Beyond linearity: Machine learning and probabilistic models for thickness degradation and failure prediction in LPG pressure vessels

This study evaluates three modelling approaches, linear regression, support vector regression with a polynomial kernel (SVR-Poly), and Gaussian process regression (GPR), for predicting corrosion-induced wall thickness degradation and estimating remaining service life in four LPG pressure vessels under coastal operating conditions. Six annual ultrasonic thickness measurements (2006–2011) were collected at component-specific critical monitoring locations (CMLs) selected per API 510 on the basis of minimum recorded thickness, with the shell governing Vessels 1 and 2 and the head governing Vessels 3 and 4. A preliminary linearity assessment using R 2 , RMSE, and the coefficient of variation of year-to-year corrosion rates classified vessels 1 and 4 as quasi-linear and vessels 2 and 3 as non-linear, with Vessel 2 recording a pronounced late-stage acceleration from 0.05 to 0.25 mm/year. GPR achieved the highest predictive accuracy, with an average R 2 of 0.9727 against 0.9269 for linear regression and 0.9381 for SVR-Poly, reducing RMSE by 40.6% and MAE by 27.6% relative to linear regression. The largest improvement occurred at Vessel 2, where GPR raised R 2 from 0.8232 to 0.9218. Failure year predictions diverged by up to four years between models on the most non-linear vessel. Bootstrap resampling over 1000 iterations produced vessel-specific uncertainty estimates that correlated directly with corrosion rate variability. Integration into a semi-quantitative risk-based inspection framework showed that model choice shifted risk classification across three categories for Vessel 1, while Vessels 2, 3, and 4 reached critical status under all models.