Interpretable failure prediction modeling of hydrogen-blended natural gas pipelines containing a crack-in-corrosion defect

In this study, a CIGWO-GPR machine learning (ML) model is developed, integrating Circle chaotic mapping, grey wolf optimization, and Gaussian process regression. The proposed hybrid model is developed to predict the failure pressure of X80 hydrogen-blended natural gas pipelines (HBNG) containing a crack-in-corrosion defect (With a hydrogen blending ratio below 15 %). The model interpretability is enhanced with SHapley Additive exPlanations (SHAP), which captures the contributions of global and local features. The results demonstrate that the hydrogen blending ratio and geometries of corrosion and crack highly correlate with corrosion- and crack-induced failure pressure. The CIGWO algorithm significantly improves the model accuracy and global search capability of GPR models, which has superior performance for high-dimensional nonlinear data. The SHAP analysis indicates that the hydrogen blending ratio and corrosion depth are the critical factors for crack-induced failure pressure (P_crack) and corrosion-induced failure pressure (P_corrosion), respectively. The proposed method can serve as a valuable reference for artificial intelligence-based integrity management of HBNG pipelines.