Development of machine learning-based burst capacity models for pipelines containing dent-gouges with synthetic full-scale burst test data generated using tabular generative adversarial network
Ze He, Wenxing Zhou
Abstract
This study develops accurate burst capacity models for steel oil and gas pipelines containing dent-gouge damages through the innovative use of a deep learning algorithm, the tabular generative adversarial network (TGAN), and three machine learning (ML) algorithms, the random forest, extra tree and Gaussian process regression. TGAN is employed to expand a limited number (88) of full-scale burst tests of dent-gouged pipe specimens available in the literature by generating synthetic test data. A dataset consisting of 62 real and 438 synthetic test data is then employed to train the three ML models to predict the dent-gouge burst capacity. Based on a validation dataset of 26 real test data, the accuracy of the ML models is shown to be markedly higher than that of a well-known semi-empirical engineering model: coefficients of variation of the test-to-predicted ratios for the ML models are below 15% compared with about 45% for the engineering model. Uncertainty quantification of predictions by the ML models is also carried out. The present study demonstrates the promising potential and effectiveness of combining deep learning algorithms and ML models to improve the integrity assessment practice for oil and gas pipelines.