In a resource-constrained world, water companies are already using AI to target the biggest and most wasteful leaks for repair. Leaks are just the just the first step on the road to ending water scarcity. AI is taking intelligence-led decision-making to the next level.
Advances in artificial intelligence (AI) mean that the technology will ultimately play a role in all elements of the water use cycle. The systems involved in the distribution, consumption, and collection and treatment of water and wastewater naturally lend themselves to being structured or processed using AI or an intelligent algorithm.
Historically these systems have all be run by people: an approach that has worked very well. In a risk-averse industry with a long legacy and a legal requirement to deliver safe drinking water, there is some resistance to change how risk is managed.
Where a system is not under stress, this is not necessarily an issue. The service can still be delivered in the same way. The challenge is that as physical demand starts to grow, the environment becomes more stressed. Withdrawing water becomes more difficult and losses start to increase, and operators have to make better use of the available resource.
“This is where AI is becoming part of the story.”
With business as usual no longer an option some action is needed to correct that trajectory. Optimising those systems requires a combination of technology – monitoring systems, Internet of Things (IoT) equipment and some systems that make decisions based on that data to get the best results. Despite resistance to change, the reality is that a system that understands the parameters will make far better decisions than any human ever will.
Listening for leaks
Listening for leaks has long been part of the water company arsenal for resource conservation and reducing non-revenue water losses. Digital acoustic technology has been around for 30 years in the water industry. Using acoustics in water dates back to Roman times when people used sticks to listen to water pipes. It’s still a manual process in lots of cases, but the technology now available is simply more sensitive than human hearing and across a wider range of frequencies.
The difficulty is there is a cacophony of noises on the network coming from pumping systems, valves, hydrants, leaks, air conditioning systems and so on. The key to the use of acoustics in managing water systems more effectively is not just picking up those noises but characterising them and identifying leakage. That is where AI is becoming part of the story.
While a good human engineer might recognise particular noises as leaks to a degree of accuracy for a few hours a day, AI systems are relentless. They can take this mass of data and process it quickly, spitting out results in a few milliseconds. It does that 24 hours a day. Taking all the various noises, classifying them and then training an AI system to recognize them is at the core of the unique FIDO approach.
Unlike conventional smart leak detection technologies, FIDO’s unique deep learning algorithm is a totally sensor-agnostic technology. The first step was to get data and FIDO worked with lots of companies to get a wealth of acoustic data and then train the Ai algorithm to process and characterise it.
“FIDO AI is able to rank leaks within a distribution network by size, even in the face of multiple sources of background noise.”
To do this, the company evolved several techniques, such as adding noise to develop different versions of the data set and then applied different AI modelling techniques. Their AI model can request a data set, consider it and produce an output to test, before selecting another set of data to retrain itself across multiple data sets and using multiple AI models all working together. Each one continues to optimise and tune itself over time.
Advancing the model
Identifying leaks is just the first step though. There is in fact a very good correlation between the noises made and the size of any particular leak. Advancing the model now allows it to identify a leak, it’s location and its size too.
FIDO AI can rank leaks within a distribution network by size, even in the face of multiple sources of background noise. It is also able to make this distinction regardless of the materials used throughout the network, such cast iron, polypropylene, even stainless steel.
In a resource-constrained business like a water network operator, in most cases it is not commercially viable to repair every single leak. The FIDO model is detecting water leaks and leak size in real time from any audio or vibration file, with more than 92% accuracy. This gives a leak ranking capability that can determine if a leak is worth fixing immediately or if it is changing over time, getting bigger for example.
Water distribution companies are then able to target their investment at the most costly problems first by addressing the biggest leaks. And, by identifying the biggest leaks the AI system can perhaps halve the number of repairs undertaken while still achieving the same results in terms of non-revenue water losses.
Use less, lose less, make more
The total coverage of acoustic sensors across water networks globally is not that high currently – probably covering less than 5% of the network. However, a lot is possible from the data already available. As more data is gathered across the US, Europe and elsewhere there are more kinds of noises to explore and analyse making models ever more precise.
The AI approach is also evolving though and is now on the threshold of another advance. By analysing the noise of the water as it exits the pipe, AI is able characterise the leak still further by identifying the different types of leak – crack, pin hole or joint failure for example.
“In the next stage of AI analysis, these sophisticated models are starting to characterize entire networks.”
Characterizing those types of different leaks also allows the AI model to identify those which are expected to get worse over time. Recognising the type of leak supports more accurate data-led decision making and therefore better targeting of repair team resources.
The AI model also keeps the history of faults so also knows if there had been 10 failures on a short section over the last six months, say. By taking a holistic view of the data and understanding the full picture, AI can also start to direct the type of repair, replacing an entire section of failure-prone pipework rather than making repeated repairs within the same area.
AI systems in the future
This hints at further advances that are on the horizon as AI moves from real-time leak detection and into the future. In the next stage of AI analysis these sophisticated models are starting to characterize entire networks. By understanding what materials and systems are being used in the network, and that particular elements are likely to fail over a certain period of time depending on the weather and other influential factors, AI is able to rank the risk of failure for a particular section of network.
With that data and smart analytics it is possible to pinpoint the more likely areas where problems arise. On a 20,000 km network being able to narrow down the target areas for leak detection is a big advantage. That’s the start of a steering process that is a level above looking for actual current leaks and instead looks for problems in advance. It helps enhance deployment of acoustic detectors, for example, and is another route to targeting investment for maximum effect.
AI can offer far more than leak detection when it comes to curbing water losses though. Non-revenue water (NRW), also includes water theft. This can be a major issue, one recent example saw around 65% of a water utility’s lost water was from leaks, the other 35% was from water theft. By analysing the water usage data and consumption patterns, AI could potentially identify where water is being drawn from the network but there is no corresponding revenue stream and target water theft as well as leaks. For AI the key is all data – long-term partnerships between utilities and FIDO will mean more data becomes available that can deliver actionable AI. This will eventually help deliver insights into a multitude of areas of the network that are currently opaque.
Currently the industry is collecting data, understanding it and making intelligence-led deployment decisions. Yet, the ultimate solution is a fully automated system. In the idealised future an AI system will be able to alert operators to a leak, locate it, rank its urgency, notify customers, shut down the relevant section of the network, schedule the work and then send out some kind of in-pipe robot ROV that can swim to the location, fix it and swim back again.
It’s a vision of the future but the mission is to rid the world of water scarcity. The technology is starting to evolve towards that closed loop cycle. Characterizing leaks is the first step.
Neil Edwards is the CTO of FIDO Tech.