Starlink: Collision risks

[FutureSpaceTourist]

https://twitter.com/profhughlewis/status/1569986222294024192

Latest analysis for #Starlink & #OneWeb shows these two constellations accounted for 42% of all close approaches within 5 km predicted by #SOCRATES at the end of August, with Starlink alone accounting for 29%. [1/n]

On average, #SOCRATES predicts that each #Starlink satellite will now experience 1 close approach within 5 km with a non-Starlink object every day, and each #OneWeb satellite will experience 3.4 close approaches with a non-OneWeb object every day. These rates are increasing [2/n]

Here's the same data from [2/n] plotted with respect to the number of satellites in each constellation in orbit, clearly showing #SOCRATES predicts that #OneWeb satellites experience more close approaches (within 5 km) per satellite than the #Starlink satellites [3/n]

Focusing on the #SOCRATES predictions for close approaches with max. collision probability of at least 1E-5, we see that the two constellations accounted for 30% of all such close approaches in August, with #Starlink alone accounting for 20%. [4/n]

We see a rising trend through time in predictions of the average number of such close approaches being experienced by each satellite on a daily basis, with #OneWeb satellites seeing a rate that is now 5 times greater than the #Starlink satellites. at 0.05 per sat per day [5/n]

Here's the equivalent #SOCRATES data showing the average number of conjunctions with max. collision probability of at least 1E-5 per satellite per day [6/n]

Across all of the predicted conjunctions within 5 km for #Starlink just under half (47%) involve a debris object, one-quarter (23%) involve another non-Starlink payload & just over one-fifth (22%) involve another Starlink satellite (likely 'ignored' by the operator) [7/n]

For #OneWeb, the picture is very different. Just under 70% of all the conjunctions involve a non-OneWeb payload with only one-quarter (26%) involving a debris object. [8/n]

Hence the predictions from #SOCRATES suggest that #OneWeb has experienced about 33% more conjunctions with other payloads than #Starlink despite the constellations substantially smaller size. This is likely due to the relatively long orbit raising process through LEO. [9/n]

Taking a snapshot (from a #SOCRATES report generated on 12 September) the close approaches within 5 km involving #Starlink and #OneWeb can be seen clearly, with the #Starlink shells around 550 km & the OneWeb shells around 1200 km dominating. [10/n]

Here's the same #SOCRATES data plotted using a logarithmic y-axis, which distorts but enables a little more clarity at the lower counts. [11/n]

And this is the same data now shown as the proportion of all close approaches at each altitude. #Starlink shells around 350 km & 550 km, and #OneWeb shells between 1100 km & 1200 km become readily apparent, as do close approaches occurring during orbit raising. [12/n]

My thanks to @TSKelso (@CelesTrak) for the #SOCRATES data & support, and to @planet4589 for #Starlink & #OneWeb summary statistics. Hope this thread has been useful! [13/13; fin]

[steveleach]

How close does Starlink's collision avoidance let things get before they take action? If it is less than 5km then those numbers are all meaningless, presumably.

[steveleach]

Also, I noticed that the post at the start of this thread used a threshold of 1km. Does anyone know the reason for the switch to a 5km threshold? Is it just that 1km no longer paints a scary enough picture?

[jimvela]

A separation distance isn't super useful without an associated uncertainty... 
I wonder how good the orbit determination is that SpaceX is using for its automated responses.

[OceanCat]

A separation distance isn't super useful without an associated uncertainty... 
I wonder how good the orbit determination is that SpaceX is using for its automated responses.

They provide position and velocity covariance if you are curious. Instructions here.

How close does Starlink's collision avoidance let things get before they take action?

It's based on probability of collision not distance threshold. Greater than 1 in 100,000 triggers collision avoidance maneuver.

[niwax]

A separation distance isn't super useful without an associated uncertainty... 
I wonder how good the orbit determination is that SpaceX is using for its automated responses.

The ephemeris SpaceX hands out to sources like Celestrak is accurate to within about 50m. When you have proper tracking and information sharing, you can fit a hell of a lot of satellites up there.

[Robotbeat]

A separation distance isn't super useful without an associated uncertainty... 
I wonder how good the orbit determination is that SpaceX is using for its automated responses.

The ephemeris SpaceX hands out to sources like Celestrak is accurate to within about 50m. When you have proper tracking and information sharing, you can fit a hell of a lot of satellites up there.

