Using advection-diffusion model and auto-encoders to reduce adversarial sensor data predictions in water leak management

Accurate water leak localisation remains a significant challenge in pipeline maintenance, water distribution networks, and building water systems. Traditional methods using pressure or acoustic monitoring often lack precision, leading to wasted water and infrastructure damage. This paper proposes a novel methodology to enhance leak localisation accuracy by combining machine learning and physical modelling. The strengths of two techniques are leveraged: advection-diffusion models predict the spread of a leak over time based on the movement of water in a pipeline or a building's water system, and autoencoders, a type of neural network, learn encoded representations of the current sensor data under normal operating conditions along with common adversarial patterns. In the proposed work, autoencoders are combined with advection-diffusion models to achieve significantly improved performance in water leak localisation. Experiments underscore that this combined approach can help reduce false alarms in locating water leaks, preventing further damage and reducing water waste. The proposed model shows improved performance in terms of precision, recall, and accuracy over existing systems. The improved performance is due to the effective hybridisation of physical and neural network models for the reduction of false alarm rates in water leak localisation.