Interpretable machine learning for maximum corrosion depth and influence factor analysis

Abstract

We have employed interpretable methods to uncover the black-box model of the machine learning (ML) for predicting the maximum pitting depth (dmax) of oil and gas pipelines. Ensemble learning (EL) is found to have higher accuracy compared with several classical ML models, and the determination coefficient of the adaptive boosting (AdaBoost) model reaches 0.96 after optimizing the features and hyperparameters. In this work, the running framework of the model was clearly displayed by visualization tool, and Shapley Additive exPlanations (SHAP) values were used to visually interpret the model locally and globally to help understand the predictive logic and the contribution of features. Furthermore, the accumulated local effect (ALE) successfully explains how the features affect the corrosion depth and interact with one another.