Adaptive Leak Detection for Water-Distribution Networks Under Realistic Label Availability: Distribution-Shift Analysis, an Edge-IoT Prototype, and Network-Level Evaluation

We evaluate rolling-window XGBoost retraining (Adaptive XGBoost, AXGB) for leak detection on the simulated BattLeDIM benchmark, with an edge prototype and network simulation. Hourly supervised retraining on 6 h label windows (AXGB-S) reaches F1 = 0.993, but 99.1% of its refits repeat the last label; last-label persistence reaches 0.999. With a one-day label delay, AXGB-S (0.928), persistence (0.934) and an Adaptive Random Forest (0.931) detect leaks only after labels arrive. The shift defeating static XGBoost (F1 = 0.448) coincides with growing unlabelled background leakage, an important though not isolated explanation; metered demand decreased. A label-free, alarm-gated adaptive inflow model (AXGB-U) detected one of two held-out 2018 leaks after 1.2 h (0.36 false alarms per week) and two of four large abrupt 2019 leaks, after 0.8 and 12.8 h. An Arduino–Raspberry Pi prototype demonstrated the edge architecture once. In OMNeT++/INET simulations, injected loss reaches the detector almost one-to-one; sub-second delay has no effect. In a closed-loop channel model, AXGB-U detected the first test leak in every run up to 50% loss, but false alarms rose to 3 per week and four runs detected it after 27–28 h. Uncapped 1 Hz alarm bursts cost 53 times the 5 min cadence (30 min cap: 1.35 times).