Classification of failure modes of pipelines containing longitudinal surface cracks using mechanics-based and machine learning models

Abstract This paper applies the mechanics-based approach and five machine learning algorithms to classify the failure mode (leak or rupture) of steel oil and gas pipelines containing longitudinally oriented surface cracks. The mechanics-based approach compares the nominal hoop stress remote from the surface crack at failure and the remote nominal hoop stress to cause unstable longitudinal propagation of the through-wall crack to predict the failure mode. The employed machine learning algorithms consist of three single learning algorithms, namely naïve Bayes, support vector machine and decision tree; and two ensemble learning algorithms, namely random forest and gradient boosting. The classification accuracy of the mechanics-based approach and machine learning algorithms is evaluated based on 250 full-scale burst tests of pipe specimens collected from the open literature. The analysis results reveal that the mechanics-based approach leads to highly biased classifications: many leaks erroneously classified as ruptures. The machine learning algorithms lead to markedly improved accuracy. The random forest and gradient boosting models result in the classification accuracy of over 95% for ruptures and leaks, with the accuracy of the decision tree and support vector machine models somewhat lower. This study demonstrates the value of employing machine learning models to improve the integrity management practice of oil and gas pipelines.