Corrosion-like Defect Severity Estimation in Pipelines Using Convolutional Neural Networks

The study of methodologies for the effective monitoring of structural damage and leaks in oil and gas pipelines has been a relevant topic in recent years due to the importance of avoiding fatalities, environmental disasters, and economic losses. In this context, the present study aims at developing a deep learning-based framework for estimating the severity of corrosion-like defects in pipelines. The dataset used in this work was generated by finite element simulation of a simple cylinder with different defect depths. The simulation was carried out with the commercial software On-scale™. In this way, 100 samples were obtained representing diverse structural conditions. A Convolutional Neural Network (CNN) architecture based on LeNet-5 was proposed to address the problem. This architecture was trained and tested 100 times with a Monte Carlo Cross-Validation procedure. Results show that the proposed architecture achieved good performance, with mean RMSE and R2 of 0.4448 and 0.9637, respectively.