Moving forward in water distribution network leak identification through an innovative features engineering step

In Italy an average of 40% of the water in transit in Water Distribution Networks (WDNs) is lost, which represents a significant economic loss, and, because of the corresponding energy waste, also a considerable environmental damage. Over the years various methods have been developed for detecting a...

Celý popis

Uloženo v:

Podrobná bibliografie
Vydáno v:	Digital signal processing Ročník 168; s. 105603
Hlavní autoři:	Damonti, Elvio, Bernasconi, Giancarlo
Médium:	Journal Article
Jazyk:	angličtina
Vydáno:	Elsevier Inc 01.01.2026
Témata:	CNN Overcomplete Autoencoder inference errors analysis Convolutional neural network (CNN) Overcomplete Autoencoder application Leak identification in Water Distribution Networks (WDN) Machine learning methods for water leak detection and localization Water pressure data features engineering Water pressure data features engineering Convolutional neural network (CNN) Overcomplete Autoencoder application Machine learning methods for water leak detection and localization CNN Overcomplete Autoencoder inference errors analysis Leak identification in Water Distribution Networks (WDN)
ISSN:	1051-2004
On-line přístup:	Získat plný text
Tagy:	Přidat tag Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!

Popis
Shrnutí:	In Italy an average of 40% of the water in transit in Water Distribution Networks (WDNs) is lost, which represents a significant economic loss, and, because of the corresponding energy waste, also a considerable environmental damage. Over the years various methods have been developed for detecting and locating leaks in WDNs, and, in particular, several algorithms have been implemented that use Convolutional Neural Networks (CNNs). They are all based on a training phase run on a representative subset of the WDN data, and the main differences between the various implementations are in the data pre-processing and in the CNN configuration. This paper proposes a new fully Data-Driven approach, where a preliminar features engineering step, performed by a visual analysis of specific data patterns, both in the time domain and in the Fourier domain, allowed us to conceive and identify two paramount features engineering steps: a new effective data pre-processing algorythm and a new configuration of CNN that uses an Overcomplete Autoencoder (Overcomplete AE) topology with residual blocks. These two steps, described in detail in this paper, allowed us to better highlight and identify the anomalies caused by leaks in WDN pressures time series and they permitted, in association with a new original automatic analysis of the reconstruction error made by the Autoencoder, to achieve results that are on the top of the current state of art. Specifically, the whole innovative method is presented in detail exploiting publicly available data, so to be easily reproducible, and, more specifically, for this purpose, the benchmark was run on a synthetic 'LeakDB' dataset of 500 scenarios and the outcomes were then validated through different data obtained from a second and more complex synthetic 'LeakDB' dataset containing 1000 scenarios. Both these datasets are related to a District Metering Area (DMA) of the Hanoi WDN and both are publicly available.
ISSN:	1051-2004
DOI:	10.1016/j.dsp.2025.105603