Fast and Accurate Water Leak Detection in Distribution Networks Using Deep Anomaly Models

Researchers from Politecnico di Milano propose a data-driven water leak detection method that treats leaks as anomalies in pressure data, using a convolutional autoencoder for feature extraction and a one-class SVM trained only on leak-free conditions. Tested on multiple simulated water distribution networks, the approach achieves high detection accuracy, faster leak identification, and strong robustness and generalization compared with existing methods.

Researchers from the Politecnico di Milano, Italy, examine one of the most persistent problems facing urban infrastructure systems: undetected water leaks in water distribution networks (WDNs). Framed within the broader pressures of climate change, population growth, and aging infrastructure, the paper emphasizes that water leakage is both an environmental and economic crisis. Globally, nearly one-third of treated urban water is lost before reaching consumers, driving up operational costs, wasting energy, and exacerbating water scarcity. Against this backdrop, the authors argue that rapid and reliable leak detection is no longer optional but essential for sustainable water management. They position data-driven monitoring approaches as a promising alternative to traditional hydraulic or hardware-based solutions, which often require detailed pipe-level information and are difficult to scale in real-world networks.

Rethinking Leak Detection as Anomaly Detection

Rather than modeling leaks directly, the paper introduces a conceptual shift: treating leaks as anomalies in pressure behavior. The proposed method relies exclusively on pressure measurements collected at network nodes, minimal knowledge of network topology, and historical data from normal, leak-free operation. This design choice addresses a key limitation of many existing methods, namely, the scarcity of labeled leak data. By learning what “normal” looks like and flagging deviations from it, the system can detect leaks without ever being trained on leak examples. This makes the approach especially suitable for operational settings where leaks are rare, unpredictable, and difficult to document in advance.

Learning Normal Network Behavior with Deep Features

At the heart of the proposed pipeline is a convolutional autoencoder trained solely on no-leak data. Pressure readings from all nodes are continuously collected and organized into sliding time windows spanning one full week. This window length is chosen to capture daily and weekly pressure patterns while remaining robust to slower seasonal trends. Each window forms a matrix that preserves both spatial information across nodes and temporal evolution over time.

The autoencoder learns to reconstruct these pressure matrices, forcing its internal encoder to compress them into a compact, informative representation. After training, only the encoder is retained and used as a feature extractor. Through stacked one-dimensional convolutional layers, the encoder transforms high-dimensional pressure data into low-dimensional embeddings that summarize normal network dynamics. This feature extraction step is critical, as it reduces complexity and enables effective anomaly detection on large-scale networks.

One-Class SVM for Timely Leak Detection

The extracted embeddings are analyzed using a one-class Support Vector Machine (SVM), a model designed specifically for anomaly detection when only normal data are available. The one-class SVM learns a boundary that encloses the feature representations of leak-free operation. During deployment, new pressure windows are encoded and evaluated against this boundary. Embeddings that fall outside it are interpreted as anomalies and flagged as potential leaks.

To improve temporal stability, the anomaly scores are smoothed using a moving average filter, and a threshold is applied to generate a binary leak indicator. When this indicator crosses the threshold, a leak is declared, and the corresponding time is recorded as the estimated leak start. This design allows the system not only to detect the presence of a leak but also to estimate when it began, a capability that is crucial for minimizing water loss and repair costs.

Strong Results, Robustness, and Generalization

The method is evaluated through extensive simulations on the Modena water distribution network, a realistic benchmark with 268 nodes and 317 pipes. Using EPANET, the authors generate 500 scenarios covering both leak and no-leak conditions, with randomized leak locations, sizes, and pipe parameters. The results show a detection accuracy of 92 percent and an average detection delay of just over 40 hours, significantly outperforming prior deep-learning-based approaches.

The study also examines robustness to noise by injecting Gaussian noise at different signal-to-noise ratios and finds that the method remains effective at moderate noise levels. Importantly, the authors test generalization by applying the trained model to entirely different networks, including the Hanoi and Pescara systems. Even without retraining the feature extractor, the method maintains solid performance, demonstrating its adaptability across diverse network topologies.

Overall, the paper presents a compelling, scalable solution for modern water infrastructure monitoring. By combining deep feature learning with anomaly-based detection, it advances the state of the art in leak detection while laying the groundwork for future extensions such as real-world deployment, leak localization, and enhanced noise handling.