Intelligent prediction model for pitting corrosion risk in pipelines using developed ResNet and feature reconstruction with interpretability analysis

Abstract

In coastal and offshore environments, oil and gas pipelines are subjected to harsh environmental conditions, including high temperatures, humidity, and salt fog, which accelerate corrosion and deterioration. These factors significantly constrain pipeline lifespan, increase maintenance costs, and pose safety risks. Accurate prediction of corrosion rates is critical for optimizing site selection, construction, and operational strategies—forming a cornerstone of corrosion management in pipeline systems. While existing models predominantly prioritize predictive accuracy, their exploration of the relationships between influencing factors and pipeline pitting depths remains limited. To address this gap, this study introduces an enhanced residual neural network—integrating feature reconstruction—to evaluate pipeline pitting risks. Utilizing Kernel Principal Component Analysis (KPCA) and empirical formulas, the approach identifies key factors most closely correlated with pitting depths. Validation via practical engineering cases demonstrates that the proposed D-ResNet model achieves a MAE of 0.4616, MAPE of 0.3830, and RMSE of 0.5896—reducing errors by 31.6 %, 32.1 %, and 34.9 %, respectively, relative to baseline models. Furthermore, the BowTie framework incorporates SHAP (Shapley Additive exPlanations) analysis to enable interpretable risk characterization, revealing underlying mechanisms and providing a comprehensive methodological basis for lifecycle pipeline integrity management.