Collision Avoidance Challenge (2019)
Deep Learning Approach
Comptetition organized by European Space Agency (ESA)
Today, active collision avoidance among orbiting satellites has become a routine task in space operations, relying on validated, accurate and timely space surveillance data. For a typical satellite in Low Earth Orbit, hundreds of alerts are issued every week corresponding to possible close encounters between a satellite and another space object (in the form of conjunction data messages CDMs). After automatic processing and filtering, there remain about 2 actionable alerts per spacecraft and week, requiring detailed follow-up by an analyst. On average, at the European Space Agency, more than one collision avoidance manoeuvre is performed per satellite and year.
In this challenge, you are tasked to build a model to predict the final collision risk estimate between a given satellite and a space object (e.g. another satellite, space debris, etc). To do so, you will have access to a database of real-world conjunction data messages (CDMs) carefully prepared at ESA.
Project developed with Sergio Pérez
Deep Learning Approaches
- Fully Connected Network
- Convolutional Neural Networks
- Recurrent Neural Networks
- Autoecoder Network
- Siamese Network (Manhattan LSTM)
