Leak Detection and Localization in Water Networks Using Convolutional Neural Networks with a Modified Loss Function

Leaks in water distribution networks can cause considerable water wastage, infrastructure damage, and contamination. Various solutions have been developed for leak detection. Previous researchers have applied deep learning, utilizing traditional loss functions, such as F1 score or accuracy. However, these approaches may oversimplify the real challenge of quickly repairing leaks. This study aims to develop a Convolutional Neural Network (CNN) for leak detection, trained with the specific objective of optimizing leak detection. A synthetic data set is developed by simulating leaks using EPANET in Python. The leakage detection model was trained with a modified loss function based on the distance to the actual leak. Results show that the use of the modified loss function led to better leak detection accuracy than with the traditional loss function. These findings provide insights into strategies for using novel deep-learning models to efficiently detect leaks.