Exactly this. And when you DO have a “conjunction,” you only require a very small movement to move out of the way.

I think it might be a good idea, as LEO fills up, to have like an ADS-B-type requirement for satellites. If everything’s position (and orientation) is known precisely down to less than a meter, conjunctions will be incredibly rare and you can avoid collisions very easily (even just changing both satellites’ orientation would often be enough, and then maybe move a couple meters to either side). Objects which lose ADS-B-whatever position transmitting capability could be swept out by deorbit tugs (with position tracked by dedicated radar beams in the meantime).

There’s no reason you couldn’t have literally millions of satellites in LEO with sufficient precision and maneuverability. It’s almost entirely empty space up there. When we talk about collision avoidance, we’re really talking about avoiding the tracking error cones, not the physical object (necessarily). The objects are almost point-like in comparison.

[steveleach]

How close does Starlink's collision avoidance let things get before they take action?

It's based on probability of collision not distance threshold. Greater than 1 in 100,000 triggers collision avoidance maneuver.

Yeah, I know that, but the quoted headline was about distance, so I'm interested in what sort of separation distances that might result in.

Assuming randomly passing through a circle of space, with a 5m radius target in the centre, I think you'd need the circle to be about 1.5k radius or less to have have a 1 in 100,000 chance of collision. A 5km radius gives 1 in a million.

So does that mean that 90% of passes within 5km would not trigger Starlink collision avoidance? And when avoidance is triggered, it still wouldn't need to adjust to pass more than 5k away.

Or are my assumptions or calculations messed up?

[OceanCat]

Yeah, I know that, but the quoted headline was about distance, so I'm interested in what sort of separation distances that might result in.

Assuming randomly passing through a circle of space, with a 5m radius target in the centre, I think you'd need the circle to be about 1.5k radius or less to have have a 1 in 100,000 chance of collision. A 5km radius gives 1 in a million.

So does that mean that 90% of passes within 5km would not trigger Starlink collision avoidance? And when avoidance is triggered, it still wouldn't need to adjust to pass more than 5k away.

Or are my assumptions or calculations messed up?

The modelling is more sophisticated than randomly passing through a circle of space. It's more like normal distribution within an elongated 3d ellipsoid. An ellipsoid around a Starlink satellite must be much smaller than the ellipsoids around two debris pictured below thanks to onboard GPS.

According to the data Prof. Lewis posted earlier 97% of conjunctions within 5 km and 50% within 1 km didn't trigger a CAM.

[steveleach]

Yeah, I know that, but the quoted headline was about distance, so I'm interested in what sort of separation distances that might result in.

Assuming randomly passing through a circle of space, with a 5m radius target in the centre, I think you'd need the circle to be about 1.5k radius or less to have have a 1 in 100,000 chance of collision. A 5km radius gives 1 in a million.

So does that mean that 90% of passes within 5km would not trigger Starlink collision avoidance? And when avoidance is triggered, it still wouldn't need to adjust to pass more than 5k away.

Or are my assumptions or calculations messed up?

The modelling is more sophisticated than randomly passing through a circle of space. It's more like normal distribution within an elongated 3d ellipsoid. An ellipsoid around a Starlink satellite must be much smaller than the ellipsoids around two debris pictured below thanks to onboard GPS.

According to the data Prof. Lewis posted earlier 97% of conjunctions within 5 km and 50% within 1 km didn't trigger a CAM.

Yep, I get that it isn't that simple, but I wanted a ballpark.  Where did you see that 97% figure?

Also, would that mean that out of those 30,000 Starlink conjunctions, roughly 900 resulted in the Starlink taking avoiding action? And would they still be likely to pass within 5k, but at a much lower probability of collision, if they did? And that 29,100 times the Starlinks didn't even bother adjusting?

[OceanCat]

Yeah, I know that, but the quoted headline was about distance, so I'm interested in what sort of separation distances that might result in.

Assuming randomly passing through a circle of space, with a 5m radius target in the centre, I think you'd need the circle to be about 1.5k radius or less to have have a 1 in 100,000 chance of collision. A 5km radius gives 1 in a million.

So does that mean that 90% of passes within 5km would not trigger Starlink collision avoidance? And when avoidance is triggered, it still wouldn't need to adjust to pass more than 5k away.

