An open research programme on leak detection and localisation in water distribution networks, run on the one public benchmark with a realistic network, two years of SCADA data and a published leak schedule: BattLeDIM's L-Town (Vrachimis et al. 2022; Zenodo 4017659, CC-BY-4.0), which infra-waterloss could not reach and replaced with Net3. It asks how much of a detector's reported skill is above what random alarms would score, and whether a data-driven detector can carry a false-alarm rate fixed before the year is seen.
Step 1 (2026-11-12) builds the harness: SHA-256-checked loader, ground truth from the leak-flow files (onsets match Melnikov 2025's table), matching (an alarm counts for the running leak with the latest onset within 30 days) and distances along the pipe graph. Detector D1 fits each of the 33 pressure sensors on hour-of-week, tank level and the three inlet flows over a trailing two-week window ending a week back, runs a one-sided CUSUM on the standardised residuals, and re-anchors its window after each alarm (an earlier hold-off version re-alarmed on the same leak; a unit test caught it).
Tuned on 2018 only (at most 2 false alarms), on 2019 it detects 13 of the 19 new leaks with 2 false alarms, a median delay of 104 h (11 to 341), and 16 of BattLeDIM's 23 counting the 3 of 4 carried-over leaks it found in 2018; without localisation, so not comparable with the published 15 to 17 of 23, which required a reported pipe within 300 m. Against the same 26 alarms placed at random (1,000 draws, 14 days apart, random sensors), the detection count is near chance (median 11, 95% range 8 to 13, P = 0.105), as is detection within 72 h (4 against 2, P = 0.145); but the alarming sensor is within 300 m along the pipes of 5 detected leaks against a chance median of 0 (P < 0.001). L-Town is a simulated network with simulated leaks; 2018 labels were used for tuning, which competitors did not have; nothing here is an operational recommendation.
Show the rest of the summary (30 more paragraphs, 2,538 words)
Step 2 (2026-11-12) reimplements LILA's identification (Daniel et al. 2022: pairwise sensor regressions on the inlet flow, a most-affected-sensor reconstruction error, Montgomery's CUSUM) and runs it blind on D1's trailing window; the authors' released notebook instead writes in each known leak's sensor and calibration windows. Tuned on 2018 (delta 1, Cthr 20), it detects 14 of 19 with 1 false alarm and 26 alarms: P = 0.019 against chance on the count, which refutes the programme's pre-registered H1 as written, but thinly: 2 of the 13 settings with 20 or more alarms clear 0.05, and a disclosed post-hoc change of the most-affected-sensor tie-break (ties occur in 46% of hours, because the inlet flow drives every sensor) gives 13 of 19 and P = 0.265.
Its most affected sensor is within 300 m of 7 of the 14 leaks (chance median 0, P < 0.001), and every one of the 27 settings beats chance on that count. Zenodo was unreachable; the inputs came from public GitHub copies whose SHA-256s match Zenodo's.
Step 6 (2026-11-13) tests the practitioner's minimum-night-flow statistic on the district water balance (p227 + p235 less the tank T1's storage change, least hourly flow from 00:00 to 06:00, a CUSUM on its trailing median and MAD), pre-registered as hypothesis H4: that it detects incipient leaks sooner than pressure CUSUM. Tuned on 2018 the same way, it detects 3 of the 9 new 2019 leaks under 12 m³/h, against 7 for D1 and 7 for LILA, and is earlier than D1 on 1 of them: H4 refuted.
The nightly minimum's two-week robust sd is 10.7 m³/h, the size of the leaks; a post-hoc look shows Friday and Saturday nights 10 to 15 m³/h above the week, and removing that pattern leaves 7.1 m³/h. Over all 19 new leaks it finds 10 with 6 false alarms, at chance on the count (P = 0.80).
