This thesis investigates the problem of leak detection and localization in water distribution networks using a two-stage deep learning framework. Traditional approaches—such as rule-based diagnostics or statistical analysis—are often limited by their scalability, adaptability, and reliance on handcrafted thresholds. To address these challenges, we propose a two-stage architecture that decouples temporal and spatial reasoning: (1) a Multi-Layer Perceptron (MLP) is used to perform real-time leak detection by analyzing sliding windows of pressure and flow sensor data; and (2) a Gated Graph Neural Network (GGNN) is then applied to localize the leak node using graph-structured data derived from the network topology. Given its central role in spatial leak localization, we further enhance the GGNN by incorporating gated recurrent units (GRUs) within the message passing process to capture both localized disturbances and global propagation effects caused by leaks. To assess the impact of architectural choices, we evaluate several GGNN variants—including GNN-Res, GNN-NP, and GNN-LSTM—to study the trade-offs between model complexity and localization performance. Experiments on a simulated 272-node water distribution network with 100 diverse leak scenarios demonstrate that the proposed pipeline achieves high accuracy across both tasks: the MLP detects leak onset with low false positives, while the GGNN achieves Top-1 localization accuracy of 79.4%, Top-5 accuracy of 94.3%, and AUC of 0.9435. The framework also shows robustness to sensor noise, data sparsity, and topological variation. Overall, this work introduces an interpretable and scalable framework for real-time leak detection and localization, offering promising applications in smart water infrastructure monitoring and fault response systems.Future work will explore real-world deployment, integration with control systems, and broader adaptability to varying network conditions.
Questa tesi affronta il problema della rilevazione e localizzazione delle perdite nelle reti di distribuzione idrica mediante un framework di apprendimento profondo a due stadi. Gli approcci tradizionali, come le diagnosi basate su regole o le analisi statistiche, risultano spesso limitati in termini di scalabilità, adattabilità e dipendenza da soglie definite manualmente. Per superare tali limitazioni, proponiamo un’architettura a due stadi che separa il ragionamento temporale da quello spaziale: (1) un Multi-Layer Perceptron (MLP) viene utilizzato per rilevare in tempo reale l’insorgenza delle perdite, analizzando finestre temporali dei dati di pressione e portata acquisiti dai sensori; (2) una Gated Graph Neural Network (GGNN) viene successivamente applicata per localizzare il nodo interessato dalla perdita, sfruttando dati strutturati a grafo derivati dalla topologia della rete. Data la sua rilevanza nella localizzazione spaziale, la GGNN è stata ulteriormente potenziata integrando Gated Recurrent Units (GRU) all’interno del processo di propagazione dei messaggi, al fine di catturare sia disturbi locali sia effetti di propagazione globali causati dalle perdite. Per valutare l’impatto delle scelte architetturali, sono state analizzate diverse varianti della GGNN—tra cui GNN-Res, GNN-NP e GNN-LSTM—allo scopo di studiare i compromessi tra complessità del modello e accuratezza della localizzazione. Gli esperimenti condotti su una rete idrica simulata composta da 272 nodi e 100 scenari di perdita differenti dimostrano che il framework proposto raggiunge un’elevata accuratezza in entrambi i compiti: l’MLP rileva tempestivamente l’insorgenza delle perdite con un basso tasso di falsi positivi, mentre la GGNN ottiene un’accuratezza Top-1 pari al 79,4%, una Top-5 pari al 94,3% e un AUC di 0,9435. Il sistema si dimostra inoltre robusto rispetto al rumore dei sensori, alla scarsità dei dati e alle variazioni topologiche. Nel complesso, questo lavoro propone un framework interpretabile e scalabile per la rilevazione e localizzazione in tempo reale delle perdite, con promettenti applicazioni nei sistemi intelligenti di monitoraggio delle infrastrutture idriche e nella gestione dei guasti. I futuri sviluppi riguarderanno la sperimentazione in contesti reali, l’integrazione con sistemi di controllo e l’adattamento a condizioni operative più eterogenee.
Water pipeline leak detection and localization
NIU, ZIXIN
2024/2025
Abstract
This thesis investigates the problem of leak detection and localization in water distribution networks using a two-stage deep learning framework. Traditional approaches—such as rule-based diagnostics or statistical analysis—are often limited by their scalability, adaptability, and reliance on handcrafted thresholds. To address these challenges, we propose a two-stage architecture that decouples temporal and spatial reasoning: (1) a Multi-Layer Perceptron (MLP) is used to perform real-time leak detection by analyzing sliding windows of pressure and flow sensor data; and (2) a Gated Graph Neural Network (GGNN) is then applied to localize the leak node using graph-structured data derived from the network topology. Given its central role in spatial leak localization, we further enhance the GGNN by incorporating gated recurrent units (GRUs) within the message passing process to capture both localized disturbances and global propagation effects caused by leaks. To assess the impact of architectural choices, we evaluate several GGNN variants—including GNN-Res, GNN-NP, and GNN-LSTM—to study the trade-offs between model complexity and localization performance. Experiments on a simulated 272-node water distribution network with 100 diverse leak scenarios demonstrate that the proposed pipeline achieves high accuracy across both tasks: the MLP detects leak onset with low false positives, while the GGNN achieves Top-1 localization accuracy of 79.4%, Top-5 accuracy of 94.3%, and AUC of 0.9435. The framework also shows robustness to sensor noise, data sparsity, and topological variation. Overall, this work introduces an interpretable and scalable framework for real-time leak detection and localization, offering promising applications in smart water infrastructure monitoring and fault response systems.Future work will explore real-world deployment, integration with control systems, and broader adaptability to varying network conditions.| File | Dimensione | Formato | |
|---|---|---|---|
|
2025_07_ZIXIN NIU_Thesis.pdf
accessibile in internet per tutti
Descrizione: 2025_07_ZIXIN NIU_Thesis
Dimensione
1.93 MB
Formato
Adobe PDF
|
1.93 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/240349