Or are my assumptions or calculations messed up?

The modelling is more sophisticated than randomly passing through a circle of space. It's more like normal distribution within an elongated 3d ellipsoid. An ellipsoid around a Starlink satellite must be much smaller than the ellipsoids around two debris pictured below thanks to onboard GPS.

According to the data Prof. Lewis posted earlier 97% of conjunctions within 5 km and 50% within 1 km didn't trigger a CAM.

Yep, I get that it isn't that simple, but I wanted a ballpark.  Where did you see that 97% figure?

Also, would that mean that out of those 30,000 Starlink conjunctions, roughly 900 resulted in the Starlink taking avoiding action? And would they still be likely to pass within 5k, but at a much lower probability of collision, if they did? And that 29,100 times the Starlinks didn't even bother adjusting?

There are three graphs in the linked thread showing 97,080 conjunctions within 5 km,  5,794 conjunctions within 1 km, and 2,884 conjunctions with 1 in 100,000 probability of collision. I derived the percentage from those numbers. At 19th tweet he shows that SOCRATES data allows estimating the number of Starlink CAMs fairly accurately.

Yes to all three other questions.

[meekGee]

A separation distance isn't super useful without an associated uncertainty... 
I wonder how good the orbit determination is that SpaceX is using for its automated responses.

The ephemeris SpaceX hands out to sources like Celestrak is accurate to within about 50m. When you have proper tracking and information sharing, you can fit a hell of a lot of satellites up there.

Exactly this. And when you DO have a “conjunction,” you only require a very small movement to move out of the way.

I think it might be a good idea, as LEO fills up, to have like an ADS-B-type requirement for satellites. If everything’s position (and orientation) is known precisely down to less than a meter, conjunctions will be incredibly rare and you can avoid collisions very easily (even just changing both satellites’ orientation would often be enough, and then maybe move a couple meters to either side). Objects which lose ADS-B-whatever position transmitting capability could be swept out by deorbit tugs (with position tracked by dedicated radar beams in the meantime).

There’s no reason you couldn’t have literally millions of satellites in LEO with sufficient precision and maneuverability. It’s almost entirely empty space up there. When we talk about collision avoidance, we’re really talking about avoiding the tracking error cones, not the physical object (necessarily). The objects are almost point-like in comparison.

If everyone is playing constructively.  Geo-politics suck, and you may need to be really vigilant about other satellites.  I wouldn't trust a cooperative system without an additional passive tracking system.

[steveleach]

Yeah, I know that, but the quoted headline was about distance, so I'm interested in what sort of separation distances that might result in.

Assuming randomly passing through a circle of space, with a 5m radius target in the centre, I think you'd need the circle to be about 1.5k radius or less to have have a 1 in 100,000 chance of collision. A 5km radius gives 1 in a million.

So does that mean that 90% of passes within 5km would not trigger Starlink collision avoidance? And when avoidance is triggered, it still wouldn't need to adjust to pass more than 5k away.

Or are my assumptions or calculations messed up?

The modelling is more sophisticated than randomly passing through a circle of space. It's more like normal distribution within an elongated 3d ellipsoid. An ellipsoid around a Starlink satellite must be much smaller than the ellipsoids around two debris pictured below thanks to onboard GPS.

According to the data Prof. Lewis posted earlier 97% of conjunctions within 5 km and 50% within 1 km didn't trigger a CAM.

Yep, I get that it isn't that simple, but I wanted a ballpark.  Where did you see that 97% figure?

Also, would that mean that out of those 30,000 Starlink conjunctions, roughly 900 resulted in the Starlink taking avoiding action? And would they still be likely to pass within 5k, but at a much lower probability of collision, if they did? And that 29,100 times the Starlinks didn't even bother adjusting?

There are three graphs in the linked thread showing 97,080 conjunctions within 5 km,  5,794 conjunctions within 1 km, and 2,884 conjunctions with 1 in 100,000 probability of collision. I derived the percentage from those numbers. At 19th tweet he shows that SOCRATES data allows estimating the number of Starlink CAMs fairly accurately.

Yes to all three other questions.

Thanks for that.

This does all make it sound like the "30,000 conjunctions within 5km" comment is just clickbait, but maybe I'm missing something.

[MP99]

I also wonder if SpaceX will start investing in their own tracking capabilities for space debris in order to reduce the uncertainty and reduce the need for maneuvering.

