KAI engine · reference case
Which manhole did the pollution come from?
A real combined sewer network of about 995 manholes, run in EPA SWMM. The engine finds the source of a release from sensor data alone, says how sure it is, and shows where extra sensors help most. The same method runs on an industrial park's sewer once its network is mapped.
How often the engine names the right source
Share of blind test releases where the engine's first guess was the exact manhole (darker), or within 500 m of it, the same few streets for an inspector (lighter). Sensor sites and the confidence calibration were fitted on separate training releases; with 53 test releases, each figure is good to about ±7 points.
- Right first guess (exact manhole)
- First guess within 500 m
One blind release, traced
1.35 kg released at manhole G71F314. The engine saw only the sensor data below, without knowing where or when.
- Possible source
- Likely source (size = probability)
- True release point
- Sensor
Most likely sources
- 1.G71F314✓ true source27 %
- 2.G72F76624 %
- 3.G80F13P20 %
- 4.G80F1109 %
- 5.G80F5209 %
- 6.G73F2745 %
- Mass released
- 1.35 kg (est. 1.24 kg)
- Release time error
- 0 min
- Inlet sensor
- Sensor 1
- Sensor 2
- Sensor 3
- Sensor 4
Why calibrating the network model matters
The same blind releases, traced twice: once with the network model's flows matching the real ones (a model calibrated with flow and level data), and once with every manhole's inflow off by about 25 % (an uncalibrated model). Right first guess, exact manhole.
| Sensors | Calibrated model | Flows off by ~25 % |
|---|---|---|
| Inlet, online probe | 40 % | 21 % |
| Inlet + 5 existing utility sites | 34 % | 40 % |
| Inlet + 4 chosen sites | 60 % | 59 % |
Differences of less than about 10 points are within the noise of 53 test releases. The clearest gains come from where the sensors are; a calibrated model then lifts the share of first guesses within 500 m from 62 % to 76 %. This is why a park water diagnostic maps the sewer and calibrates the model against a few weeks of flow data.
Are the probabilities honest?
When the engine says it is this sure of its first guess, how often is it right? (5 sensors, blind releases)
| Engine's confidence | Releases | Actually right |
|---|---|---|
| 0–50 % | 27 | 44 % |
| 50–80 % | 14 | 71 % |
| 80–95 % | 2 | 50 % |
| 95–100 % | 10 | 80 % |
When the source is right, the release time is typically within 5 minutes and the mass within 9 %.
What is real and what is simulated
- Real: the sewer network of Bellinge and Brændekilde (Odense, Denmark): pipes, manholes, pumps, basins, overflows and control rules, and the utility's dry-weather inflows at 449 points (12.4 L/s on average).
- Real: the first sensor candidates are the utility's own monitoring sites.
- Simulated: the pollution releases. This network has no public pollutant records, so each test puts a known mass into a manhole at a random time, and the engine has to find it from noisy sensor data (10 % noise). Test releases last 15 to 90 minutes and run with each manhole's inflow changed at random (by about 25 %), so the model that finds the source is not the one that made the data.
- Left out: 6 candidate manholes at the tops of almost-dry pipes, where SWMM's water-quality routing loses tracer at night, and their test releases.
- Not used: the Bellinge rain-gauge records, which are licensed for non-commercial research only. This case is dry weather, when industrial spills are most often traced.
How it works
- EPA SWMM runs the real network with a tracer released at every possible source, at every hour of the day.
- For any sensor data, the engine tests every source and release time, fits the released mass, and turns the fit into a probability for each source.
- The probabilities are calibrated on training releases, because the model is never exactly the real network.
- Sensor sites are added one at a time, each time where they most improve the probability given to the true source.
Data: The Bellinge data set (Nedergaard Pedersen et al., 2021, ESSD 13, 4779–4798), doi.org/10.5194/essd-13-4779-2021, licence CC BY 4.0. Model file pinned to a fixed version (SHA-256 1ac98b81fbce…). Built 2026-10-21 by the KAI engine.