Failure prediction for water pipes

This thesis focuses on predicting the future condition of pipes in water supply networks based on their previous performance using statistical analysis. The contemporary methods developed to solve this problem are reviewed and a number of novel statistical analyses and new probabilistic techniques that enhance the failure prediction accuracy and uncertainty modelling are introduced. When a complete history of water pipes failures is available, the statistical analysis will efficiently provide an accurate formulation of the relationship between failure frequencies and the factors contributing to the overall structural deterioration of the pipes. This result can then be effectively utilised to predict the future failures of the pipes. In practice, however, a complete dataset collected on a regular basis for each pipe of water networks is very costly and not readily available. In such circumstances, more sophisticated statistically derived models are required. In this thesis, a failure history of water mains provided by City West Water PTY LTD (CWW) is studied and analysed as a typical database that is usually available for water supply networks. This database is also used for comprehensive simulation and evaluation purposes. An intelligent statistical reliability model based on artificial neural networks is proposed for reliability estimation of pipes similar in terms of material, diameter, location, etc. Application of this model to the CWW failure dataset shows that it substantially outperforms existing statistical reliability models based on lognormal and Weibull distributions. In the next step, the ensemble of failures of each group of similar pipes (called a pipe class) are studied as a random process and demonstrated to be non-stationary because of the time-varying environmental factors that affect the pipe failure processes. This thesis concludes with suggesting a new non-parametric probabilistic technique developed to capture the non-stationary process of pipe failures despite the lack of information about time-variant factors which is typical of the data available in water distribution systems. The predictions are updated automatically and therefore take the gradual time-variant factors into account. The output of this novel non-parametric auto-updating technique is a confidence interval that represents a range of possible number of failures occurring in a given period of time in the future with a given confidence. The results of evaluation of this method for prediction of failures in the CWW failure database show that, in 95% of the cases, the actual number of failures is within the confidence interval given by the suggested technique.