A novel approach for reliability assessment of corroded offshore pipelines using machine learning and random sampling
Abstract
Offshore pipelines are critical for hydrocarbon transport but face significant reliability challenges due to internal corrosion, particularly from interacting defects. This study introduces a novel framework integrating Latin Hypercube Sampling (LHS), Finite Element Analysis (FEA), and machine learning (ML) to assess the reliability of offshore pipelines with longitudinally aligned corrosion defects. Using LHS, 200 initial samples were generated, producing a dataset of 266,200 points through FEA in ABAQUS, capturing Maximum Von-Mises Stress (MVMS) across 121 internal pressure levels (1-30 MPa) and 11 defect spacing intervals (0-2 root Dt ). Five ML models-Linear Regression, Stochastic Gradient Descent, K-Nearest Neighbors, Decision Tree Regression (DTR), and Neural Network (NN)-were trained and optimized using k-fold cross-validation and grid search. Model performance was evaluated through statistical metrics (RMSE, MAE, R-2), probabilistic assessments using Kernel Density Estimation (KDE), code-based comparisons against DNV-RP-F101 standards, and learning curve analyses to ensure robustness against overfitting. The NN model was identified as the most effective, achieving an RMSE 15 % lower than DTR, with 86.6 % of prediction errors within +/- 10 MPa. Permutation importance analysis highlighted pipeline thickness, material properties, and defect depth as key predictors of MVMS. The framework's efficacy was demonstrated by expanding the dataset from 200 to 2000 samples in a case study, revealing up to 8 % improvement in reliability assessment accuracy. This framework enhances pipeline integrity management by enabling precise, cost-effective reliability assessments, with implications for optimizing maintenance strategies in the oil and gas industry.