Tree-based machine learning models for predicting the maximum depth of corrosion defects based on historical in-line inspection data

• Investigates the application of tree-based supervised machine learning techniques to predict the maximum depth of corrosion defects based on historical ILI dataset. • Historical ILI data from a 250 km pipeline in the oil and gas industry is utilized, comprising four ILI runs conducted two years apart, totaling over 3 million defects. • Analyzes four tree-based machine learning techniques: XGBoost, Random Forests, DART, and LightGBM with linear trees, employing six predictors representing the inspection year, defect location, circumferential orientation, and defect dimensions. The models will be trained using the Tree-structured Parzen Estimator (TPE) search algorithm and hyperband pruning with 500 trials for hyperparameter tuning. The performance of these models will be compared to a naïve forecasting benchmark to assess their capability in modeling ILI data. • Establishes an alternative pipeline corrosion prediction framework that prioritizes the use of ILI datasets to develop service-agnostic models applicable to all pipelines in the oil and gas industry, as well as other industries utilizing similar techniques. Oil and gas pipelines are the primary means of fluid transportation in the industry due to their efficiency, reliability, and cost-effectiveness. However, pipeline corrosion poses significant risks, leading to loss of containment, operational interruptions, and potential loss of life if undetected or unmitigated. The introduction of novel and corrosive fluid, such as hydrogen and carbon dioxide (CO2), is expected to exacerbate corrosion-related issues. Consequently, pipeline inspection techniques, particularly In-Line Inspection (ILI), are increasingly vital for corrosion monitoring. This paper establishes a framework for utilizing ILI data to develop machine learning models for corrosion prediction. Four tree-based machine learning techniques—eXtreme Gradient Boosting (XGBoost), Dropouts meet multiple Additive Regression Trees (DART), Light Gradient-Boosting Machine (LightGBM) with linear trees, and random forests—were employed to predict the maximum depth of corrosion defects based exclusively on historical ILI data. All models significantly outperformed the Naïve forecasting benchmark, with the best model achieving a root mean square error (RMSE) of 0.368 mm on the test set, surpassing the benchmark by 41.5%. Furthermore, the accuracy of these models exceeded that of most service-based corrosion prediction models in the literature. This was achieved using an ILI dataset that exhibited extreme variations in maximum depth distribution and changes in reporting criteria. These findings indicate that the proposed framework is a superior alternative to service-based models, leveraging the vast amount of ILI data available to pipeline operators. Additionally, the service-agnostic framework supports the integration of process and service parameters, further enhancing the efficacy of machine learning models for corrosion prediction.