Metal-loss defect depth inversion in oil and gas pipelines based on Bayesian regularisation neural network

Abstract

Defect depth inversion is generally considered as a challenge in magnetic flux leakage (MFL) testing and evaluation because of its strong non-linearity and low prediction accuracy. Current inversion models focus on the inversion accuracy of specific datasets, ignoring consideration of the generalisation ability of inversion models under different conditions. In order to solve such problems, this paper proposes a novel pipeline defect inversion method based on a Bayesian regularisation neural network (BRNN) model. This method consists of two parts. Firstly, three domain features are extracted and a Boruta algorithm is introduced to reduce the feature dimension and obtain the best feature subset. Secondly, in order to approximate the complex non-linear relationship between multi-dimensional features and defect depth, a back-propagation neural network (BPNN) model based on Levenberg-Marquardt optimisation and a Bayesian learning algorithm is constructed. The model can effectively find a close global minimum and overcome the phenomena of overfitting and overtraining. In order to evaluate the performance of the proposed defect inversion method, a comparative experiment is carried out with other well-known inversion algorithms. The results obtained confirm that the inversion method can improve the prediction accuracy of defect depth. More importantly, this method enhances the generalisation ability of defect inversion problems with different sample sets.