Polaris Project - Introduction
The aim of the Polaris project is to analyze a satellite set of telemetry to understand links/dependencies among different subsystems and between the spacecraft and its context. A data-driven analysis should be able to demonstrate the understanding of the links between the different behaviour changes of each telemetry within a satellite or within a set of external sources of information (mission plan, solar aspect angles, ephemerides, etc.). Machine learning can be used to learn the dependencies and correlations happening within a spacecraft. The acquired knowledge should be stored in a dependency graph (e.g. Bayesian network) in order to be able to compare changes in the future and reason on the graph when auto-diagnosing.
From discussions with the community we decided to split the polaris project in three parts:
- Polaris-Fetch: a tool to fetch data from telemetry databases (such as SatNOGS) and prepare context information for machine learning algorithms. Usually preparing regular/irregular timeseries.
- Polaris-Learn: Feature engineering and selection along with machine learning algorithms to create prediction models, anomaly detection models and many other knowledge and situation awareness techniques.
- Polaris-Viz: data visualization focused on understanding what ML algorithms have output and why (explainability). It is also where user's interactions are designed to be able to learn from operators (polaris users).
|Polaris Presentation at the Cubesat Developer's Workshop 2020 - by Hugh Brown|