| City: |
Scarborough, Ontario, Canada
|
| Organization: |
University of Guelph
|
| Project Start Date: |
Unspecified
|
| Project End Date: |
2012
|
| Reference: |
Harvey, R., McBean, E. A., and Gharabaghi, B., (2014), “Predicting the timing of water main failure using artificial neural networks”, Journal of Water Resources Planning and Management, 140(4), 425-434.
|
| Problem: |
Water main failures in an ageing infrastructure can cause interruptions in water access, contaminant intrusion, damage to nearby infrastructures, disruption in the transportation system, and financial burden. Therefore, Greater Toronto Area, having a relatively old water system and about 1300 breaks a year, has planned to invest in water mains rehabilitation at a proper time. Scarborough is one of the regions with the highest rate of breaks, especially during cold months. Predicting the time of failure by using data mining would provide insight into analyzing, inspecting, and planning maintenance and rehabilitation practices. |
| Technical Solution: |
- Random partitioning the dataset into training, testing, and holdout subsets for each pipe material (asbestos cement, cast iron, and ductile iron)
- Training three Neural Networks for different pipe materials using back-propagation with momentum algorithm
- Determination of the optimal network architectures using a double-loop technique
- Using the holdout subset and the early-stopping cross-validation technique to ensure reaching the global optimal solution
- Standardizing the value of attributes and time to failure by subtracting the mean and dividing by the standard deviation
- Using importance analysis to evaluate the role of each attribute on the trained model.
|
| Datasets Used: |
Scarborough’s historical dataset of water main failure, the Scarborough district, 1962-2005
|
| Outcome: |
- Failure history has a strong impact on failure behaviour and failure prediction over time.
- Using cement mortar lining and cathodic protection could increase the lifetime of an aging water network.
- The correlation coefficient as a performance metric shows the acceptable ability of the model to predict time to failure of the water mains (correlation coefficients of 0.70 for asbestos cement, 0.7 for cast iron, and 0.81 for ductile iron).
- The prediction results are the expected values and could be considered as one part of the available information on rehabilitation decision-making practices.
|
| Issues that arose: |
- No model is developed for PVC pipes due to the small sample size of failures
- The model is unable to predict the location of the breaks along the pipes.
- As the corrosion seems to be an important cause for water mains failure in Scarborough, data on the resistivity and corrosivity of the soil are important missing attributes in this study.
- Since about 60% of water main failures in Scarborough occur during cold months, temperature data, forest layer data, and burial depth data would improve the model.
- Data on the location of the pipes and the adjacent municipal infrastructures, pipe wall thickness, and water pressure should be considered in future studies.
|
| Status: |
Terminated.
|
| Entered by: |
October 30, 2020: Negin Ahadzadeh, negin.ahadzadeh@mail.utoronto.ca.
|