Adaptive approach for estimation of pipeline corrosion defects via Bayesian inference

Abstract

A framework to construct a model that predicts the corrosion defect distribution using a small amount of observation data is proposed in this study. A time-dependent generalized extreme value distribution was employed to consider the changing corrosion growth rate with time, and model parameters were estimated via Bayesian inferences to develop a robust prediction model. The model parameters were updated when a new batch of inspection data was available; previous data were not directly used but they indirectly assisted parameter estimation in the form of a prior distribution. In addition, an artificial data point representing a larger defect depth was added to the inspection data to ensure a conservative estimation of the model parameters and higher reliability of the model. The model was verified under three different cases, and the results showed that the suggested parameter estimation allowed the prediction model to adapt to the changing defect depth distribution in all three tested cases: 1) inspection data are available without measurement errors, 2) inspection data are available with measurement errors, and 3) the properties of the underground environment are drastically changed.