Indirect Inspection-based Bayesian Machine Learning Model for Probabilistic Coating Defect Severity Interpretation

Abstract Corrosion pit nucleation and propagation in underground pipelines vary by geographic position along the pipeline right of way (RoW). Effective pipeline integrity management requires both the modeling of corrosion initiation/propagation and the integration of field-acquired data. To mitigate external corrosion, onshore pipelines use barrier coatings and either sacrificial or impressed current cathodic protection (CP) systems. Pipeline integrity is monitored through direct and/or indirect assessments. In this work, we focused on leveraging indirect methods, such as close interval potential surveys (CIPS) and direct current voltage gradient (DCVG). CIPS evaluates CP effectiveness by identifying deficiencies in CP-equipped structures via pipe to soil potential and detecting potential 'hot spots' in those with insufficient CP. Meanwhile, the DCVG method identifies coating flaws by measuring voltage gradients using reference electrodes placed in the soil. But the reliability of defect detection is a practical concern. To enhance corrosion assessment reliability, we propose a Bayesian framework combining indirect inspection data, corrosion science, and artificial intelligence. Using Stochastic Variational Inference, the framework predicts corrosion severity and rates. Validation against inline inspection data demonstrates its accuracy and confidence in identifying corrosion-prone areas, thus improving pipeline integrity management and reducing failure risks.