Electronic Thesis and Dissertation Repository

Thesis Format

Integrated Article

Degree

Doctor of Philosophy

Program

Civil and Environmental Engineering

Supervisor

Zhou, Wenxing

Abstract

Dent-gouges and corrosions are two of the well-known failure mechanisms that threaten the structural integrity management of oil and gas pipelines. Dent-gouges or corrosions markedly reduce the burst capacity of pipelines as a result of localized wall thickness reduction. Fitness-for-service (FFS) assessment is commonly employed to maintain the integrity of in-service pipelines containing flaws and the burst capacity evaluation is central to the FFS assessment. As the predictive accuracy of existing FFS models is generally very poor, the use of machine learning (ML) tools provides a viable option to develop burst capacity models with high accuracy. The main objective of the present thesis is to facilitate the FFS assessment of dent-gouges and corrosions based on ML tools.

The first study proposes an improved burst capacity model for pipelines containing dent-gouges based on European Pipeline Research Group (EPRG) burst capacity model using full-scale burst tests by adding a correction term. The Gaussian process regression (GPR) is employed to quantify the correction term, which is a function of six non-dimensional random variables incorporating the effect of pipe and geometric properties, sizes of dent-gouges, and internal pressure loading condition. The accuracy of the improved EPRG model, i.e. EPRG-C model, is validated based on the comparison between the test and predicted burst capacities corresponding to the test data, and shown to be markedly greater than that of the EPRG model, suggesting the high effectiveness of the correction term.

The second study presents a limit state-based assessment (LSBA) framework for pipelines containing dent-gouges to achieve reliability consistent outcomes. The LSBA is formulated based on the EPRG-C model proposed in the first study by assigning appropriate partial safety factors to key variables as well as the internal pressure. The calibration of partial safety factors is carried out by making the outcomes of LSBA are consistent with those of the reliability-based assessment given different pre-selected allowable failure probabilities. The failure probabilities corresponding to extensive assessment cases covering wide ranges of pipe geometric and material properties, sizes of dent-gouges and the model error are evaluated using the first-order reliability method. The validity of the calibrated partial safety factors is demonstrated using independent assessment cases and two illustrative examples. The advantages of LSBA over the deterministic assessment procedure in terms of achieving reliability-consistent assessment outcomes is further demonstrated.

The third study employs a deep learning algorithm tabular generative adversarial network (TGAN) to generate synthetic burst tests by capturing the joint probability distribution based on real full-scale burst test data of corroded pipelines. Two other ML tools, random forest (RF) and extra tree (ET), are used to tune the hyper-parameters and validate the credibility of TGAN-generated data. A simple criterion is proposed to eliminate the outliers contained in the synthetic data. The results indicate that the synthetic burst test data match well with the real data, suggesting that TGAN can accurately capture the joint probability distribution of real test data and generate credible synthetic data.

The fourth study develops new ML-based burst capacity models for dent-gouges with combined real and synthetic full-scale burst tests. The synthetic burst test data are generated using TGAN framework, which is proposed in the third study. The results of which are used as the basis combined with the real burst tests to develop ML burst capacity models based on three ML tools, i.e. RF, ET and GPR. The proposed models are shown to be more accurate than the models developed using real test data only. The analysis result further indicates that trained models are markedly more accurate than the semi-empirical EPRG model widely employed in the pipeline industry.

Summary for Lay Audience

Pipelines buried underground are commonly considered as the most efficient means to transport large quantities of oil and gas products such as the carbon dioxide and crude oil. However, the structural integrity of pipelines is often threatened by various defects. The occurrence of such defects leads to the reduction of the pressure containment capacity (i.e. burst capacity). Dent-gouges and corrosions are two of the most commonly observed failure mechanisms in practice, which pose direct threats to the pipeline integrity. The fitness-for-service (FFS) assessment is generally employed in pipeline industry to ensure the integrity of pipelines. This research aims at improving the accuracy of FFS assessment models for pipe specimens containing dent-gouges and corrosions using machine learning (ML) tools.

Empirical and semi-empirical burst capacity models have been proposed in the literature to predict the burst capacity of pipelines containing dent-gouges, e.g. the dent-gouge fracture model adopted by the European Pipeline Research Group (EPRG). However, as the simulation of a dent-gouge is very complex, the EPRG model is associated with considerable errors. This study employs the machine learning (ML) tools to improve the accuracy of the EPRG model based on the full-scale burst test data. The improved EPRG model can then be employed as the basis to develop limit state-based assessment framework to identify critical dent-gouge defects for mitigation to facilitate the performance-based pipeline integrity management. However, because of the high cost to conduct full-scale burst tests, the number of available pipe segments reported in the literature is scarce, which casts doubt on the model credibility from a practical standpoint. To address this limitation, the tabular generative adversarial network (TGAN) is used to generate synthetic full-scale burst tests by capturing the joint probability. By combing real and credible TGAN-generated synthetic data of dent-gouges, large quantities of full-scale burst tests are used to develop and validate ML burst capacity models.

COinS