A Deep Convolutional Neural Network for Predicting the Failure Response of High-pressure Gas Pipes subject to Pitting Corrosion

S. Soghrati Ohio State University, United States

Keywords: corrosion, convolutional neural network, automated modeling, finite element analysis

Summary:

Pitting corrosion is one of the main causes of mechanical failure in high pressure gas pipelines in aggressive environment. Various techniques such as ultrasound pig equipment can be used to characterize the size and shape of pits along the pipe length. However, linking the pits morphological and statistical features to the failure response of the pipe remains a challenge, as pits are often much smaller (few hundred microns) than characteristic length scales of the pipe. Further, the inherent scattering in their size distribution and spatial arrangement necessitates simulating the failure response of thousands of statistical volume element (SVEs) in different regions to determine the corresponding loss of strength and toughness due to pitting attack. In this work, we implement an automated computational framework allowing the virtual reconstruction of realistic 3-D SVEs of the pipe with various shapes, sizes, and spatial arrangement of pits. A novel meshing algorithm is then implemented to build high-fidelity finite element (FE) models of each SVE and simulate its nonlinear failure response using a continuum ductile damage model. Despite the automation of this process, each simulation requires several minutes on massively parallel platforms. This huge computational burden makes it practically impossible to fully characterize the failure response even across a short length (e.g., one kilometer) of the pipe. To address this challenge, this automated reconstruction-meshing-simulation framework is employed to train a deep convolutional neural network (ConvNet). Several algorithmic aspects pertaining to reducing the bias and variance in the resulting ConvNet predictions, including data augmentation, enhancing input imaging data, and optimal network architecture are discussed. We show that while the predictions made by the proposed ConvNet yields an average error of less than 5% compared to high-fidelity FE simulations, the corresponding computational cost is dropped by more than 5 orders of magnitude.