Low-grade fault detection via vision transformer and U-net fusion model

Accurate identification of low-grade faults is crucial for the exploration and development of residual oil, as well as for predicting and resolving injection-extraction conflicts. Low-grade faults have small offset distances, strong concealment, and short lateral extensions, resulting in poor prediction by traditional semantic segmentation networks seen in the literature. The U-net model focuses on detailed information but weakly captures global context, whereas the Vision Transformer (ViT) excels in global perception with its ability to incorporate information from all image locations. Therefore, this study proposes a fault recognition method, VTRUnet, which fuses ViTs and U-net networks. The encoder is constructed using the ViT module to efficiently learn 3D seismic data sequences, which in turn effectively captures long-range visual context by capturing global dependencies in the image using multiple heads of attention. The decoder is then constructed using U-net to extract local information. The features extracted from the encoder are directly merged with the U-net-based decoder through a skip connection to maximize the preservation of details and semantic information in the original input data. The joint loss function is used to address the sample imbalance problem and enhance the generalization ability of the model. Finally, the constructed model was applied to fault recognition on synthetic and actual data, which was superior to that of traditional U-net and SE-UNet architectures. The results show that the method can effectively solve the problems of blurred fault cross-location, thickened contour edges, recognition discontinuity, and omission of dense regions and improve low-grade fault recognition accuracy.