Generative Adversarial Network Informed Burst Failure Risk Analysis of Oil and Gas Pipelines

Machine Learning (ML) applications in pipeline failure risk prediction have recently shown promising results. However, obtaining actual data to train ML models is a significant challenge due to safety concerns. To overcome this limitation, this study employed a Generative Adversarial Network (GAN)-based framework to generate synthetic data based on a subset of experimental test data compiled from the literature. The burst failure risk of corroded oil and gas pipelines was determined using probabilistic approaches where pipelines were classified into two groups: (1) low risk (pf: 0−0.5) and (2) high risk (pf: >0.5). Two Random Forest (RF) models were trained and validated using a subset (70%) of actual experimental data and combined actual and synthetic data. The outcomes of this study revealed that synthetically generated data improve the accuracy of ML models and could be an alternative to actual data for developing failure risk prediction without compromising the safety of real data.