INCREASING THE ROBUSTNESS OF CLASSIFICATION ALGORITHMS TO QUANTIFY LEAKS THROUGH OPTIMIZATION

Abstract

Leaks in water pipeline networks have cost billions of dollars each year. Robust leak quantification (to detect and to localize) methods are needed to minimize the lost. We quantify leaks by classifying their locations using machine learning algorithms, namely Support Vector Machine and C4.5. The algorithms are chosen due to their high performance in classification. We simulate leaks at different positions at different sizes and use the data to train the algorithms. We tune the algorithm by optimizing the algorithms' parameters in the training process. Then, we tested the algorithms' models against real observation data. We also experimented with noisy data, due to sensor inaccuracies, that often happen in real situations. Lastly, we compared the two algorithms to investigate how accurate and robust they localize leaks with noisy data. We found that C4.5 is more robust against noisy data than SVM.