Generation of synthetic full-scale burst test data for corroded pipelines using the tabular generative adversarial network

Abstract

This paper employs a deep learning tool, the tabular generative adversarial network (TGAN), to generate synthetic full-scale burst test data for corroded pipe specimens by capturing the joint probability distribution of five dimensionless random variables characterizing a database of 258 real full-scale burst tests collected from the literature. A simple criterion is proposed to identify outliers contained in the synthetic dataset. Two machine learning models, the random forest and extra tree, are trained using the real and synthetic test data to predict the burst capacity of corroded pipelines for the purpose of tuning the hyper-parameters of TGAN and also validating the credibility of the synthetic data. The generated synthetic data are shown to accurately capture the joint probability distribution of the real test data. This study provides a viable option to effectively generate a large number of high-quality synthetic full-scale test data to facilitate the development of engineering critical assessment models employed in the pipeline integrity management practice.