Controlled Leaks in an Andean Transmission Main Exposed the Limits of Changepoint Detection
Leak detection in pressurized water mains sounds straightforward until the signal is too small, too brief, and too noisy to separate from normal operating variation. In an operational transmission main crossing steep terrain near Loja, Ecuador, investigators staged seven controlled leaks and asked five unsupervised methods to find the onset from 30-second flow and pressure records. The result was sobering: none of the detectors performed better than chance at identifying the leak onset, even though the flow signal shifted measurably across the seven events as a group.
The setting itself helps explain why this problem is so difficult. The pipeline ran through an Andean landscape with a 600 m experimental segment, an average longitudinal slope of 11%, and an elevation difference of 66.28 m between endpoints. In that kind of system, pressure responses are shaped not only by the leak but by wave propagation, boundary conditions, pump dynamics, and topographic head changes. A small leak can be swallowed by the hydraulics before any algorithm gets a clean look at it.
Flow Shifted Across Events, Pressure Did Not
Across the seven trials, the leak condition was created by valve settings equal to 3% to 5% of nominal flow. Relative to the leak-free observations immediately beforehand, flow decreased in every trial, with a mean change of −0.50%. A sign test gave p = 0.016, and a placebo test placed against random changepoints gave p = 0.003, confirming that the aggregate shift was real even if it was too small to be resolved reliably event by event.
Pressure behaved differently. Recorded at 0.1-m resolution, it showed no change across the controlled leak events, and the placebo test was essentially null with p = 0.99. The divergence matters because flow and pressure do not respond in the same way to a leak. Flow reflects mass continuity and the balance between supply and loss, while pressure is filtered through local impedance and the geometry of the network. On steep terrain, that filtering can mute the local signal that a detector is trying to catch.
One line captures the operational problem: no single leak exceeded its own detection threshold.
The flow change was only visible as a seven-event mean, with a signal-to-noise ratio of −3.3. In other words, the leak signal existed, but it was too faint to stand out from the background on a per-event basis. For a field engineer or a monitoring system, that is the difference between an actionable alarm and an ambiguous drift in routine operations.
Five Unsupervised Methods Missed the Leak Onsets
Five unsupervised approaches were tested on both flow and pressure series: PELT with Bayesian and Akaike penalties, Bayesian online changepoint detection, an OLS-CUSUM structural-break procedure, and a moving-average ±3σ chart. The study then compared each detector with a chance-level reference built by placing the same number of changepoints at random.
None of the five methods found leak onsets more often than chance for any method-signal pair, with pooled p values above 0.40 throughout. PELT, in both its Bayesian and Akaike forms, appeared aggressive on pressure and flagged six of seven events, but the apparent sensitivity came with a heavy penalty: 560 to 960 false alarms per day and precision between 13% and 33%. In a live utility setting, that alarm burden would quickly drown operators in false positives.
BOCD performed better on pressure than the other methods by average precision, reaching 0.54, but the gain did not survive Holm correction when set against the chance reference, with adjusted p = 0.14. The method therefore looked better on paper than in statistical comparison. Sequential changepoint methods can be excellent when a shift persists long enough for the posterior to accumulate evidence; they are far less persuasive when the perturbation is short-lived and sparse.
Short Leak Duration, Sparse Sampling, and the Physics of Missed Detections
The trials lasted 90 to 120 seconds, while the data were sampled every 30 seconds. That leaves only a handful of observations to characterize each event, and it is a harsh regime for any detector that depends on temporal accumulation. A leak that lasts only three or four samples may never build enough statistical weight to cross a threshold, especially when the underlying network is already operating under changing hydraulic conditions.
The pressure findings fit the hydraulics. In a complex transmission main, pressure is not a simple mirror of mass loss. It is shaped by local impedance, elevation, and boundary effects, so a small or brief leak may have a muted or delayed footprint. Flow is more directly linked to continuity, which explains why the leak left a detectable aggregate signature in flow but not in pressure.
There is a practical asymmetry here. For short leaks, the best detector may still be the one that acknowledges defeat quickly rather than one that keeps producing confident false alarms.
What an Optimal Sequential Detector Could Achieve
To bound performance, the study calculated how long an optimal sequential detector tuned to the observed flow shift would need to detect the event at one false alarm per day. The answer was 12 to 29 minutes, which is 8 to 20 times the duration of the leak trials themselves. That estimate is powerful because it sets a ceiling on what any method could reasonably do on these records: if the signal persists, detection becomes plausible; if it lasts only a couple of minutes, the odds collapse.
This is not a failure of mathematics so much as a mismatch between signal duration and sampling regime. A persistent leak producing the same flow change would be detectable within tens of minutes. A 90- to 120-second leak sampled every 30 seconds is another matter entirely, and the study’s null result across five methods is exactly what the signal physics would predict.
The field comparison therefore lands on a hard conclusion: chance-level performance is the expected outcome for leaks of this duration at this sampling rate. Pushing detectors harder without changing the data quality, the sampling cadence, or the hydraulic context would mostly increase false alarms.
Why the Field Setting Matters More Than a Simulation
Much of the leak-detection literature rests on simulated networks, synthetic events, or single-leak scenarios. This experiment used an operational pipeline in mountainous Ecuador, where soil instability, seismic activity, and steep topography complicate routine monitoring. Those conditions matter because they are the conditions utilities actually face. Simulations can be clean; field pipelines are not.
The study also provides a reproducible R pipeline and publicly available data through Zenodo, which should make it easier for other groups to test alternative approaches against the same records. The value of the work is not just in documenting failure. It shows where detection methods lose traction when leak duration is brief, sample density is sparse, and the hydraulic system is topographically complex.
For utilities, the message is concrete. A detector tuned for persistent shifts may still be useful, but it will not rescue a monitoring program that relies on coarse sampling to catch short leaks at onset. For researchers, the next step is not another abstract benchmark on idealized data. It is a better fit between hydraulic reality, sensor resolution, and the event duration that the algorithm is asked to detect.