Deep diveInfrastructure
Water leak detection with AI: how flow, acoustic and field data fit together
Water leak detection AI works best as a triage layer across signals a utility already has: district meter flows, night flow trends, acoustic logger alerts, pressure transients and customer reports. Models flag where leakage is probably rising and how urgent each area is; field crews confirm the leak, and their findings become the labels the models learn from. This page explains each method, what it can and cannot see, and how to combine them.
On this page
- Where leakage sits in the IWA water balance
- Leakage terms worth agreeing across operations and analytics
- Leak detection methods compared by what they can localize
- How signals, models and field crews form one detection loop
- Keeping anomaly detection honest through seasons and network changes
- Field records that turn crew visits into usable training labels
- Reporting rules that define what success means
- A hypothetical utility working through dozens of DMAs
- Questions and answers
- Sources
Where leakage sits in the IWA water balance
The water balance developed by the International Water Association and adopted by AWWA divides the volume a utility puts into supply into authorized consumption and water losses2. Losses split again. Apparent losses are water that reaches customers but is not measured or billed correctly: meter under-registration, data handling errors and unauthorized use. Real losses are physical: leaks and bursts on mains and service connections, and overflows at storage tanks.
Non-revenue water adds unbilled authorized consumption, such as firefighting and mains flushing, to both kinds of loss. That distinction matters before any sensor is bought. Leak detection only reduces real losses; apparent losses need meter testing, billing analytics and enforcement. A utility that skips the audit can spend heavily on acoustic sensors in a network whose main problem is aging customer meters.
Leakage terms worth agreeing across operations and analytics
- District metered area (DMA)
- A discrete zone of the network supplied through metered inlets with boundary valves closed, so its inflow can be measured and compared over time.
- Minimum night flow
- The lowest inflow to a DMA during the night, when customer use is lowest and leakage forms the largest share of flow.
- Legitimate night use
- Customer and non-domestic consumption expected during the night, estimated and subtracted from minimum night flow to approximate leakage.
- Background leakage
- Many small, often undetectable leaks at joints and fittings that run continuously and are managed mainly through pressure rather than repair.
- Infrastructure Leakage Index (ILI)
- The ratio of current annual real losses to the unavoidable annual real losses for that network, used as a performance indicator in the IWA and AWWA methodology2.
- Leak noise correlator
- Equipment that compares the timing of leak noise at two sensors on either side of a leak to calculate its position along the pipe.
- Pressure transient
- A rapid pressure wave in the network, caused by a burst, a pump trip or fast valve operation, recorded by high-rate pressure loggers.
Leak detection methods compared by what they can localize
| Method | What it detects | How precisely it localizes | Main limitation |
|---|---|---|---|
| DMA flow and night flow analysis | Rising leakage and new bursts in a zone | To the DMA, sometimes to a sub-zone after step testing | Depends on DMA integrity and good estimates of night use |
| Fixed acoustic loggers | Persistent leak noise near fittings and valves | To a group of loggers, often a few streets | Works poorly on plastic and large-diameter pipe |
| Correlators and ground microphones | Leak noise between two contact points | To a short length of pipe, enough to excavate | Needs skilled operators and accurate pipe records |
| Fiber-optic acoustic sensing | Noise and vibration along a cable route | Along the fiber, often to a short distance | Only where suitable fiber runs along the main |
| Pressure transient monitoring | Bursts and damaging surges as they happen | Between loggers, using arrival times | Needs high-rate loggers and timing synchronization |
| Satellite and aerial survey | Likely leak areas from soil moisture or thermal signatures | To areas that still need ground survey | Affected by soil type, vegetation and recent weather |
No single method finds every leak. Most mature programs use zone-level analytics to decide where to send acoustic crews.
How signals, models and field crews form one detection loop
- Network signals
DMA inflows, pressures, logger alerts, customer contacts and work orders.
- Context features
Weather, day type, holidays, valve operations and known boundary changes.
- Anomaly scoring
Residuals from expected flow and noise baselines, combined per zone.
- Triage queue
Zones ranked by estimated volume, confidence and consequence.
- Field verification
Crews step test, correlate and pinpoint, or record that nothing was found.
- Repair record
Location, cause, pipe material and estimated flow before and after repair.
Keeping anomaly detection honest through seasons and network changes
A useful leakage model predicts what each DMA's flow should be, given temperature, day of week, holidays and recent trend, and scores the gap between expected and measured flow. A burst shows as a sharp step in that gap; a slowly developing leak shows as drift in night flow over weeks. Separate detectors for the two patterns usually work better than one model asked to catch both.
Most false alarms come from the network, not the algorithm. An open boundary valve merges two DMAs; a large industrial user changes shifts; an inlet meter drifts. Data quality checks on meters, a feed of valve operations from the work management system and a register of large users remove many of them before scoring.
Combining signals is where machine learning earns its place. A zone with rising night flow, two acoustic loggers reporting persistent noise and a cluster of low-pressure complaints is a stronger candidate than any one signal suggests. Ranking zones on estimated volume and confidence keeps crews on the leaks that matter, a pattern our predictive monitoring work applies to other networks too. Leak detection from acoustic sensors and flow analytics is one of the water use cases on our infrastructure page1.
Field records that turn crew visits into usable training labels
Reporting rules that define what success means
In North America, AWWA's M36 manual sets out the IWA/AWWA water audit method, guidance on AWWA's free audit software and performance indicators for reporting losses2. In England and Wales, Ofwat sets leakage performance commitments for each water company through its price reviews, with financial incentives tied to performance3. In the EU, the recast Drinking Water Directive (EU) 2020/2184 requires member states to assess leakage for larger suppliers, using the ILI rating method or another appropriate method, and opens the way to an EU leakage threshold and national action plans4.
These frameworks reward verified volume reductions, not alerts. That is why the field loop matters: a detection system should report leaks confirmed and repaired, and the change in measured losses, in the same units the audit or regulator uses.
A hypothetical utility working through dozens of DMAs
Questions and answers
Can AI find leaks without installing new sensors?
Often it can start without them. DMA inflow meters, pressure loggers, customer contacts and work orders already contain useful signal, and zone-level anomaly detection on that data can rank where to send crews. New sensors become worthwhile once you know which zones or pipe materials the existing data cannot resolve.
Why do leak detection models raise so many false alarms?
Usually because the network changed and the model was not told: a boundary valve left open, a large customer changing its pattern, a meter drifting or a pressure scheme adjusted. Feeding valve operations and large-user data into the model and checking meter health before scoring removes many of them. Field outcomes then show which remaining alerts are worth acting on.
Is satellite leak detection accurate enough to replace acoustic surveys?
No. Satellite and aerial methods narrow down areas where leakage is likely, which helps plan surveys in large or rural networks. They do not pinpoint a leak to the point of excavation, and their results depend on soil, vegetation and weather. Ground acoustic methods are still needed to confirm and locate the leak.
Does leak detection AI help with apparent losses?
Not directly. Apparent losses come from meter inaccuracy, billing and data handling errors and unauthorized use, so they need different analytics, such as detecting under-registering customer meters or mismatches between metered and billed volumes. Many utilities run both programs from the same water audit so the two are not confused.
Sources
- Infrastructure: water system intelligence — ColdAI
- M36 Water Audits and Loss Control Programs, Fifth Edition — American Water Works Association · checked 10 October 2026
- PR24 business plan table guidance, part 1: Outcomes — Ofwat · checked 10 October 2026
- Directive (EU) 2020/2184 on the quality of water intended for human consumption (recast) — EUR-Lex · checked 10 October 2026