Machine learning approach for remaining useful life prediction for oil pipelines under data scarcity: A risk-based inspection framework

Accurate prediction of the remaining useful life for ageing oil pipelines is critical for integrity management; however, conventional machine learning methods are limited by sparse inspection data. This study introduces a support vector regression-based approach for corrosion-induced wall thickness prediction and risk-based inspection prioritization, applied to eight oil transmission pipelines, each with only three inspection records spanning 15–32 years. Using a polynomial kernel with hyperparameter optimization, support vector regression was compared against the linear extrapolation model standard in industry practice. Although both models fit the training data comparably well, their predicted failure years diverged by up to 12.5 years, indicating a non-linear degradation behaviour that linear extrapolation cannot capture even when the training error is small. These predictions were integrated into a first-order second-moment probability-of-failure model, propagating corrosion rate and prediction uncertainty into quantitative risk scores and inspection priorities for practical risk-based inspection planning.