Deep Learning For Cubesat: Behavior Segmentation With Collection Of Contextual Information

The Polaris project used in-spacecraft data to learn and predict its behavior. The project aimed at adding external sources of data, such as orbit propagation, solar events, magnetic events, and other elements of space weather, to better predict the behavior of the spacecraft in any space environment.

This was achieved by creating modules to collect data from various sources, converting it to a time series and learning features from the converted data. Storage of the data in a DBMS (InfluxDB) was also added.

An auto-encoder based technique to detect anomalies in the space weather was also added. This module implements this research paper: https://arxiv.org/abs/1801.05394

Improvements to the visualization module to make it more user friendly and interactive was also done. Ability to train on any satellite telemetry (with/without normalization to SI units) was also added.