This work proposes a reliable leakage detection methodology for water distribution networks based on machine learning techniques. The design is developed through real data acquisition from a municipal area of a water distribution network. We propose to combine both unsupervised learning (K-means and cluster validation techniques) and supervised learning (LVQ-type algorithms) for the efficient design of prototype-based classifiers. We investigated several metrics aiming to define the optimal number of clusters, in which we succeeded in reporting attractive classification accuracies (approximately 90%) on scenarios of severely limited number of prototypes.