Data augmentation using conditional generative adversarial network (cGAN): Application for prediction of corrosion pit depth and testing using neural network
Haile Woldesellasse, Solomon Tesfamariam
Abstract
Machine learning (ML) based algorithms, due to their ability to model nonlinear and complex relationship, have been used in predicting corrosion pit depth in oil and gas pipelines. Class imbalance and data scarcity are the challenging problems while training ML models. This paper utilized a conditional generative adversarial network (cGAN) to handle class imbalance problem in a corrosion dataset by generating new samples. Utility of the cGAN data augmentation is evaluated by training an artificial neural network (ANN) model. In addition, random oversampling and Borderline-SMOTE data generating techniques are used for comparison with cGAN. The testing accuracy of the ANN model increased greatly when trained by the cGAN based augmented dataset and this model performance improvement can be useful for a pipeline integrity management.