Modelling Cable and Pipe Failures from Excavation Works

Abstract

Cables and pipes are critical infrastructure systems (CISs) which are mostly located in the very crowded subsurface. Especially in urban areas, a typical road includes five to ten infrastructure systems which are all owned and managed by different entities. The CISs are spatially interdependent as these are highly interconnected due to the close spatial proximity. Despite the critical function of cables and pipes, over 30,000 cable and pipe failures from excavation works are reported in the Netherlands yearly. Multiple studies have been conducted to reduce the risk of excavation damage. These studies have mainly focused on the impact side. Remarkable as from an extensive cooperation between the network operators and other stakeholders, a guideline (CROW500) was formed that seeks to prevent cable and pipe damage from excavation works.
The objective of this thesis is to develop a model to accurately predict cable and pipe failures from excavation works, considering spatial interdependencies. The associated research question is: “What method can predict the influence of spatial interdependencies on the probability of failure from excavation works on the cables and pipes of subsurface utility operators?”
Predictive analyses are frequently used for enhanced decision making by subsurface utility operators, whereby cable and pipe failures, having large impact in this sector, are relatively rare. This thesis explores the possibilities of modelling rare event data within the subsurface utility industry, through which specific situations, such as failures from excavation works, can be considered. A case study within Evides Water Company was conducted whereby 107,000 non-failures and 180 failures from excavation works were collected. In terms of statistical techniques, alternative models to logistic regression and Bayesian logistic are considered. Two approaches, involving weighting and synthetic minority oversampling have been examined to compensate the imbalanced classifiers in the data set. Balancing is done by over-, under sampling, as well as by weighting, which aim to increase the accuracy of the models. At algorithm level, under sampling and weighting combined were tested and found to improve the balanced accuracy to 0.66 with 0.38 of the failures predicted. At data level, over and under sampling by SMOTE resulted in 0.58 of failures predicted and a balanced accuracy of 0.61. These results proved that logistic regression for network operators can predict failures in specific situations with reasonable accuracy.