Optimal leakage detection and classification of the water distribution network based on the machine learning approach

Water distribution networks are operated to supply tap water stably, but they suffer from leakage due to various factors, resulting in poor performance. In Korea, IoT (Internet of Things)-based smart sensors are installed to reduce these leaks, and a large amount of data is collected. In this study, to solve the data uncertainty and verification of the model’s field applicability, the eXtreme Gradient Boosting (XGBoost) model, which is based on boosting, was selected. This choice was made through hyperparameter optimization using leak detection data obtained from leak detection sensors, as well as data refined through on-site leak detection. The area under the curve_receiver operating characteristic value for each leak class using the XGB model was 0.9955 for class 0, 0.9956 for class 1, and 0.9985 for class 2. In addition, the partial dependence plot of 120 Hz, which has the highest variable importance, was derived from analyzing the influence of the leak detection and classification model on the amount of leak vibration change at 120 Hz. Therefore, it is expectged the suggested methods can be used to build a GIS-based monitoring system for real-time leak detection by linking the developed model and the location of the leak detection sensor.