An Intelligent Identification Method for Offshore Drilling Kick Conditions Based on Remotely Operated Vehicle Images: Coupled Convolutional Neural Network and Generative Adversarial Network

Summary During shallow open-circuit drilling, utilizing remotely operated vehicle (ROV) images could be a feasible approach for kick monitoring. However, it faces challenges that include poor image quality and the complexity of overflow image discrimination, which has not been well studied so far. In this work, we propose an intelligent kick recognition model based on ROV images that couples convolutional neural network (CNN) and generative adversarial network (GAN). It consists mainly of (i) a GAN model for data augmentation of kick image data set to address the issues of poor image quality and data imbalance in ROV images and (ii) a ResNet model incorporating convolutional block attention module (CBAM) for kick images classification and recognition to improve recognition accuracy and efficiency. Extensive experimental results demonstrate that our method exhibits real-time performance and high accuracy in identifying kicks. The diagnosis time for a single kick image is reduced to as low as 0.015 seconds, enabling real-time judgment. Furthermore, the accuracy of our model increases to 91.04%, whereas the accuracies of traditional models including Inception V3, AlexNet, VGG, and GoogLeNet are 87.73%, 86.64%, 85.42%, and 80.86%, respectively. After data augmentation by GAN, the accuracy of our method can be improved by 11.21%.