The numbers, and their limits.
2,500 simulated cases the system had never seen, on a 20 km line with 6 sensors. We built the standard methods ourselves and ran everything on the same cases. Then we broke the model on purpose.
Same cases, same rules
How far off is the location?
Typical distance from the real leak. Shorter is better.
How many leaks does it catch?
Out of 100 real leaks. Longer is better.
All alarms tuned to the same false-alarm rate (5 in 100) on separate data. The standard methods are our own implementations.
closer to the leak than a moving-horizon estimator when the model is wrong.
of the time it names the right cause: leak, valve, sensor or pump.
of wrong-model cases where it says “do not trust my model”. It says so in 0.2% of normal cases.
Less product on the ground
Average spill per leak over 15 minutes. 100 simulated leaks.
No response
Today: alarm, operator shuts down after 2 minutes
Checked recommendation, acted on in 15 seconds
Fair warning: most of this gain comes from reacting in 15 seconds instead of 120. With the same 2-minute delay, a better alarm alone saved little (5.5 vs 5.2 m³). No pressure limit was broken in any case.
What is proven. What is not.
Proven, in simulation
- Finds 97% of leaks at 7 false alarms per 100 quiet cases. Typical location error 190 m.
- Tells a leak from a valve fault, a bad sensor and a weak pump.
- Keeps working when the pipeline model is wrong.
- Every recommendation passes a fixed safety rulebook first.
- 12,800 cases. Every number is reproducible from the code.
Not proven yet
- No real pipeline data has been used. None.
- Tiny leaks (under 1% of flow): it catches 68%. A plain neural network catches more.
- Pump trips still cause about one false alarm in five.
- Two faults at once are often read as one.
- One straight 20 km line, one product. No branches, no batches, no gas.
- Nodes are in design; none is built.
The next number comes from your line.
A month of your history, replayed offline, gives the one figure that matters: false alarms per month on real operations. We will show it to you whether it is good or bad.