Analysis of machine learning models and data sources to forecast burst pressure of petroleum corroded pipelines: A comprehensive review

A comprehensive evaluation of the integrity of oil and gas pipelines subjected to corrosion defect is required for forecasting health & safety actions. If corrosion is ignored, it may have significant repercussions on a person's health, finances, and the environment. The preponderance of failure pressure prediction research uses numerical simulations and industry-specific codes. However, the complexity and magnitude of deteriorated pipe systems make machine learning based technologies such as artificial neural networks, support vector machines, deep neural networks, and hybrid supervised learning models more suited. Current ML research techniques that predict burst pressure lack a comprehensive review. This research aims to evaluate the present ML techniques (methodology, variables, datasets, and bibliometric analysis; Most active researchers, journals, regions around the world and institution). Based on the results the most widely used machine learning model is ANN followed by SVM but still they have some major limitations such as overfitting and generalization, but other machine learning models such as random forest and ensemble models along with theory guided models have been utilized even though still a very little research has been carried out. The most commonly datasets used to build these models are either experimental or numerical simulations conspiring of inputs and outputs as geometry and pipe material-based parameters such as pipe diameter, material grade, defect depth-breadth-length, and wall thickness. Most of these datasets have been built by known institutions such as PETROBRAS, KOGAS, BRITISH PETROLEUM and Waterloo university. In addition, this analysis revealed research limitations and inadequacies, including data availability, accuracy, and validation. Finally, some future recommendations and opinions are presented such as collaboration between institutions to share the dataset, providing more practical models such as physics informed machine learning and digital twins in the field.