine their orientation in space.