AI Leak Detection and Non-Revenue Water Reduction

Acoustic leak detection guided by machine learning reaches over 93 percent accuracy in locating underground water leaks. Conventional walking surveys, where a technician moves along a route listening with a ground microphone, come in as low as 45 percent. That gap is the entire business case, and it is unusually easy to translate into money.

Every unnecessary excavation costs equipment, labour, road reinstatement and traffic disruption. Halving the number of dry digs changes a utility's cost structure without changing anything else about how it operates.

The problem is finding, not fixing

Water utilities are generally good at repairing leaks. The expensive, slow part is knowing where they are. A distribution network is buried, largely unobservable, and most leaks never surface — they drain quietly into the subsoil for months or years.

The industry term for this loss is non-revenue water: water that is produced and pumped but never billed, because it leaked, was mis-metered, or was taken without measurement. It is a standard operational metric worldwide, not an unusual condition, and reducing it is a normal part of utility management.

How the AI approach differs

Two techniques do most of the work, and they answer different questions.

Acoustic classification deals with sound. Sensors placed on the network continuously record the acoustic signature of the pipe. A leak generates a characteristic noise, but so does a passing truck, a pump, a partially closed valve and ordinary consumption. A trained classifier separates the leak signature from that background — which is exactly the task humans do inconsistently, because it depends on the operator's experience and attention on the day.

Failure prediction deals with probability. A model trained on pipe material, diameter, installation year, soil conditions, operating pressure and incident history estimates which segments are most likely to fail next. This shifts maintenance from reactive to planned, which is almost always cheaper.

The two combine well: prediction narrows where to look, acoustics confirms what is actually there.

A realistic implementation path

Confirm the data that already exists

Three things matter most: a network map with pipe material and diameter, several years of incident history, and inlet pressure and flow data by district. Most utilities hold more of this than they expect, though often across disconnected systems. This same dataset can later form the infrastructure layer of a broader city model — see Urban digital twins for city decision-making.

Choose one district, not the whole network

A single district metered area with reasonably reliable inlet measurement is the right unit. It is large enough for the result to be meaningful and small enough to instrument and finish. The same narrow-scope reasoning applies in AI-based traffic management.

Measure the baseline first

Non-revenue water in the pilot district, measured over a defined period before anything is installed. Without it, no improvement can be demonstrated and the project ends in argument rather than evidence.

Define success in operational terms

Not the number of leaks found. The measures that matter are the reduction in non-revenue water in that district, and the number of excavations avoided. Both convert directly into currency, which makes reporting straightforward and makes the case for expansion self-evident.

Train the people who will operate it

The most durable part of the project is the knowledge that stays inside the organisation. Operations staff need to read model output, prioritise alerts, and — importantly — recognise the conditions under which the model is not reliable.

A short AI training course built around this specific application, together with an AI consultant working alongside the team through the first months, is usually what separates a system that gets used from one that only produces reports nobody opens.

Frequently asked questions

Does this require replacing existing meters?

Generally no. The approach builds on district metering and acoustic sensors placed on the existing network. Meter replacement is a separate programme with its own economics and should not be bundled into a detection pilot.

What if the network map is incomplete or out of date?

This is the normal starting condition rather than a blocker. A pilot can begin with the best available map, and one useful side effect of the work is that it surfaces and corrects mapping errors in the pilot district.

How is this different from what technicians already do?

It changes coverage and consistency. A survey team can only cover a limited length of network per shift, and detection quality varies with individual experience. Fixed sensors monitor continuously and apply the same classification standard at three in the morning as at midday.

The short version

The technology here is mature rather than experimental, and the metric it improves — non-revenue water — is already tracked by most utilities, so the result is measurable from day one. Starting does not require a large programme: one district, one baseline figure and a pilot of a few months is enough to establish what this approach is worth for a given network.

Qatreh is based at the Alborz Science and Technology Park in Karaj and works with organisations across:

Data science and predictive infrastructure modelling — network data analysis, failure prediction and predictive maintenance
AI consulting for business — selecting the process with the strongest return and designing the pilot around it
AI training courses and AI consulting for education — building capability among operational staff
Computer vision and robotics — automated facility monitoring and image processing
AI automation and customer-service chatbots

If you work in water network management and want to know which district to start with, talk to us.