Fault Recognition Method and Application Based on Generative Adversarial Network

Abstract

In view of the limitation of generalization ability faced by deep learning in fault identification, especially in the case of complex underground geological conditions and variable seismic data characteristics, it is often ineffective to directly use the network based on synthetic data training for fault prediction of real data. To overcome this challenge, this study proposes an innovative solution, which uses generative adversarial network-UNet (GAN-UNet) to extract features from data in depth. The network employs a U-net architecture as the backbone to simultaneously extract all features from forward-modeled synthetic data and real seismic data. These features are utilized as inputs for both the fault classifier and discriminator. The fault classifier distinguishes between fault and non-fault segments, while the discriminator employs adversarial mechanisms to differentiate whether input features originate from real seismic data or synthetic data. Once the discriminator, after training, cannot accurately discern the precise source of features, the network model has effectively uncovered the fundamental shared features between the two datasets. This approach demonstrates effective fault recognition in practical seismic data. To verify the effectiveness of the method, we applied it to the actual seismic data sets of the North Sea F3 block and the western deep basin. The experimental results show that compared with the traditional deep learning method, this method shows significant advantages in fault recognition. It not only improves the accuracy of fault identification, but also enhances the adaptability of the model to complex geological conditions.