Probabilistic failure prediction for deteriorating pipelines: Nonpararnetric approach

Component failures in water distribution systems are usually predicted by parametric models where the model parameters are determined by projecting the past failure rates of the component to the future. This paper shows that in such techniques, failures are implicitly assumed to be stationary random processes. However, due to the nonstationary nature of some influencing factors, this assumption may lead to inaccurate predictions. A new nonparametric technique is developed for failure prediction of classes of pipes considering this nonstationary process. The presented technique uses limited data that are typical to the databases of water distribution systems. In this method, maximum likelihood estimates of the probability of future failures are calculated and used, both to predict the number of failures occurring within a specified period of time in future, and to provide some lower and upper bounds (confidence intervals) for the estimations. This technique is applied to predict the failures of water pipes in western suburbs of Melbourne. Results of the predictions are compared with the empirical results from a failure record. Deviation of these predictions from empirical measures in terms of both rejection rates and mean-square errors of predictions are acceptable.