How AI is learning to hear water disappear
Western and Southern Europe, parts of East Asia, North Africa, and sections of North America experienced an unusually hot and dry summer in 2026.
Households were asked to leave cars unwashed and stop watering gardens from the mains. Farmers faced irrigation restrictions as dry soils and low river levels damaged crops. Conditions have since improved in some places, but groundwater, reservoirs and ecosystems can take much longer to recover than surface soils and smaller rivers.
Yet even as consumers and farmers are urged to use less, vast quantities of treated water continue to disappear beneath the streets.
A hidden leak rarely announces itself with a fountain bursting through the asphalt. Water may drain quietly into the soil or find its way into a sewer, leaving no puddle or collapsed road. It can continue for weeks before somebody notices lower pressure or the damage finally reaches the surface.
Acoustic AI water leak detection offers another way to find it. Sensors attached to pipes, valves and hydrants record the vibrations created by escaping water. Machine-learning software then tries to separate the acoustic signature of a leak from the ordinary noise of a working city.
That technology received a major vote of confidence in southern England this summer. Millions of households faced temporary restrictions on using hosepipes, while water companies searched for leaks in ageing networks that lose hundreds of millions of litres every day.
In August, Thames Water said it had expanded its network of acoustic sensors from 21,000 in 2025 to more than 75,000. The announcement followed the driest July in England since records began approximately 190 years ago.
Thames Water says the denser sensor network has nearly doubled the accuracy of its leak detection and helped it find 88 per cent more leaks in July 2026 than during the same month a year earlier. The company also reported that it was repairing more than 1,000 leaks a week.
Separately, the utility is using an artificial-intelligence system called Origin Orbit to analyse satellite data for signs of underground water. A 19-week trial identified more than 800 leaks, which Thames Water estimates were collectively losing 8.7 million litres a day.
The two technologies approach the same problem from opposite directions. Satellites look down for changes in moisture, temperature and soil. Acoustic sensors listen through pipes and the surrounding ground.
Together, they are giving water networks a way to report that they are in trouble before the street caves in.
How acoustic water leak detection works
Even when a leak is invisible, it is rarely silent.
Water inside a municipal pipe is under pressure. When it squeezes through a crack, damaged joint or faulty seal, it creates vibrations that can travel through the pipe wall, the water inside it and the surrounding ground.
Engineers have listened for them for decades. A technician might place a listening stick against a valve or use a ground microphone to search for the characteristic hiss of escaping water.
Acoustic correlators compare recordings taken at two points along a pipe. If the noise reaches one sensor slightly earlier than the other, software can estimate where between them it began.
It is a proven technique, but a city’s water network is considerably noisier than it appears from the pavement.
Pumps start and stop. Valves move. Customers open taps. Trains rumble past and roadworks shake the ground. Somewhere within that underground orchestra may be the quieter, steadier note of a leak.
Acoustic AI is being developed to pick it out.
Small Internet of Things (IoT) sensors are attached to hydrants, valves, meters or accessible sections of pipe. Some detect vibrations through the pipe wall. Others use hydrophones; these are underwater microphones that listen directly to the water inside.
Software breaks the recordings into patterns of frequency, intensity and time. A lorry passing overhead produces a short burst. A pump may switch on at regular intervals. A leak generally creates a more persistent signal.
A machine-learning model trained on recordings of known leaks can estimate whether a new sound has the same acoustic fingerprint. More advanced systems may filter out traffic and machinery, calculate how much water is escaping or rank suspected leaks by urgency.
What does AI add?
Not every digital listening system qualifies as AI. Noise loggers and acoustic correlators existed long before the current boom.
Machine learning adds the ability to classify complicated signals rather than simply record them or calculate where they originated. Some platforms also combine sound with pressure sensors, flow measurements and digital maps.
A faint noise becomes more convincing if overnight water flow is simultaneously rising and a nearby pressure sensor has recorded an unusual event. It is a little like a doctor considering several symptoms instead of making a diagnosis from one.
“AI leak detection” can still describe very different systems, from genuinely trained audio models to established correlation technology sold under a more fashionable label.
Three ways to teach pipes to listen
Oxford-based FIDO Tech has developed deep-learning models to analyse recordings from water networks. Its software is designed to recognise leaks and estimate their size, helping utilities decide which repairs should come first.
One FIDO sensor found a leak at a fire hydrant in San Tan Valley, Arizona. From the street, nothing looked wrong. Yet a faulty seal was allowing between three and seven US gallons to escape every minute; equivalent, according to the project partners, to the normal daily water use of about 43 households.
Without continuous monitoring, the problem might have remained hidden until the hydrant’s next maintenance inspection, potentially years later.