Step 7 (2026-11-13) runs the two fixes step 6 named, pre-registered: a 2018 weekday profile and the night balance net of the 82 AMR meters (8.2% of district flow). Together they halve the night-to-night spread to 5.1 m³/h, yet the combined statistic still detects 3 of the 9 small leaks (AMR-net alone: 6 of 9, median delay 473 h): H4 refuted again. The ground-truth leak flows show why: five of the nine are ramps leaking under 0.5 m³/h for a week and 1.3 to 6 m³/h at day 30, invisible to any trailing-baseline statistic inside the 30-day match; and the pressure detectors were credited with four of them at the hour the leak carried 0.01 to 0.22 m³/h, chance matches rather than detections. A detection-only score should credit an alarm only once the leak's flow passes a stated level.
Step 8 (2026-11-13) rescores all six detectors, unchanged, under a pre-registered flow-gated match: an alarm counts for a leak only from the hour its own flow reaches φ = 1 m³/h (swept 0.5 to 5), with the chance level under the same rule. D1 falls to 11 of 19 (P = 0.319), LILA to 13 (P = 0.013, but 0.057 at 2 m³/h), and the AMR-net night flow rises to 15 of 19 with 3 false alarms (P = 0.001, below 0.005 at every gate).
Two of three written predictions failed. The swing is two or three leaks, from alarms before a ramp's gate and from which of two overlapping leaks an alarm credits: on L-Town the detection-only verdict depends on the matching rule, while the 300 m location score beats chance under every rule.
Step 9 (2026-11-13) rescores the six detectors under six matching rules (14, 30 or 60-day windows, with and without the 1 m³/h gate), chance recomputed under each. D1 beats chance (P < 0.05) under 2 of the 6 rules, night flow M1 under 4, M3 under the 3 gated ones, LILA under all 6 but only 2 after a Holm correction across the six detectors; the 300 m location score beats chance under all 6. At the primary rule no pair of detectors differs on the 19 leaks (exact sign tests, smallest Holm-adjusted P 0.59; M3's 4-leak lead over D1 has a bootstrap interval of 0 to +8). Three of four pre-registered predictions held. A detection-only comparison on BattLeDIM needs its rule, its chance level, a multiplicity correction and a paired test stated, and more leaks than one test year provides.
Step 10 (2026-11-13) cross-fits the thresholds: each detector's grid is re-selected on 2019 by the same rule and scored once on 2018's 13 leaks, then pooled with the forward fold for 32 held-out leaks. Five of six thresholds move with the tuning year. Under the gated rule no pooled count survives a Holm correction across the six detectors (smallest 0.27); M3, 15 of 19 in 2019, finds 6 of 13 in 2018 against 8 for D1 and 9 for LILA.
Ungated, LILA's 24 of 32 survives (Holm P = 0.024). No pair of detectors differs on the 32 leaks (smallest Holm-adjusted P 0.86). Three of four pre-registered predictions held; the one that failed was that M3 would beat chance after correction.
Step 11 (2026-11-13) localises leaks with WNTR sensitivities on the nominal L-Town model (782 junctions, 24 h mean pressure change at the 33 sensors, best Pearson correlation with the residual): 31% of D1's and 36% of LILA's detected 2019 leaks fall within 300 m (pre-registered H3 inconclusive), and 47% of all 32 leaks localised at half-peak flow, about ten times what a random junction scores (4%). The sensor whose pressure fell most does as well (13 of 32; paired P = 0.75), so the uncalibrated nominal model adds nothing on this benchmark; slow-ramp leaks are 0.6 to 2.8 km off for every method.
Step 12 (2026-11-14) tries the two textbook fixes, hour-by-hour signatures (4,692 EPANET runs) and a leak size from the inflow balance: they put 12 of 32 leaks within 300 m against step 11’s 15 and the sensor’s 13, so H3 is refuted under the pre-registered primary (23% and 29%). The inflow balance sizes only 12 of 32 leaks within a factor of 2, and the nominal model’s tank-level pump control adds a 0.9 m artefact at one hour of the day.
