The problem of fault diagnosis in potable water supply networks is addressed in this paper. Two different fault diagnosis approaches are proposed to deal with this problem. The first one is based on a model-based approach exploiting a priori information regarding physical/temporal relations existing among the measured variables in the monitored system, providing fault detection and isolation capabilities by means of the residuals generated using these measured variables and their estimations. This a priori information is provided by the topology and the physical relations between the elements constituting the system. Alternatively, the second approach relies on a data-driven solution meant to exploit the spatial and temporal relationships present in the acquired data streams in order to detect and isolate faults. Relationships between data streams are modelled using sequences of linear dynamic time-invariant models, whose estimated coefficients are used to feed a hidden Markov model. Afterwards, a cognitive method based on a functional graph representation of the system isolates the fault when existing. Finally, a performance comparison between these two approaches is carried out using the Barcelona water supply network, showing successful and complementary results which suggest the integrated usage in order to improve the results achieved by each one separately.