PHYSICS-INFORMED HYBRID LEARNING FOR PREDICTIVE LEAK DIAGNOSTICS IN HIGH-RISE WATER SYSTEMS

Código

IAHR0003

Autores

CHENG Shu, Oussama CHOURA, Moez LOUATI

Tema

Topic

Resumo

High-rise building water distribution systems (WDS) present significant leak diagnostic challenges due to complex transient flow behaviours (e.g., unpredictable flow and demand (randomness and density - number of branches/consumption nodes), transient waves interactions) and scarce real-world leak data. To address limitations of conventional methods and enable continuous monitoring without supply interruption, this study proposes a physics-informed hybrid learning framework for leak detection, localization, and quantification. Our core innovation integrates convolutional neural networks (CNNs) with physics constraints derived from transient flow governing equations, fusing hydraulic mechanisms with multi-sensor data. This physics constrained deep learning architecture leverages numerical simulations for data augmentation and is being validated using experimental data from a living lab at HKUST. The framework aims to advance transient flow based diagnostics by significantly reducing data dependency and improving accuracy for small leaks in multi floor systems, supporting proactive maintenance, water conservation, and sustainable urban water management through quantified environmental benefits.