Step 13 (2026-11-14) reimplements SICAMS (Melnikov 2025) and scores the paper’s own published detection times on the same harness: 13 of the 19 new 2019 leaks at 27 alarms against a chance median of 10 (P = 0.057; 14 at P = 0.026 with its one stated false alarm added), with 9 within 72 h (P < 0.001). An automatic SICAMS, tuned on 2018 by rule, detects 13 with 26 false alarms (P = 0.82): one false alarm in 2018 becomes 26 in 2019 as leaks pile up; its alarming sensor is within 300 m for 5 (P < 0.001).
Step 14 (2026-11-14) simulates the leak-free network from L-TOWN_Real.inp, the organisers’ generating model: it reproduces the published pressures to their 0.01 m rounding (the data carry no sensor noise), and on it the frozen pressure detectors raise 0 (D1-F) and 1 (LILA-auto) alarms in 716 days against 26 each on the real data, while the night-flow detectors raise as many as before (M3 31 against 29), and M3’s leak-free alarms are credited with 14 of the 19 new 2019 leaks, so its earlier detection scores carry no evidence; one deterministic history, without demand uncertainty.
Step 15 (2026-11-15) localises the 32 leaks on the oracle residual (published minus leak-free pressure, taken against each leak’s pre-onset level): the nominal-model signatures put 16 of 32 within 300 m against 15 on step 11’s regression residual (paired P = 1.0) and 14 for the most-affected sensor (P = 0.73), though the best signature correlates with the true leak effect at a median r = 0.97; without the pre-onset baseline the overlapping leaks swamp it (2 of 32).
Every leak shows at some sensor within a median 0 h of onset (at most 20 h), while the pressure CUSUM and LILA alarmed a median 104 and 138 h after onset: the residual limits detection delay. Oracle bounds, not methods.
Step 16 (2026-11-15) localises with the generator’s own model, L-TOWN_Real.inp, restarted on each leak’s day with its true tank and pump state and the leak’s true size (25,056 EPANET runs): 15 of 32 within 300 m, against 16 for the nominal model and 14 for the most-affected sensor (P = 1.0 for both); all five pre-registered predictions failed. Post hoc, abrupt leaks localise 11 of 15 with any method and incipient leaks 4 of 17: a slow leak reaches half its peak a median five weeks after onset, by which time other leaks’ flows have changed by 2.7 times its own, so its residual is mostly other leaks. This withdraws step 15’s reading that the nominal model limits localisation; the residual does.
Step 17 (2026-11-15) tests that reading: every leak pipe of L-TOWN_Real.inp is split at its midpoint and draws its published flow, which reproduces the published pressures exactly (median difference 0.0000 m), so published minus the replay without the target leak is that leak’s own effect. On it the generator’s-model localiser puts 27 of 32 within 300 m (median error 56 m; 14 gained and 2 lost against step 16, P = 0.004), incipient leaks 14 of 17, the nominal model at a fixed 5 m³/h 28 of 32 and the most-affected sensor 20 (P = 0.065); four of five pre-registered predictions held. An oracle bound: the open problem is a residual a utility can compute that removes the other leaks.
Step 18 (2026-11-15) tries the simplest such residual, with no model and no leak list in it: the 24 h mean pressure from the leak’s localisation hour minus the same hours a week earlier. The nominal model puts 12 of 32 within 300 m (abrupt 8 of 15, incipient 4 of 17), no better than the regression residual’s 15 (P = 0.45) or the most-affected sensor’s 12, and 16 short of the oracle; removing demand change with the leak-free model does not help (11), and lags of 24, 72 and 336 h give 8 to 12.
Three of five pre-registered predictions failed. Over a week the other leaks’ flows change by about as much as a slow leak’s own (median 2.7 against 2.3 m³/h), and only a third of the slow leak’s own change is left in the window: no lag separates them.
Step 19 (2026-11-16) models the other leaks instead: a non-negative greedy fit of the leaks’ pressure effect over the 782 nominal signatures, refitted a week later from the same atoms, placing the leak at the junction whose fitted flow rose most. It also finds 11 of 32 (slow leaks 1 of 17); four of five predictions failed. The fit leaves only 4.6% of the signal, and, seen after the run, its atoms land within 300 m of 10 of the 17 slow leaks against 5.5 by chance (P = 0.016): the slow leak is found, but a one-week rise cannot tell it from the others.
