LEAK DETECTION IN WATER SUPPLY NETWORKS USING MACHINE LEARNING

This study analyzed the performance of the Decision Tree Classifier algorithm in detecting leaks in the Net-1 network. Experiments with individual variables and their combinations were explored using the LeakDB database. Comparative analyses included learning curves and stratified cross-validation to assess performance on algorithm training and test phases for anomaly detection. No evidence of overfitting was found. Performance was measured with precision, accuracy, recall, F1 score, and confusion matrix. The most promising experiment combined pressure, flow, and demand, reached a precision of 96.64%, an accuracy of 97.5%, a recall of 100%, and an F1 of 98.29%. The demand variable stood out as crucial to the model's performance.