Water main decision support system seizes advantages
Development of the evaluation methodology
The use of statistical analysis to predict failure has a lot of benefit in the ranking or prioritizing
process, especially since it is an objective evaluation and there is little interaction required by the
end user. However, the process of ranking and prioritizing projects may involve much more
complex issues such as politics and other areas that require sound engineering judgment. To
consider these issues, a more subjective ranking system may be required. A more subjective
ranking process is also required to consider events that may not have adequate historical data to
support a statistical analysis. The evaluation methodology adopted in the PIPES application
combines the benefits of the two approaches—an objective evaluation and a subjective
evaluation.
Obiective Evaluation
The purpose of the statistical analysis is to determine if any combination of available data (e.g.,
pipe age, diameter, leak history, installation date, soil type, zoning, etc.) could be used to predict
the time and probability of leaks in water pipes. The final output of the model is a listing of main
segments prioritized on the probability of failure for a given time period.
Data Collection and Formatting
Data were collected from Seattle Public Utilities’ GIS and Seattle City Light (SCL) GIS. The
data were formatted for use in the statistical software package S-PLUS (Version 3.1 for
Windows). Quality control techniques were used to eliminate bad data, such as those pipes with
leak dates prior to the installation date, and those pipes which could not be classified into
categories of interest for this study
Model Selection
Standard regression analysis, whether linear or non-linear, is inappropriate for projects that
contain censored data. Censored data, in the case of water pipes, are data about pipes which
have not yet leaked. That is, they have not had a failure within the time period beginning with
their installation and ending with the time of this study. Information on censored pipes is as
important as information on pipes which have leaked within the study period. In a standard
regression analysis, the time until the first leak for these censored pipes would either have to be
ignored (i.e., the data discarded) or the failure time estimated as some time in the future. Both
methods would result in biased model parameter estimates and poor predictions. For example, if
censored data are ignored, the time until the first leak will be grossly underestimated. Estimating
the failure time for censored pipes (perhaps as the censored time plus a constant) could lead to
bias in either direction. Failure time analysis is a statistical technique for estimating failure times
when some data are censored (Weston 1996).
Failure time analysis was used to account for the unobserved or censored leak times, since most
of the pipes in the study have never leaked. Standard regression techniques would model the
time until the first leak as some function of the independent variables. Failure time analysis
models the probability that the pipe will fail before a certain time as some fi.mction of the
independent variables. These probabilities for the first and subsequent leaks in water pipes were
modeled using various techniques. The Weibull regression model was determined to be the most
appropriate model, based on the available data (Weston 1996).
The survival function for the Weibull distribution:
The variables length, epoch, material and press remained significant as predictors of second
leak time. No independent variables were found to significantly affect the time of third or
subsequent leaks (Weston 1996).
