Why generic AI struggles with infrastructure, and what actually works for water networks

Artificial intelligence is everywhere right now. From chatbots to image generators, the term “AI” has become shorthand for a wide range of technologies, many of which have little relevance to the practical challenges of operating infrastructure.

For water utilities, the objective is not novelty. It is better decisions about complex physical systems.

Yet much of the current conversation around AI is dominated by tools designed to interpret language, generate images, or automate digital workflows. These technologies can be powerful in the right context, but they are not designed to understand the behaviour of buried pipe networks.

Detecting emerging leaks, interpreting acoustic signals, and understanding changes in network behaviour require a very different kind of intelligence.

These challenges have been the focus of more than a decade of collaborative research between water utilities and researchers, including the peer-reviewed studies underpinning Adelitics’ analytical methods.

Infrastructure is a signal interpretation problem

Water networks generate large volumes of data. Acoustic sensors, fibre-optic monitoring systems and operational telemetry all capture signals that reflect what is happening inside the network.

Within those signals may be early indicators of:

• emerging leaks

• changes in pipe condition

• operational events such as pumps or valves

• background noise from normal system activity

Interpreting these signals is challenging because water networks are complex and dynamic environments. Pipe materials behave differently, demand fluctuates throughout the day, and acoustic conditions vary across the network.

The challenge is not collecting more data. It is understanding which signals matter and what they mean.

Machine learning provides a systematic way to interpret complex sensor signals within operating networks.

Not all AI is designed for physical systems

Much of the recent attention around AI has been driven by large language models (LLMs) and other generative tools trained on enormous collections of internet text and images.

These systems are designed to recognise patterns in language and visual data.

Water networks, however, produce sensor signals, not sentences.

Understanding those signals requires machine learning approaches designed for time-series data, signal processing and physical system behaviour, rather than generative AI.

How machine learning is used in water networks

Modern monitoring systems collect acoustic and operational data from across a network. Within this data are patterns associated with different types of events.

Machine learning models can be trained to recognise these patterns when provided with large labelled datasets of real network behaviour.

At Adelitics, we have built bespoke supervised machine learning models using labelled acoustic event datasets developed over more than a decade of work with utilities and researchers.

These datasets include examples of confirmed leaks, operational noise and other acoustic events recorded within operating networks.

By learning how these signals behave in practice, machine learning models can help identify anomalies and emerging changes that may indicate developing problems.

Importantly, these models are trained on water network data, not generic datasets.

From anomalies to operational decisions

Detecting anomalies is only the first step.

Utilities operating large monitoring systems must also decide where to investigate first and how to allocate limited field resources.

To support this process, Adelitics combines the outputs of multiple machine learning models within a structured analytical framework.

This framework integrates model outputs with contextual network information to prioritise where utilities should investigate or intervene first.

Why domain-specific AI matters

Infrastructure systems are governed by physical laws, material behaviour and operational constraints.

Generic AI tools trained on internet-scale datasets cannot capture these realities.

Effective analytics for water networks require models trained on infrastructure data and informed by engineering understanding of how pipe systems behave under real operational conditions.

Adelitics has been developing and refining these models for more than ten years through collaboration with utilities and researchers. The resulting analytical framework reflects the realities of operating water networks rather than generic data science approaches.

Built for infrastructure, not hype

The current surge in interest around AI has led many technology products to adopt the label.

At Adelitics, our approach is intentionally more specific.

We do not rely on proprietary generative AI platforms or large language models. Instead, the platform uses our own machine learning models and analytical framework developed specifically for water network monitoring.

These models have been refined through field deployments, labelled datasets and peer-reviewed research.

The aim is simple: helping utilities convert monitoring data into clear, operationally useful insight that supports earlier intervention and extends asset life.

In infrastructure systems, the value of AI is not measured by the sophistication of the algorithm, but by whether it helps operators make earlier and better decisions in the real world.

Why this matters for utilities

When applied to real monitoring systems, domain-specific machine learning can help utilities:

• detect emerging anomalies earlier

• reduce time spent investigating false alarms

• prioritise inspection and repair activities

• extend the life of ageing pipe assets

The practical shift is from reactive repair after failure to earlier intervention guided by signal change.

Figure: Types of AI and where infrastructure analytics fits

Recent advances in AI have focused on language and image generation. Infrastructure monitoring relies on signal-based machine learning trained on labelled sensor data from real physical systems. Adelitics builds on this foundation with domain-specific models and the structured decision framework to prioritise actions that help prevent pipe failures.