Leak detection in water & gas networks

Buried pipes leak silently for months. Utilities historically find leaks through periodic night-flow analysis, manual listening rounds or customer complaints. AI combines permanent acoustic sensors, district-meter flow/pressure data, hydraulic models and (for gas) methane measurements to raise early, located and ranked leak alarms, so crews are dispatched to the right street instead of surveying the whole network. Most of the work is done by classical signal processing, statistics and hydraulic models; machine learning mainly improves noise rejection, ranking and explanation.

Problem & value

• Who: water utilities and Stadtwerke, gas distribution system operators, municipalities with small networks, industrial sites and campuses with private water/steam/compressed-air/gas piping.
• Today: minimum-night-flow analysis per DMA, periodic acoustic surveys with ground microphones and correlators, and reactive repair after visible bursts. Leaks that never surface are found late. Gas networks add safety obligations (explosion risk) and, in the EU, methane-emission rules.
• AI value: continuous scoring of flow, pressure and acoustic streams detects small emerging leaks earlier, separates leak noise from traffic and pump noise, narrows the search area and ranks candidates by estimated loss and consequence.
• Regulatory pull: the recast EU Drinking Water Directive (EU) 2020/2184 introduces leakage reduction measures: Member States reported national leakage assessments by January 2026, the Commission is to set an EU leakage threshold by January 2028, and national action plans are due by January 2030 [3]. Methane rules (Regulation (EU) 2024/1787) add leak detection and repair duties for gas infrastructure; check the exact survey and repair deadlines for your asset class.
• Benefit figures: vendors publish case studies (for example Echologics lists PUB Singapore, West Virginia American Water and Las Vegas Valley Water District) but the cited page gives no quantified results [4]; treat savings as site-specific and measure your own baseline.

Variants

• Water distribution (DMA analytics): statistical or ML detection on district-meter flow and pressure; the BattLeDIM benchmark covers exactly this task [1][2].
• Acoustic monitoring: permanent noise loggers or in-pipe/valve-mounted sensors for distribution and transmission mains; AI classifies leak versus background noise [4][5].
• Gas distribution: handheld/vehicle methane detectors (e.g. Sewerin SR-LD 800 detects methane from up to 200 m) plus pressure/flow analytics for the network [5]; satellite methane quantification for facilities (GHGSat DATA.SAT) [6].
• Satellite and aerial screening of water mains: remote-sensing providers screen regions for likely leak areas to prioritise ground work (product claims should be validated by pilot).
• Industrial/private variants: compressed-air leak detection by ultrasonic cameras in plants (see ANOM-005), smart-home water flow monitors that shut off a valve on abnormal flow.
• Boundary: pipe-condition and remaining-life prediction belong to PRED-046/PRED-006; theft and grid anomalies in electricity to ANOM-008.

How it works

sensors (flow, pressure, acoustic, AMI) → ingestion/streaming → cleaning and DMA water balance → anomaly detection and leak classification → localisation (hydraulic model + correlation) → ranked alerts → field confirmation → repair and feedback labels

• Statistical/hydraulic layer: night-flow baselines, residuals between measured and modelled pressure/flow (EPANET-style model), pressure-wave analysis for bursts. Wikipedia's overview lists the internal methods (balancing, statistical, real-time transient models) and external ones (acoustic, fibre-optic DAS/DTS, thermal) [10].
• ML layer: unsupervised or semi-supervised anomaly detectors on multivariate time series; audio classifiers on hydrophone or accelerometer spectra; graph-based localisation over the pipe network. Pretrained time-series foundation models such as Chronos-Bolt and TimesFM 2.5 (both Apache-2.0) give zero-shot demand forecasts whose residuals can serve as a baseline [7][8].
• LLM layer (optional): summarise an alarm with its evidence, draft the work order, answer operators' questions over procedures. It never decides on shutting valves.
• Human in the loop: control-room engineer validates alarms; field crew confirms with correlator or ground microphone before excavation; any automatic valve action requires a classic, certified safety layer.

Data preparation

Leak detection needs SCADA flow, pressure and reservoir-level time series, DMA inlet meters and smart-meter (AMI) consumption, plus the GIS pipe register (material, age, diameter) and the hydraulic model. In a typical utility these sit in the SCADA historian, the GIS, the billing/ERP system and, for older assets, in paper plans and Excel lists. Typical gaps are unsynchronised timestamps, sensor drift and gaps, unmapped DMA boundaries, pipe records without age or material, and no labelled history of confirmed leaks.

