Pipeline leak detection using hydraulic transients and domain-guided machine learning

In this paper, we present a novel domain-guided framework for training machine learning (ML) models, particularly neural networks (NN), to detect multiple leaks in pipelines using hydraulic transient datasets. Central to our approach is the utilization of a ‘leak function’ as the target variable in the ML model, where leak properties exhibit a pulse-like signature. This framework allows ML models to be trained without the typical constraints on the number of leaks and avoids the necessity of retraining separate models for different scenarios. To enhance model robustness against operational uncertainties, we employ iterative refinement techniques. Starting with a baseline NN model, each subsequent refinement incorporates factors such as valve closure time, pre-transient flow properties, and the capacity to handle an increased number of leaks. We validate the accuracy of our framework through numerical simulations and experimental data from a laboratory-scale Viscoelastic pipe system. Furthermore, we demonstrate the framework's applicability to pipeline networks in two configurations: a series connection of two pipes and a tree-type network comprising three pipes. Our results indicate satisfactory detection accuracy and system adaptability, showcasing the potential of our framework for practical ML applications in leak detection.