Innovative Methodologies for Leak Detection Using Hydrophone Data and Machine Learning Techniques

Abstract

The aim of this research is to develop innovative methodologies for leak detection using hydrophone data and machine learning techniques. The database used has been shared publicly and contains several recordings from various sensors, including hydrophones. The data was collected in a laboratory environment simulating real potable water distribution network topologies. The study focused on recordings without water demand, as they reflect the ideal conditions for leak detection. An outlier analysis revealed that their number was significantly higher in pipes without leaks compared to those with leaks. Two models, Random Forest and SVM, were employed to classify the data based on a threshold defined from this analysis. Using the number of outliers as a discriminant feature, thanks to the acoustic pressure measured by hydrophones, allowed for accurate leak detection. The models showed promising performance, with an average accuracy of 0.95 for the Random Forest model and 0.9 for the SVM model. The methodologies presented in this study proved effective for leak detection using hydrophone data. The combination of outlier analysis and machine learning techniques offers considerable potential to improve leak detection in water distribution networks.