Artificial Intelligence and Data-Driven Leakage Management for Sustainable and Resilient Water Distribution Systems: A Critical Review

Water is the lifeblood of a modern city, flowing through a vast, hidden network of pipes that delivers it to homes, schools, and businesses. Yet, this system is constantly under siege from a silent thief: leakage. When pipes crack or joints fail, treated water escapes into the ground before it ever reaches a tap. This loss is not just a matter of wasted water; it represents a massive drain on energy used to pump and treat that fluid, increases the cost of service, and can weaken the structural integrity of the entire network. For decades, finding these leaks has been a difficult game of chance, often relying on workers walking the streets listening for hissing sounds or waiting for a pipe to burst completely. The challenge is that leaks can be tiny, buried deep underground, and their signals are easily drowned out by the normal noise of daily water use.

In response to this persistent problem, a new wave of technology has emerged, promising to turn the water network into a smart, self-aware system. This approach relies on artificial intelligence and data-driven models to listen to the pipes, analyze patterns, and pinpoint exactly where water is escaping. Researchers are now teaching computers to recognize the subtle fingerprints of a leak, whether it is a sudden burst or a slow, steady drip. However, as these technologies advance, a critical question remains: do the sophisticated algorithms that work perfectly in computer simulations actually hold up when faced with the messy, noisy reality of a real city?

A comprehensive review by researchers at the University of Ottawa and the National Research Council of Canada sets out to answer this question. The team did not build a new machine or test a single new sensor. Instead, they acted as investigators of the research itself, gathering and analyzing dozens of studies published between 2010 and 2026. Their goal was to map the entire landscape of how artificial intelligence is being used to manage water leaks, to see what methods work best, and to identify where the field is still falling short. They examined everything from simple statistical tools that look for unusual changes in water flow to complex neural networks that mimic the human brain to process sound and vibration data.

The review reveals that there is no single "magic bullet" algorithm that solves every leakage problem. The effectiveness of a computer model depends entirely on the context in which it is used. For instance, a method that excels at detecting a sudden, loud pipe burst using pressure sensors might fail completely when trying to find a slow, quiet leak using acoustic sensors. The researchers found that while many studies report incredibly high success rates, these numbers often come from controlled environments. In computer simulations, where the data is clean and the rules are perfect, algorithms can achieve near-perfect accuracy. But when these same models are tested against real-world data from actual water networks, their performance often drops. The real world introduces variables that simulations struggle to capture: the varying materials of pipes, the background noise of traffic and construction, and the unpredictable changes in how much water people use at different times of the day.

One of the most significant findings of the review is the gap between the laboratory and the field. Many of the most promising studies rely on data generated by computer models of water networks, such as the widely used EPANET software. While these simulations are valuable for testing ideas, they cannot fully replicate the complexity of a real city. The researchers noted that studies using real-world data, such as recordings from actual acoustic sensors in cities like Hong Kong or Singapore, tend to show more modest but more reliable results. These real-world tests reveal that noise and sensor placement are critical factors; a model might work well in one neighborhood but fail in another simply because the soil or pipe material changes how sound travels.

The review also highlights a growing shift in how researchers are approaching the problem. Early efforts focused on simple detection—telling the difference between a leak and no leak. Today, the focus is expanding to include precise localization, which means finding the exact pipe or section where the leak is happening, and severity estimation, which determines how big the leak is. Newer techniques are emerging, such as using "digital twins," which are virtual replicas of the physical water network that update in real time with sensor data. These twins allow engineers to test scenarios and predict where leaks might occur before they happen. There is also a rise in hybrid approaches that combine the laws of physics with machine learning, ensuring that the computer's predictions make sense according to the known behavior of water flow.

Despite these advances, the researchers identified several persistent gaps that prevent these technologies from being widely adopted by water utilities. A major issue is the lack of standardized testing. Because every study uses different data, different sensors, and different ways of measuring success, it is difficult to compare one method to another. Furthermore, many studies do not report enough detail about how they prepared their data or tuned their models, making it hard for other scientists to repeat the work. Another critical challenge is the "class imbalance" problem: in a typical water network, leaks are rare events compared to normal water flow. This means a computer could be trained to simply guess "no leak" every time and still appear to be 99% accurate, while missing every single actual leak. The review argues that success should be measured by how well a system finds the rare leaks and how few false alarms it raises, rather than by a simple overall accuracy score.

The authors also look toward the future, noting the potential of generative artificial intelligence and large language models to help operators make decisions. These tools could act as intelligent assistants, pulling together data from sensors, historical records, and maintenance logs to suggest the best course of action. However, the review cautions that these advanced systems must be used carefully. Unlike a simple detector that just sounds an alarm, a decision-making system must be grounded in physical reality to avoid "hallucinating" problems that do not exist or suggesting actions that could damage the network. The ultimate goal is not just to build smarter algorithms, but to create a reliable system that utilities can trust to save water, reduce energy costs, and keep the infrastructure resilient.

In the end, the review concludes that while artificial intelligence offers powerful new tools for managing water leaks, the technology is not yet a finished product. The path forward requires a move away from isolated experiments in simulated environments and toward rigorous testing in real cities. It demands a focus on transparency, where researchers clearly report how their models perform under difficult conditions and how they handle the inevitable noise of the real world. By connecting these advanced analytical methods with the practical needs of water utilities, the field can move toward a future where water distribution systems are not only efficient but also truly sustainable and resilient. The technology is ready to learn, but it must be taught with the same care and realism that engineers apply to the physical pipes themselves.