A deep learning model to predict the failure response of steel pipes under pitting corrosion

• Published in: Computational mechanics Vol. 71; no. 2; pp. 295 - 310
• Main Authors: Ji, Mingshi, Yang, Ming, Soghrati, Soheil
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01.02.2023; Springer; Springer Nature B.V
• Subjects: Classical and Continuum Physics; Computational Science and Engineering; Corrosion; Deep learning; Engineering; Failure; Finite element method; Gas pipelines; Hoop stress; Inspection; Mesh generation; Natural gas; Optimization; Original Paper; Petroleum pipelines; Pitting (corrosion); Steel; Steel pipes; Tensile stress; Theoretical and Applied Mechanics; Deep learning; Finite element method; Pitting corrosion; Squeeze-and-excitation residual network; Pipe
• ISSN: 0178-7675; 1432-0924
• DOI: 10.1007/s00466-022-02238-y

Pitting corrosion is one of the major causes of failure in high-pressure oil and gas pipelines. Various inspection techniques can be used to characterize the morphology of corrosion pits, which must be linked to the risk of failure to develop proper maintenance strategies. While numerical techniques such as the finite element method can accurately predict this risk, the labor and computational cost associated with these methods render their application unfeasible over hundreds of miles of a pipeline. In this manuscript, we introduce a deep learning approach relying on the squeeze-and-excitation residual network (SE-ResNet) to predict the strength and toughness of statistical volume elements (SVEs) of a corroded pipe. An automated microstructure reconstruction and mesh generation framework is utilized to synthesize the training data for this model by simulating the failure response of 10,000 SVEs subject to a tensile load (hoop stress). A Bayesian optimization approach is utilized to determine the optimal combination of hyperparameters for the SE-ResNet model, followed by a k -fold cross-validation of the model. We show that the trained SE-ResNet can accurately predict the failure response of corroded pipe SVEs with a maximum error of < 1 % . Moreover, a comparison between the proposed model with several other well-known DL architectures shows that it yields superior accuracy and efficiency.