Leakage detection in a real water distribution network through a federated prototype-based model

Abstract

This thesis explores important an intricate trade-off in intelligent systems: the efficiency of leakage detection tasks, the preservation of sensitive information, and the interpretability of the intelligent system. Through a comprehensive analysis of these interconnected domains, we aim to design a low-cost algorithmic solution that can efficiently learn from industrial data collected by distributed sensors while preserving sensitive information. We focus on the following topics: (i) efficient leakage detection on water distribution networks (WDNs); and (ii) formulation of federated prototype-based models. In particular, this thesis mainly focuses on proposing an efficient and low-complexity distributed solution for identifying potential leaks in water distribution networks in municipal areas while ensuring the privacy of the hydraulic data. To this end, we explore and extend existing theories and methods from prototype-based learning and federated learning. We consider a hydraulic dataset, which includes water pressure and flow measurements obtained from pumps within district-metered areas in Stockholm, Sweden. In the first part of this thesis, we have considered a traditional learning paradigm, which contemplates the existence of a server responsible to collect and process all samples from every device and build machine learning models. Our goal is validating our assumption that prototypebased models (PBMs) can efficiently detect leakages in water distribution networks. The feature extraction, PBMs and clustering validation techniques we propose have low complexity, and the numerical experiments show the potential of using machine learning strategies in leakage detection for monitored water distribution networks. Specifically, our experiments show that the proposed learning strategies are able to obtain correct classification rates up to 93.98%. In the second part of this thesis, we have investigated the task of detecting leakages in WDNs by proposing a distributed approach to prototype-based models. In particular, we extended our initial formulation by designing a solution grounded on federated learning. Moreover, we further extended the previous analysis by including water flow measurements. Since we have utilized a federated modeling paradigm, we have considered that every device is responsible for processing their collected samples and generating local models. Meanwhile, the server only processes local models to building an efficient global model. Specifically, our experiments show that the proposed optimized learning method can obtain higher detection rates at each station than the conventional centralized approach, e.g., improvements of purity rates up to 7.6% in one of the pumping stations, which increased the minimum values from 92.13% obtained through centralized learning, to 99.11%, obtained via federated learning.