Perhaps as a natural extension to this, what if each Starlink sat had situational awareness RADAR / LIDAR over short distances. To throw a number out there to start a conversation - say 30km, which would be a few seconds of tracking?

Each sat would report any pings, including return time-of-flight, return brightness and tracking of rate-of-direction-change. Maybe increase the broadcast power once a target is identified, to get the best/longest tracking of the object as it recedes. I'd assume these would be downlinked for analysis. Perhaps the data could be forwarded raw to reduce onboard analysis.

If there is a collision, the rest of the constellation could rapidly identify any new items of debris, which could be a big improvement over the rate that they can be catalogued, and orbital elements nailed down.

It could also decrease the uncertainty of position of known items after a CAM, or an approach which doesn't need a CAM. Maybe the sensitivity could be increased when a close approach is expected. The system could be turned on only when an approach is expected, but that would stop it identifying smaller items that aren't currently tracked, which could be a valuable extension of existing tracking.


I also wonder about whether a Starlink should respond if something is going to impact despite all CAMs. Could the Starlink somehow perform a pre-emptive "destruction" (analogous to a launch destruct), which cleanly separates itself into two mostly-intact halves, specifically to maximise the eccentricity of those two halves.

The intention would be to minimise the perigee of the two Starlink debris clouds, which would minimise the on orbit lifetime. Think of it like an airbag which accepts the inevitability of the crash, but the intent is "save the on-orbit environment", rather than "save the passengers".

How much might this reduce the lifetime of the debris? I'd also appreciate any thoughts what would happen if the impactor hits one of the halves after Separation, IE the separation fails to avoid the collision?


I believe that the ideal dV for the separation would be prograde and retrograde to minimise perigee, but that would seem to threaten the leading and trailing sats in the same shell. Thoughts?

I would also expect that any close approach would be sent out via all available ISL as an emergency priority in real time, which would provide a real time telemetry "black box" in case of a collision. Should also set initial estimates of the energy and direction of any debris.

Any feedback appreciated.

[Asteroza]

I also wonder if SpaceX will start investing in their own tracking capabilities for space debris in order to reduce the uncertainty and reduce the need for maneuvering.

Perhaps as a natural extension to this, what if each Starlink sat had situational awareness RADAR / LIDAR over short distances. To throw a number out there to start a conversation - say 30km, which would be a few seconds of tracking?

Each sat would report any pings, including return time-of-flight, return brightness and tracking of rate-of-direction-change. Maybe increase the broadcast power once a target is identified, to get the best/longest tracking of the object as it recedes. I'd assume these would be downlinked for analysis. Perhaps the data could be forwarded raw to reduce onboard analysis.

If there is a collision, the rest of the constellation could rapidly identify any new items of debris, which could be a big improvement over the rate that they can be catalogued, and orbital elements nailed down.

It could also decrease the uncertainty of position of known items after a CAM, or an approach which doesn't need a CAM. Maybe the sensitivity could be increased when a close approach is expected. The system could be turned on only when an approach is expected, but that would stop it identifying smaller items that aren't currently tracked, which could be a valuable extension of existing tracking.


I also wonder about whether a Starlink should respond if something is going to impact despite all CAMs. Could the Starlink somehow perform a pre-emptive "destruction" (analogous to a launch destruct), which cleanly separates itself into two mostly-intact halves, specifically to maximise the eccentricity of those two halves.

The intention would be to minimise the perigee of the two Starlink debris clouds, which would minimise the on orbit lifetime. Think of it like an airbag which accepts the inevitability of the crash, but the intent is "save the on-orbit environment", rather than "save the passengers".

How much might this reduce the lifetime of the debris? I'd also appreciate any thoughts what would happen if the impactor hits one of the halves after Separation, IE the separation fails to avoid the collision?


I believe that the ideal dV for the separation would be prograde and retrograde to minimise perigee, but that would seem to threaten the leading and trailing sats in the same shell. Thoughts?

I would also expect that any close approach would be sent out via all available ISL as an emergency priority in real time, which would provide a real time telemetry "black box" in case of a collision. Should also set initial estimates of the energy and direction of any debris.

Any feedback appreciated.

Dragon's LIDAR might work, but is added mass.

