Abstract

Machine learning (ML) approaches allow engineers and data scientists to develop models that learn from collected data and generalize inferences for new situations where data is not already available (eg, pipeline segments for which inline inspection (ILI) has not been conducted). Best-practice guidelines were recently developed to promote and facilitate adoption of ML for pipeline integrity applications. The guideline procedures were applied to an accompanying demonstration model for stress corrosion cracking (SCC) susceptibility prediction. The developed guidelines provide an overview of ML concepts and describe a step-by-step process for applying ML best practices in the context of pipeline integrity. They also include a review of existing published ML applications in pipeline integrity management and provide a categorized compilation of helpful references.

In the accompanying demonstration, a set of ML models were developed, following the step-by-step process described in the guidelines, for predicting SCC susceptibility to support targeted ILI inspections, direct assessment, and risk modelling. The models were trained and validated using anonymized pipeline attribute and inspection data for approximately 30,000 km of transmission pipelines from six operators in Canada and the United States. This data was also aligned with publicly-available soil data from the Soil Survey Geographic Database (SSURGO) and Canadian Soil Information Service (CanSIS) database.

The developed models include a set of classifiers using the Extreme Gradient Boosted Decision Trees algorithm (XGBoost) along with a simplified classifier based on industry-standard SCC screening criteria. Special consideration was given to appropriately splitting data for cross-validation and testing to avoid data leakage due to spatial autocorrelation of datapoints (reducing potential for artificially inflated model performance). Visualization of feature contributions to model predictions using SHapley Additive exPlanations (SHAP) values enabled additional evaluation of potential overfitting and improved model interpretability.

A performance comparison between the simple classifiers and the more complex XGBoost models was made on the basis of standard machine learning metrics. The results show that, while the simple classifier exhibits reasonable performance, the XGBoost models allow more efficient screening of SCC-susceptible pipe segments (ie, resulting in a lower false positive rate at equivalent recall).

This paper describes the development the SCC susceptibility model (including problem definition, data processing, model engineering, and performance evaluation) in the context of demonstrating the process described in the developed ML best-practice guidelines for pipeline integrity.

Authors:

McAllister, Lowell

Publisher:

ASME The American Society of Mechanical Engineers
2026 International Pipeline Conference

Year Published:

2026

Purchase Link:

Coming Soon.