Step 20 (2026-11-16) tracks the fit day by day on a residual a utility could compute from its own data (a regression frozen on 2018’s first two weeks), dates each fitted source by the day it entered and places a new leak at the newest one: 8 of 32 within 300 m (slow leaks 2 of 17), below the one-week rise on the same residual (14; P = 0.15). Five of six predictions failed: the tracker keeps a median 20 sources active and starts a new one every three to four days, so the newest is rarely the leak.
Step 21 (2026-11-16) runs that tracker unchanged on the demand-oracle residual (published pressure minus the generator’s leak-free run): a median 8 sources active instead of 20, 13 of 32 within 300 m (slow leaks 3 of 17) and 9 of the 17 slow leaks given a source dated to their own start (4 before). Five of six predictions held, but the gain over the data-only residual is not significant on 32 leaks (P = 0.27): the residual’s drift, not the dating rule, sank step 20, and slow leaks’ sources enter a median 18 days late.
Step 22 (2026-11-16) keeps that tracker and picks, among sources dated within the last 30 days, the one whose fitted flow rose fastest over two weeks: 14 of 32 within 300 m and 6 of the 17 slow leaks (newest source: 3), but 8 of 15 abrupt leaks against 10 (likely because a burst’s source jumps once and goes flat; not yet tested). Two of five predictions held; the net gain is one leak (P = 1.0), and the two rules together hit 18 of 32.
Step 23 (2026-11-17) fits each recent source’s flow with a step or a ramp and picks the one whose change is timed nearest the alarm: every slow leak’s nearby source is fitted as a ramp (8 of 8), but the rule finds 13 of 32 and 3 of the 17 slow leaks, the newest-source result again, because a still-rising ramp’s half-level day falls days before the alarm. Three of five predictions failed; the shape label, not the timing, is what to use next.
Step 24 (2026-11-17) lets the fitted shape choose: the steepest fitted ramp when one exists, else the newest source, finds 15 of 32 (6 slow leaks, 9 bursts), but in-sample.
Step 25 (2026-11-17) tests it out of sample on 64 new leaks simulated through the generator’s model and added to the published pressures: 25 of 64 within 300 m against 16 for the newest source (11 against 2 discordant, P = 0.022; chance 2.1), with 13 of 37 slow leaks against 5. All six pre-registered predictions held; the residual still removes demand with the generator’s model, and the alarm hour is still an oracle.
Step 26 (2026-11-17) removes the time oracle: the project’s own pressure detector (D1, thresholds frozen from 2018) alarms, then the same rule localises at the alarm day. On the 30 of 64 injected leaks it detects, the rule places 8 within 300 m at the alarm day against 15 at the oracle day (P = 0.065), still far above chance (1.0 expected, P = 4e-6); on the 10 real 2019 leaks it detects, 2 against 6. End to end, detect-then-localise places 8 of 64 injected and 2 of 19 real 2019 leaks.
Four of six predictions failed. A sweep after the run shows why: moving the oracle day by one day costs 2 to 6 hits, so steps 21 to 25 were tuned to an exact day no detector supplies.
Step 27 (2026-11-19) re-times the rule to the alarm: of 17 rules (a lag or a vote over days around the alarm), the best on 2018’s detected leaks, a vote over the four days before the alarm, places 4 of the 22 detected 2019 leaks within 300 m against 3 at the alarm day and 12 at the oracle day (chance 0.8); three of five predictions held, but the rules’ ranking does not carry from 2018 to 2019 (Spearman 0.09), so the oracle day’s advantage is leak-specific timing and the next detector runs on the residual, where leaks show within hours.
Produced with AI assistance (Claude). Results are from the code in this repository and are simulated or stylised unless the abstract says otherwise; nothing here implies affiliation with or endorsement by any lab, researcher or company cited. MIT licensed.