Automating Satellite Collision Avoidance
ESA[*] is preparing to use machine learning to protect satellites from the very real and growing danger of space debris.
The agency is developing a collision avoidance system that will automatically assess the risk and likelihood of in-space collisions, improve the decision making process on whether or not a maneuver is needed, and may even send the orders to at-risk satellites to get out of the way.
[...] "There is an urgent need for proper space traffic management, with clear communication protocols and more automation" says Holger Krag, Head of Space Safety at ESA.
"This is how air traffic control has worked for many decades, and now space operators need to get together to define automated maneuver coordination."
[...] Because of [the current] debris environment, it is now routine for operators in highly-trafficked orbits to spend time protecting their spacecraft from potentially catastrophic collisions with space junk, by performing "collision avoidance maneuvers"—basically sending the commands to their spacecraft to get out of the way.
Such maneuvers depend on validated, accurate and timely space surveillance data, provided for example by the US Space Surveillance Network, serving as the basis of "conjunction data messages," or CDMs, warning of possible close encounter between their spacecraft and another satellite or space object.
For a typical satellite in low-Earth orbit, hundreds of alerts are issued every week. For most, the risk of collision decreases as the week goes by and more orbital information is gathered, but for some the risk is deemed high enough that further action is required.
For ESA's current fleet of spacecraft in these low altitude orbits, about two alerts per week, per satellite, require detailed follow-up from by an analyst. This involves hours of analysis of the distance between the two objects, their likely positions in the future, uncertainties in observations and therefore in calculations and ultimately the probability of collision.
If the probability is greater than typically one in 10,000, the work of various teams is needed to prepare a collision avoidance maneuver and upload the commands to the satellite.
The maneuver must be verified to ensure it will have the expected effect, and doesn't for example bring the spacecraft closer to the object or even in harm's way of another object.
[...] Although such maneuvers ultimately protect spacecraft, they also disrupt their normal schedule, delaying or interrupting scientific observations or communications, and often use up scarce fuel, decreasing the lifetime of the mission.
The need for such avoidance maneuvers will likely increase greatly in the next few years. Not only due to huge communication constellations by SpaceX's Starllink and OneWeb, among others, but also from a burgeoning market for "smallsats" that either rideshare on a large rocket's launch or through companies like Rocket Lab which offer relatively inexpensive and frequent launches of small payloads.
[*] ESA European Space Agency.
(Score: -1, Troll) by Anonymous Coward on Thursday October 24 2019, @10:07AM
Starlink tweets were to get funding for their new AI system.
(Score: 0) by Anonymous Coward on Thursday October 24 2019, @12:50PM (1 child)
I don't think AI is the right direction.
Something more analysis based to be right by design seems more reasonable.
The basic operating principal for AI here is show it some example sequences, let it choose criteria for predicting how they came out, and hope for the best.
This could be used to describe weather prediction for the last century and we know how well that works. Not great, but better than anything else because it is a very complicated system.
The bar should be a bit higher for collision detection.
They published a contest to help in this that provided a glimpse of how they do collision avoidance.
The work from CDM's which are the odds or a particular pair of objects violating the 2 things in the same place at the same time rule.
That's a one to one intersection check.
Perhaps with these big constellations, they need to first look at the data behind the CDM's and be looking at where whole orbit trains full of objects intersect instead of single objects in orbits.
That might be something that could be done closed form for an estimated orbit pair.
That would be a many to many, or at least many to their one check.
(Score: 0) by Anonymous Coward on Thursday October 24 2019, @03:31PM
Agreed, AI isn't the right answer here, but too much is done manually that could be automated.
(Score: 2) by corey on Thursday October 24 2019, @08:48PM
For the Starlink space junk cloud.
(Score: 0) by Anonymous Coward on Friday October 25 2019, @03:12AM
I am just waiting for some small rogue nation to retaliate for sanctions by launching a ton of coarse sand into an elliptical polar orbit that intersects geosynchronous, leveling the playing field by messing things up for everyone.