Machine learning-based maximum pipeline pitting corrosion depth prediction using hybrid FVIM-BNN-XGB model

The pronounced nonlinear characteristics of corrosion depth in buried pipelines present significant challenges to the accurate characterization capabilities of traditional experimental and statistical methods. To address this challenge, the study proposes a hybrid machine learning framework. First, a multivariate feature engineering approach is employed, integrating Pearson correlation analysis, SHapley Additive exPlanations (SHAP) values, and backward stepwise feature selection (BSFS) to identify critical features, with particular emphasis on environmental factors. Subsequently, a feature extractor combining a Bayesian Neural Network (BNN) and XGBoost is constructed to capture residual patterns and enable model fusion, thereby significantly enhancing model performance. Furthermore, a five-fold cross-validation strategy is implemented to improve model stability and generalization, particularly under conditions of limited sample. Additionally, the Four Vector Intelligent Metaheuristic (FVIM) is used to optimize model parameters, minimizing weighted relative error and enhancing prediction reliability. Experimental results demonstrate that the proposed hybrid model achieves substantial improvements in predicting maximum corrosion depth (Dmax), outperforming ten existing benchmark models. This work highlights the potential of the hybrid machine learning framework in addressing highly nonlinear problems and overcoming the limitations of traditional methods, offering valuable insights for similar scientific research and practical engineering applications.