Multiresolution analysis and deep learning for corroded pipeline failure assessment

•A parametrization of real corrosion shapes using discrete wavelet transform was proposed.•Synthetic models were created to reproduce real complex corrosion profiles.•The FE automatic modeling and non-linear analysis procedure were validated by comparing with experimental results and semi-empirical models.•A solution to predict the failure pressure using a deep neural network (DNN) was proposed. The assessment of the safety of corroded pipelines is considered a vital task in the oil and gas industry. This research aims to develop an efficient system to accurately predict the burst pressure of corroded pipelines with complex corrosion profiles through hybrid models using multiresolution analysis, numerical analysis and meta-models. The work addresses the parametrization of real corrosion shapes and its use as input to a neural network system that can predict the burst pressure accurately and quickly. Ultrasonic inspections provide the thicknesses in the corrosion area and with that data the river-bottom profile (rbp) can be constructed. A set of scripts creates finite element models from the rbp, which are then subjected to non-linear analyses to determine the failure pressure. In this work, the Finite Element models and analysis procedure are validated against experimental tests and are compared to semi-empirical assessment methods. A discrete wavelet transform is performed for the parametrization of the remaining thicknesses and it is used as a filter bank to reduce the amount of data that describes the defect. The coefficients obtained from the wavelet transform are used as inputs to feed a neural network to provide a quick evaluation of the failure pressure. This neural network is trained with the results of finite element analyses performed on synthetic models built with statistical properties similar to those of real corrosion defects The results from the neural networks are obtained very fast and accurate for all cases presented in this work.