Degradation prediction and explainable analyses of the corroded subsea pipelines based on INSGAII-PGNN and SHAP algorithm

• Published in: Ocean engineering Vol. 334; p. 121495
• Main Authors: Chen, Qian, Huang, Wei, Wen, Ya, Wang, Bohong, Liu, Gang
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.08.2025
• Subjects: Bias-variance trade-off optimization; Corrosion rate prediction; Interpretability analysis; Physics-guided neural networks; Subsea pipeline degradation; Corrosion rate prediction; Interpretability analysis; Physics-guided neural networks; Bias-variance trade-off optimization; Subsea pipeline degradation
• ISSN: 0029-8018
• DOI: 10.1016/j.oceaneng.2025.121495

This study introduces a hybrid model that integrates the corrosion mechanisms and the data-driven model to predict the corrosion degradation of subsea pipelines. Firstly, a corrosion mechanism-based feature crossing method is employed to improve the quality of features, and the null importance and the recursive feature elimination method with cross-validation are adopted to select the optimal feature subset. Then, the physics-guided neural network is established to predict the corrosion rate based on the physical impact of various features on corrosion degradation, and the improved NSGA II algorithm is proposed to optimize the hyperparameters based on the variance-bias trade-off theory. Finally, the effectiveness of this model is validated through corrosion mechanisms and interpretability methods, and the effect of each feature on corrosion degradation is analyzed. Results demonstrate that the proposed integrated model can achieve a superior balance between bias and variance compared to other models, and the mean absolute percentage error can be reduced by 1.848 %, 1.680 %, 0.280 %, and 0.316 % compared to the DNN, SVR, RF, and AdaBoost model, respectively. The hybrid model can effectively prevent the data-driven model from learning the patterns conflicting with corrosion mechanisms. This research has significant implications for subsea pipelines' integrity and reliability analyses. •Typical features of pipeline corrosion are selected by the null importance and the REFCV.•The physics-guided neural network predicts subsea pipeline corrosion degradation via mechanism knowledge.•A multi-objective optimization model determines PGNN hyperparameters via variance-bias trade-off theory.•The corrosion degradation impact of each feature is quantitatively assessed via PDP, ICE, and SHAP explainability.