Deteção de anomalias em tempo real em sistemas de abastecimento de água através de gémeos digitais baseados em métodos de aprendizagem automática

AQUALEARN aims to leverage fluid dynamics principles to detect, locate and quantify different types of anomalies in WSS, both continuously and on periodic surveys. Anomalies, like leaks, blockages, air pockets, water quality issues and illegal connections, inevitably exist in WSS, due to the systems gradual deterioration and networks extensive/continuous use. These anomalies lead to frequent bursts, service disruptions, increased water and energy losses and higher O&M costs. The digital transition with the extensive WSS sensorization, monitoring all sorts of parameters (e.g., water levels, flow rates, discharges, water quality), the development of WSS digital twins for real time simulation, and major advancements in artificial intelligence and machine learning, provide an excellent environment with a robust pool of technologies that can be innovatively combined for addressing the anomaly detection in WSS. The faster and more effective detection and repair of anomalies contribute to reduce water and energy losses, service disruptions and O&M costs, while promoting an increase in water availability and energy decarbonisation. The proposed project, AQUALEARN, aims at developing, testing in the laboratory and demonstrating in field conditions advanced and novel machine learning-based methods for real time detecting, locating and quantifying a wide range of anomalies in WSS (e.g., bursts, blockages, air pockets, illegal connections, water quality degradation). It will extend beyond the state-of-the-art knowledge by introducing new techniques based on high-frequency pressure measurements to classify, locate and size anomalies, as well as a new generation of network models to enhance hydraulic and water quality simulations. Furthermore, it will implement machine learning algorithms based on big data to provide more accurate interpretations of fluid dynamic principles. Decision-making will be improved to facilitate prompt intervention to uphold water supply efficiency, infrastructure reliability, and service safety. AQUALEARN has established a set of specific objectives: (i) the construction of a comprehensive database with the most frequent anomalies, their signatures and examples of successful detection methods; (ii) the development of an experimental programme to collect the signatures of anomalies during transient events (high-frequency data); (iii) the implementation of a hydraulic transient solver for generating a large set of pressure signals for different anomalies; (iv) the development of numerical models of pilot networks for exploring the use of multiple monitored parameters to detect anomalies; the implementation and testing of novel machine-learning algorithms for (v) real time anomaly detection and approximate location, based on the continuous monitoring of multiple parameters and (vi) real time anomaly classification, location and quantification based on high-frequency pressure signals; (vii) the development of machine learning-based digital twins of pilot networks and their full-scale testing to assess data and modelling uncertainties, types of anomalies effectively detected and practical difficulties; (viii) the proposal of a well-tested and consolidated methodology for real time anomaly detection in WSS; (ix) the elaboration of guidelines for the application of the proposed methodology in real-life WSS to more rapidly and effectively detect anomalies.