Utility company EPCOR installed 4,554 FIDO sensors across its 160-square-mile San Tan service area. EPCOR reported that non-revenue water fell from 27 per cent to around 10 per cent, partly because of the deployment, and that more than 250 leaks were identified in a year.
The figures come from the companies involved rather than an independent controlled study, and FIDO was only one part of EPCOR’s water-loss programme. They nevertheless show what utilities hope to gain from continuous water-network monitoring.
Canadian company Digital Water Solutions listens from inside the network. Its hydrant.AI devices contain hydrophones and also collect pressure and temperature data.
During a six-month pilot in Ontario, ten hydrants were fitted with the technology in an area served entirely by plastic pipes. The system identified two service-line leaks that were later verified in the field.
One involved a crack about 3.8 centimetres long. The company estimated that it had allowed approximately 3.8 million litres to escape between mid-November 2022 and its repair in January 2023.
The acoustic signature was present before a resident reported the problem and disappeared after the repair, although the company’s case study does not say that the alert itself prompted the work.
Swiss specialist Gutermann has meanwhile trained its ZONESCAN AI model on a proprietary collection of real pipe recordings classified by people. The model gives a suspicious sound a probability score, allowing utilities to investigate the strongest candidates first.
Companies including Aquarius Spectrum and Mueller Water Products’ Echologics division are developing related systems. What was once specialist equipment carried around by technicians is gradually becoming a permanent sensory layer spread across cities.
Why plastic pipes are harder to monitor
Acoustic leak detection still faces some stubborn physical limits.
Sound generally travels farther through rigid metal pipes than through plastic, which absorbs more vibration. This can make leaks harder to locate in newer networks using polyvinyl chloride or high-density polyethylene.
Bigger leaks are not necessarily easier to hear. Water escaping through a narrow crack may produce a sharp, higher-frequency hiss. A larger opening can generate deeper sound that is less obvious to the human ear and harder for some instruments to isolate.
Results also depend on sensor spacing, water pressure, pipe diameter, soil and the accuracy of the utility’s network map. A model trained primarily on metal pipes beneath one city may not work equally well on a plastic network elsewhere.
One US government-backed demonstration tested an AI system that listened inside the water column. During its first phase, the model classified previously unseen simulated leaks with an overall accuracy of approximately 94 percent. After changes to its data processing, it detected all test leaks in the second phase.
Finding their exact position proved more difficult. The system could generally identify the correct pipe section, but uncertainty about pipe layouts and characteristics, along with imperfect sensor synchronisation, reduced localisation accuracy.
Detecting every leak in one stage of a controlled demonstration does not mean a system will find every leak across a city. Utilities need to know how many alerts prove genuine, how many leaks remain unheard and whether performance holds across seasons and pipe materials.
False alarms send crews on unnecessary investigations. Missed leaks continue wasting water. A system only becomes useful when it controls both risks in a working network.
Can acoustic AI detect water theft?
Not all missing water escapes into the ground.
Some reaches users but is never billed because of inaccurate meters, incomplete records, tampering or unauthorised connections. These are known as apparent losses.
Physical leaks, apparent losses and authorised but unbilled consumption together make up non-revenue water: water that enters a distribution system but produces no income for the utility.
An illegal connection could produce sound when someone cuts into a pressurised pipe. It may also create unexplained flow in part of the network. But a sensor cannot determine criminal intent from a vibration.
The same disturbance could come from a cracked service pipe, legitimate maintenance or a moving valve. Unusual consumption might indicate theft, but it could equally be caused by a faulty meter.
Acoustic data can narrow the search, but detecting unauthorised water consumption also requires meter readings, pressure and flow data, customer records and a physical inspection. AI cannot listen to a pipe and pronounce someone guilty.
Finding a leak is not fixing it
Even a perfectly located leak cannot repair itself.
A utility must decide whether the volume being lost justifies immediate action, arrange access, divert traffic, expose the pipe and complete the repair. Afterwards, somebody has to put the road back together.
Discovering hundreds of genuine leaks can therefore create a backlog of its own. A dashboard covered in red dots is only useful if the utility has enough people and money to do something about them.
This is where estimating leak size becomes almost as important as locating it. If AI can show that one pipe is losing thousands of litres an hour while another has a slow seep, repair teams can tackle the larger loss first.
As such, the technology is not removing people from water maintenance. It is helping them decide where their limited time will save the most water.
Back in San Tan Valley, the leaking hydrant did not erupt or flood the street. It simply released treated drinking water into the soil, minute after minute, without giving anyone a reason to look down.
Acoustic AI supplied that reason. The repair still required a crew, tools and money. But the pipe no longer had to fail visibly before the utility knew it was in trouble.
Liked this article? You can support our independent journalism via our page on Buy Me a Coffee. It helps keep MoveTheNeedle.news focused on depth, not clicks.