Preparation steps: (1) inventory sensors and data sources and check readiness (PREP-001, GUIDE-003); (2) export historian and ERP data via connectors or APIs (PREP-005); (3) clean the time series (resampling, outlier and gap handling, time-zone alignment) with pandas or a time-series database; (4) correct and complete GIS attributes; (5) harvest labels from repair tickets and work orders (PREP-020) to build a test set with the field crews (PREP-018). Expect roughly 40-60 percent of project effort in data work, mostly for the first network area; later DMAs reuse the pipeline. SCADA access must stay read-only and separated from the office network.

Implementation paths

Which path for whom

• SMB (small utilities, industrial site operators): Path A - a managed leak-detection service or logger product avoids building hydraulic and signal-processing know-how in-house.
• Enterprise (large utilities): Path B - leakage detection is core business and must integrate with SCADA, GIS and the hydraulic model, which off-the-shelf products rarely do fully.
• Public (municipal utilities, Stadtwerke): Path C - critical infrastructure (KRITIS) and procurement rules favour sovereign, self-operated stacks that also serve other network analytics.
• Developer: Path C - open-source time-series and anomaly-detection libraries with EPANET-style models give full control over the algorithms.

Path A — Buy (off-the-shelf)

Acoustic monitoring and leak-localisation systems from established utility-technology vendors: Echologics (EchoShore-DX/-TX, ePulse) [4], Gutermann (Germany/Switzerland), Sewerin (Germany; water and gas detection and correlators) [5], plus network-analytics platforms from the large utility software vendors. GHGSat supplies satellite methane measurements [6]. Right choice for most utilities: hardware, calibration and alarm logic come bundled. Prefer EU-hosted data and a clear data-processing agreement; sensor data rarely contains personal data except fine-grained smart-meter readings.

Path B — Build on cloud APIs

Stream SCADA/AMI data into a cloud time-series store in an EU region, run anomaly detection with managed services or forecasting models (Chronos-Bolt via AutoGluon or SageMaker JumpStart [7]), and use an LLM API (EU-hosted where required) for alarm summaries and ticket drafts. Sensible for utilities with a data team that wants to integrate analytics into existing GIS and CMMS. Keep critical alarm logic independent of external APIs.

Path C — Self-host (open source)

Hydraulic simulation with EPANET (open source) and the BattLeDIM data and scoring code (EUPL-1.2) as development benchmark [2]; Python stacks for anomaly detection; Chronos-Bolt or TimesFM as zero-shot baselines [7][8]; MQTT/OPC UA ingestion and a time-series database. CPU or small edge devices (HW1) are enough for DMA-level analytics; audio classification on edge boxes at the sensor. Realistic for municipal utilities with an OT/IT team; SCADA network separation is the main effort.

Path D — Custom training

Worthwhile when you have many labelled confirmed leaks with acoustic recordings and a homogeneous network (pipe material and soil). Train leak/no-leak audio classifiers and site-specific localisation models; use simulated leaks in the hydraulic model to augment rare labels, then validate on real events. Expect labels to be scarce and noisy.

State of the art (2026)

• Benchmarks: BattLeDIM (L-TOWN network, 2018/2019 SCADA data, CC BY 4.0) remains the standard open benchmark for DMA-level leak detection and isolation [1][2]; many published methods are evaluated only on simulated leaks.
• Practice: permanent acoustic sensors on distribution mains and valves, correlation-based localisation, DMA analytics and satellite screening are commercial products [4][5]; foundation models for time series are used as baselines rather than as turnkey leak detectors [7][8].
• Gas: methane-specific sensing (laser-based handhelds, vehicle surveys, satellites) is the detection layer; AI helps with quantification, deduplication and prioritisation [5][6].
• Limits and failure modes: false alarms from pumping, valve operations, traffic and plastic pipes that attenuate acoustic signals; sensor drift and missing data; few confirmed labels; model skill degrades after network changes (new DMA boundaries, rezoning); satellite and simulated-data results often exceed field performance; claimed savings are vendor-reported.

Shared resources & combinations

• Capabilities shared: anomaly-detection, time-series, streaming, edge-runtime (with ANOM-005, PRED-006, PRED-046); graph-ml and simulation (hydraulic model as digital twin); audio-classification; remote-sensing (with VIS-013, VIS-003).
• Hardware: one CPU server plus edge gateways (HW1) can serve DMA analytics together with industrial anomaly detection (ANOM-005); GPU only needed for audio or foundation-model workloads.
• Data pipelines: OPC UA/MQTT ingestion, asset register and GIS, unified namespace (DATA-016), CMMS work-order integration.
• Natural combinations: ANOM-005 (industrial sensors), ANOM-008 (energy/grid anomalies on the same utility's electricity or district heat), PRED-046 (asset condition and replacement planning), VIS-003 (drone inspection of exposed pipelines), ANOM-030 (KPI anomaly detection on water-loss indicators).

