Doctor of Philosophy
Civil and Environmental Engineering
Metal-loss corrosion and third-party damage (TPD) are the leading threats to the integrity of buried oil and natural gas pipelines. This thesis employs Bayesian networks (BNs) and non-parametric Bayesian networks (NPBNs) to deal with four issues with regard to the reliability-based management program of corrosion and TPD. The first study integrates the quantification of measurement errors of the ILI tools, corrosion growth modeling and reliability analysis in a single dynamic Bayesian network (DBN) model, and employs the parameter learning technique to learn the parameters of the DBN model from the ILI-reported and filed-measured corrosion depths. The second study develops the BN model to estimate the probability of a given pipeline being hit by third-party excavations by taking into account common preventative and protective measures. The parameter learning technique is employed to learn the parameters of the BN model from datasets that consist of individual cases of third-party activities. The ILIs are infeasible for a portion of buried pipelines due to various reasons, which are known as unpiggable pipelines. To assist with the corrosion assessment for the unpiggable pipelines, the third study develops a non-parametric Bayesian network (NPBN) model to predict the corrosion depth on buried pipelines using the pipeline age and local soil properties as the predictors. The last study develops an optimal sample size determination method for collecting samples to reduce the epistemic uncertainties in the probabilistic distributions of basic random variables in the reliability analysis of corroded pipelines.
Summary for Lay Audience
The buried pipelines are the most widely used mode to transport oil and natural gas. The corrosion and damage from excavation activities can lead to pipeline incidents. To manage the pipeline safety, pipeline companies need to estimate the occurrence probabilities of such incidents. This thesis uses graphical models known as Bayesian networks to enhance the corrosion and excavation damage management program. A Bayesian network (BN) consists of circles to represent events and arrows to represent the relationship between the events. Once a part of the model is observed, the probabilities of the rest of the events can be calculated. Pipeline companies routinely run inspection tools through the pipelines to detect and size corrosion defects. The thesis develops a BN model to forecast the growth of the corrosion depth and probability of failure at the specific corrosion defect using the corrosion depths reported by the inspection tools. However, the inspection tools are infeasible for a portion of pipelines due to the reasons such as small diameters and tight bends. To assist with the corrosion assessment of such pipelines, the thesis develops a BN model to predict the corrosion depth using the pipeline age and soil parameters. To prevent the pipeline from excavation damage, the pipeline industry and regulatory agencies employ a series of measures such as patrols along the pipeline, warning signs on the pipelines and burial depth. The failures of all the preventative and protective measures can lead to the pipeline being hit by the excavation machine. The present thesis develops a BN model to estimate the probability of a given pipeline being hit by an excavation event. The probabilities of the preventative and protective measures are automatedly learned from the historical data collected by the pipeline industry. To reduce the uncertainties in the corrosion management program, the pipeline industry often collects samples by performing experiments or field measurements, which are generally expensive. The fourth study in the thesis develops a method to determine the optimal sample size from the economic standpoint and apply it to two sample size determination problems in the context of corrosion management of pipelines.
Xiang, Wei, "Application of Bayesian Networks to Integrity Management of Energy Pipelines" (2019). Electronic Thesis and Dissertation Repository. 6688.