Convolutional neural network for leak location in buried pipes of underground water supply

Water leakage in underground distribution networks is one of the greatest challenges faced by supply companies around the world. Moreover, current leakage detection and location methods are labor intensive or require a very experienced or highly qualified operator. Considering this, the goal of this manuscript is to apply a Machine Learning technique, more specifically a Convolutional Neural Network (CNN) model, to simplify the process of locating water leaks in underground pipelines, calculating the distance between the sensors and the epicenter of the leakage, from measurements on the ground surface. Machine Learning techniques have a great potential to identify the signature of a leak that might be hidden in the high background noise. In this work, accelerations were measured on the ground surface of an experimental platform, varying the vibration intensity of the underground source and the relative positioning of the sensors. The input matrices of the proposed CNN were formed by the Power Spectral Density of the collected signals and were used by three sensors concurrently in the measurements. After an extensive hyper-parameter search, four models that provided the best results were selected. The best model achieved a mean absolute error of 1.01 cm in the predicted leak distance.