A water distribution system is a key part of the critical infrastructure. Its elements are monitored to obtain the information about the condition of the system. However, recorded data sets do not usually contain the information on every single event and there is no detailed information regarding the data structure. More often, the data are
aggregated, usually in a monthly manner. Details in the data concerning failure dates, the length of down time, etc.would significantly help to make more sophisticated assessment.
Having relatively poor data in hand, we employed nonparametric regression estimators—regressogram and kernel smoothers—to assess a reliability measure: the rate of occurrence of failures. We compared the nonparametric approach with widely used parametric methods, namely with the power law model and the logarithmic linear model.
All of the mentioned methods were used for the sake of modeling the rate of occurrence of failures of the water distribution system. Also the usage of the nonparametric estimates was supported by a simulation study focused on the interpolating and extrapolating predictive power of the used models. The outcomes of the study show that
for a constant and monotone failure rates the models are comparable, but for the bathtub shaped failure rate the nonparametric model overcomes the parametric ones.