Demand Model in Water Distribution Networks for Fault Detection * *Paper supported by DGAPA-UNAM IT100716, II-UNAM and Universidad Tecnologica de La Habana Jose Antonio Echeverría (CUJAE)

A water distribution network (WDN) is a dynamic system in which the demand is a nonstationary process. On the other hand, most of the successful data-driven fault detection and isolation (FDI) methods have been developed by assuming static models and stationary processes. A demand model for the WDN is formed by a periodic signal plus a stochastic variable. This model allows the transformation of data such that a stationary and extended space of data can be obtained as it is demonstrated here. This proposition is the principal contribution of this work. To illustrate the advantages of the proposal, three well-known data-driven FDI algorithms are applied for the leak detection of the Hanoi distribution network: principal component analysis (PCA), independent component analysis (ICA), and support vector data description (SVDD). The leaks were emulated with different outflow magnitudes in all nodes. As a performance index, the fault detection rate is used, and the results indicate an improvement in the general index of approximately 75% when data are periodically transformed according to the proposal.