Comment by swiftcoder
1 day ago
> But why would you transmit all this information to every train, instead of just computing the collisions in a central node?
Because central nodes can fail (as they did here, both human and machine), and providing this redundant data allows both the computer and the human-in-the-loop on the fast-moving chunk of steel to independently validate the decisions coming from the central node.
> You regularly pass other trains within <2m at high speeds, so linear offset is not all you need.
In this collision, both trains were on the same track. So while we obviously care about adjacency issues in the general case, this specific failure mode can be detected purely via a linear offset.
You can and do have redundancy in the central node already. Also that's not what failed here.
> In this collision, both trains were on the same track.
Yeah, but how do you distinguish that? So when I understand you correctly, what you want is: Find another moving chunk of metal in the radius of 2km, which moves toward you, but not in a direct or ballistic trajectory, but in a arbitrary complicated trajectory, but which can lookup in a database. Also you are operating in the real world, so there are potentially houses, trees and mountains in between. I guess that would be possible, but it would be really expensive. Your best bet would be to adapt a missile detection and targeting system from the military. A system which does exist and is somewhat similar is RCAS (https://www.dlr.de/en/kn/research-transfer/projects/next-gen...).
> Yeah, but how do you distinguish that?
By enumerating the tracks?
I feel like you are intentionally trying to turn this into a spherical cow situation by insisting on solutions that work in an information vacuum. All of this sort of problem in the real world is solved by sensor fusion from multiple data sources.
Trains are captive to a particular track - they can only change tracks at specific pre-determined junctions. With a pretty minimal sensor feed from each junction we can statically determine which track segment a given train is on. With GPS we can shrink that segment length down to a few kilometres. By sharing data persistently every single train can validate that central command-and-control is not about to run it into another train from behind...
There cases which are genuinely challenging engineering problems (i.e. derailments that impinge on adjacent tracks), but this trivial case of bumper cars is not one of them.
>> Yeah, but how do you distinguish that?
> By enumerating the tracks?
How do you distinguish that the <2km away detected moving train runs on your track on not on the one directly besides you?
> Trains are captive to a particular track - they can only change tracks at specific pre-determined junctions.
Yeah, but you do not know which one it runs on. Your entire point is that you do not trust the information the infrastructure knows. If you do allow to use information from the infrastructure, then why not use the sensors that already exist along the whole train network? It's also very unlikely that there is something electrically connecting the two rails of a track, that you would like to run over. So why wouldn't you trust that information, but some vague GPS signal?
> we can statically determine
But the current state of the art is, that you deterministically and provable know where a train is. Why would you use a more expensive, less safe system? The current focus to reduce errors in computing train conflicts is to prevent random bit flips due to radiation at the same time in three different computers running different software.
> I feel like you are intentionally trying to turn this into a spherical cow situation
The situation where you proposal makes sense, but the current system does not, is a spherical cow situation, that's my point.
4 replies →
Redundancy in a central node is an oxymoron. Safety critical systems need to work 100% of the time if any single node fails, and stop safely if there are multiple failures and it can't work properly. Doesn't matter if each individual node is 100% reliable by itself because connections fail too.
You can still have redundancy in the computation, that isn't prevented by being in the same location.
> Safety critical systems need to work 100% of the time
Railway safety must be safe 100% of the time, but not available 100% of the time. Also the point of a signal box is to make decisions about the order of train assignment to track. You absolutely do not want multiple points of truths for that.