I do wonder about debris passing through the phased array spot beams. The array isn't much different from an AESA radar in principle, but what size object at about what range needs to monitored? The array isn't that powerful, and 3cm objects are the low end of the debris tracking catalogs currently.

[Robotbeat]

The radar can get very close to debris, though. Instead of 1000km, you might get 10km or even 1km distances, and because beam divergence to target and then back actually goes as 1/r^4, that can mean a HUGE difference.

There might be some way to use the same phased array as a radar, at times when there is little demand on a satellite’s RF capacity, like when orbiting over the Pacific.

[DanClemmensen]

The radar can get very close to debris, though. Instead of 1000km, you might get 10km or even 1km distances, and because beam divergence to target and then back actually goes as 1/r^4, that can mean a HUGE difference.

There might be some way to use the same phased array as a radar, at times when there is little demand on a satellite’s RF capacity, like when orbiting over the Pacific.

LEO orbits are roughly 90 minutes for roughly 40,000 km, or about 7.4 km/s, and satellites can be coming from any direction, so closing velocities are from 0 to about 14 km/s   Your 30 km radar gives you about 2 seconds to avoid the head-on collision. Head-ons can occur for polar orbits. Your satellites won't save themselves, but they might collectively generate a collaborative database of extremely precise debris orbits to save each other.

[Robotbeat]

The radar can get very close to debris, though. Instead of 1000km, you might get 10km or even 1km distances, and because beam divergence to target and then back actually goes as 1/r^4, that can mean a HUGE difference.

There might be some way to use the same phased array as a radar, at times when there is little demand on a satellite’s RF capacity, like when orbiting over the Pacific.

LEO orbits are roughly 90 minutes for roughly 40,000 km, or about 7.4 km/s, and satellites can be coming from any direction, so closing velocities are from 0 to about 14 km/s   Your 30 km radar gives you about 2 seconds to avoid the head-on collision. Head-ons can occur for polar orbits. Your satellites won't save themselves, but they might collectively generate a collaborative database of extremely precise debris orbits to save each other.

That's right, I wasn't thinking of realtime avoidance. (It's fun to imagine what would be required for a 2 second avoidance to succeed, though.

[MP99]

The radar can get very close to debris, though. Instead of 1000km, you might get 10km or even 1km distances, and because beam divergence to target and then back actually goes as 1/r^4, that can mean a HUGE difference.

There might be some way to use the same phased array as a radar, at times when there is little demand on a satellite’s RF capacity, like when orbiting over the Pacific.

LEO orbits are roughly 90 minutes for roughly 40,000 km, or about 7.4 km/s, and satellites can be coming from any direction, so closing velocities are from 0 to about 14 km/s   Your 30 km radar gives you about 2 seconds to avoid the head-on collision. Head-ons can occur for polar orbits. Your satellites won't save themselves, but they might collectively generate a collaborative database of extremely precise debris orbits to save each other.

That was exactly my intention. Presumably could also catalogue items that are smaller than the current capability.

BTW, the 30km was just a wild number to start a conversation. Trying to balance functionality without making the orbital environment too "noisy" - something I know nothing about.

In terms of mapping objects as they pass, you might only have two seconds as it approaches, but perhaps a longer time as it recedes if you briefly boost the power. That's for new object detections. For known objects, could perhaps boost power throughout both the approach and after.

Edit: also thanks to RB for the 1/r^4 info.

[MP99]

The radar can get very close to debris, though. Instead of 1000km, you might get 10km or even 1km distances, and because beam divergence to target and then back actually goes as 1/r^4, that can mean a HUGE difference.

There might be some way to use the same phased array as a radar, at times when there is little demand on a satellite’s RF capacity, like when orbiting over the Pacific.

LEO orbits are roughly 90 minutes for roughly 40,000 km, or about 7.4 km/s, and satellites can be coming from any direction, so closing velocities are from 0 to about 14 km/s   Your 30 km radar gives you about 2 seconds to avoid the head-on collision. Head-ons can occur for polar orbits. Your satellites won't save themselves, but they might collectively generate a collaborative database of extremely precise debris orbits to save each other.

That's right, I wasn't thinking of realtime avoidance. (It's fun to imagine what would be required for a 2 second avoidance to succeed, though.

Any thoughts on sacrificially destroying itself in as clean a way as possible if a collision object is detected? Would be from previously unmapped debris, or perhaps a sat which is functioning except for propulsion, if that is a thing?