Stochastic Resonance Enhancement for Leak Detection in Pipelines Using Fluid Transients and Convolutional Neural Networks

AbstractWater losses through leakage represent a significant problem for asset management in water distribution systems. The interpretation of fluid transient pressure waves after the generation of a transient event has been previously used as a technique to locate and characterize leaks, but existing approaches are often both model-driven and limited to the existing knowledge of the system. The potential of using artificial neural networks (ANN) and fluid transient waves to detect, locate, and characterize anomalies in water pipelines has recently been proposed. However, its application in more realistic conditions (e.g., in the presence of background pressure fluctuations) has proven challenging. To address this, one alternative to enhance the response of any nonlinear system includes the introduction of artificial noise, a phenomenon known as stochastic resonance. In this paper, the enhanced detection of leaks in pressurized pipelines via the deployment of stochastic resonance is demonstrated. This paper harnesses this approach by presenting a methodology for the active inspection of pipelines using convolutional neural networks (CNNs). This methodology finds the optimal artificial noise intensity to be introduced into the training dataset for a set of CNNs. The methodology has been applied to a real pipeline in a laboratory at the University of Adelaide in which 14 transient experimental tests were conducted. The results indicated that the addition of noise to the transient pressure head training samples significantly enhances the CNN predictions for the leak location highlighting the existence of an optimum noise intensity to obtain both accurate and reliable results. When trained with the optimum noise intensity, the CNNs were able to locate leaks with an average error of 0.59% in terms of the actual location (in a 37.24-m long pipeline), demonstrating the promising potential of developing techniques based on CNNs to detect leaks and anomalies in water pipelines.