Effort, cost & KPIs

• Path A: weeks for pilots in a few DMAs, months for network rollout; cost drivers are sensor hardware, installation and subscription.
• Path B/C: months, driven by data integration (SCADA, GIS, AMI), data cleaning and alarm tuning; ongoing cost is mainly staff time.
• Path D: several months plus a long label-collection phase.
• KPIs: time from leak start to detection; share of alarms confirmed in the field (precision); leaks found per crew-day; real losses and non-revenue water (from the water balance); infrastructure leakage index; repair lead time; avoided supply interruptions; for gas, emissions detected and repaired against deadlines.

People & change

Involve management (budget, KRITIS responsibilities) and the head of network operations from the start, since they own the process. Field crews and control-room operators are the end users: involve them in the pilot DMA early, because they validate the alarms and their experience decides acceptance. IT and IT security are needed for read-only SCADA access and network segmentation (GUIDE-013); data owners in GIS and billing release the data; sensor and logger suppliers must be coordinated.

Works council: depends. Pure network data is not personal, but logging of crew assignments, locations or response times can make the system suited to monitor employees, which triggers co-determination under section 87(1) no. 6 BetrVG; inform the council in good time under section 90 (1) no. 3 in any case (GUIDE-004). Public utilities involve the staff council (Personalrat). A DPIA is normally not needed unless AMI consumption data is analysed at household level; then involve the DPO.

AI literacy: support staff with short training on what alarms mean, false positives and limits (AI Act Art. 4, GUIDE-006). Typical concern: "the model replaces the experienced listener". Address it by keeping the crew's confirmation as the decision point. Start with one pilot DMA (GUIDE-009) and communicate results openly (GUIDE-008).

Risks, regulation & compliance

• EU AI Act: classification depends on the function. Annex III point 2 lists AI systems intended as safety components in the management and operation of the supply of water, gas, heating or electricity as high-risk [9]; "safety component" means a component whose failure endangers health and safety of persons or property, or whose purpose is to prevent or mitigate such risks [11]. A leak-detection system that only advises engineers is typically not a safety component (minimal risk), while one that automatically controls shut-off valves or gas safety functions may be. Assess per deployment; document the reasoning.
• GDPR: network sensors are not personal data, but high-resolution smart-meter data can reveal household behaviour; aggregate, minimise and restrict access. Field crews' location data needs the usual employment-law basis.
• Sector law and security: drinking-water and gas supply are critical infrastructure (NIS2 and national KRITIS rules in Germany); keep SCADA segmented, use read-only data diodes or gateways where possible. Technical rules from DVGW apply to German water and gas operators. Methane regulation requires documented detection and repair.
• Operational risks: false negatives lead to continued loss, false positives to wasted excavation; never let ML alone close valves; keep a fallback to conventional surveys.

Getting started

1. Build a water balance and pick two or three DMAs with high night flow as a pilot; collect 6 to 12 months of flow, pressure and work-order history.
2. Reproduce a baseline on BattLeDIM data [1][2] with simple residual-based detection, then compare with a zero-shot foundation-model forecast [7][8].
3. Pilot a commercial acoustic or gas-detection product on the pilot area [4][5] and log precision of alarms against field confirmation.
4. Decide path by comparing alarm precision, integration effort and data-residency needs; define an AI Act assessment note for the chosen scope.
5. Connect alarms to CMMS/GIS and add feedback labels for later model improvement.

Sources

1. BattLeDIM dataset (Zenodo, CC BY 4.0): https://zenodo.org/records/4017659
2. BattLeDIM benchmark repository (KIOS Research, EUPL-1.2): https://github.com/KIOS-Research/BattLeDIM
3. European Commission, Drinking water (leakage measures and dates): https://environment.ec.europa.eu/topics/water/drinking-water_en
4. Echologics product overview (EchoShore, ePulse, case-study titles): https://www.echologics.com/
5. Sewerin product overview (water and gas leak detection, SR-LD 800): https://www.sewerin.com/en/
6. GHGSat DATA.SAT satellite methane monitoring: https://www.ghgsat.com/en/
7. Chronos-Bolt model card (Apache-2.0): https://huggingface.co/amazon/chronos-bolt-base
8. TimesFM 2.5 model card (Apache-2.0): https://huggingface.co/google/timesfm-2.5-200m-pytorch
9. EU AI Act Annex III: https://artificialintelligenceact.eu/annex/3/
10. Leak detection methods overview: https://en.wikipedia.org/wiki/Leak_detection
11. EU AI Act Article 3 definitions (safety component): https://artificialintelligenceact.eu/article/3/