Vision transformers for three‐phase flow classifications with data augmentation through generative adversarial networks

Abstract The identification of multiphase flow regimes is important for the efficient design and operation of upstream oil and gas production pipelines. Convolutional neural network (CNN) is widely used to classify flow regimes. However, CNN usually fails to classify the small and imbalanced datasets accurately. In the current work, data augmentation is carried out using Wasserstein generative adversarial networks‐gradient penalty (WGAN‐GP) to remove the class imbalance. The Vision Transformers (ViT) model pre‐trained on the ImageNet‐21k dataset has been employed for classification. The architecture of ViT is simple and robust, which can extract dynamic features of multiphase flow images. Unlike CNN, ViT has successfully distinguished all the classes and performed consistently well across all the class systems. The methodology developed for the current case has provided a novel framework for the identification and classification of imbalanced and small datasets of multiphase flow regimes occurring in upstream oil and gas pipelines.