Using deep learning to identify the severity of pipeline dents



Published May 31, 2020
Ishita Charkraborty Brent Vyvial


With the advent of machine learning, data-based models can be used to increase efficiency and reduce cost for the characterization of various anomalies in pipelines. In this work, artificial intelligence is used to classify pipeline dents directly from the in-line inspection (ILI) data according to their risk categories. A deep neural network model is built with available ILI data, and the resulting machine learning model requires only the ILI data as an input to classify dents in different risk categories. Using a machine learning based model eliminates the need for conducting detailed engineering analysis to determine the effects of dents on the integrity of the pipeline. Concepts from computer vision are used to build the deep neural network using the available data. The deep neural network model is then trained on a sub set of the available ILI data and the model is tested for accuracy on a previously unseen set of the available data. The developed model predicts risk factors associated with a dent with 94% accuracy for a previously unseen data set.

How to Cite

Charkraborty I, Vyvial B. Using deep learning to identify the severity of pipeline dents. PST [Internet]. 2020May31 [cited 2021Jan.24];4(2(4):90-6. Available from:


Download data is not yet available.
Abstract 151 | PDF file Downloads 98



Deep learning, in-line inspection, pipeline, finite element analysis.

[1] Fields, R. J., Foecke, T. J., and deWit, R., Effects of dents and gouges on integrity of pipelines, NIST document NISTR 5479, June 1993.
[2] Fowler, J. R., Criteria for Dent Acceptability in Offshore Pipeline. (1993, January 1). Offshore Technology Conference. doi:10.4043/7311-MS
[3] Alexander, C., Evaluating Damage to Onshore and Offshore Pipelines Using Data Acquired Using Inline Inspection Efforts, PPIM, 2009.
[4] Deep Learning, Ian Goodfellow, Yoshua Bengio, Aaron Courville, MIT Press, 2016.
[5] Scikit-learn: Machine Learning in Python, Pedregosa et al., JMLR 12, pp. 2825-2830, 2011.
[6] Abaqus 6.14, Dassault Systems, 2014.
Original Work