Application of Deep Learning Fault Detection and 3D Fault Structure Modeling in the Tarim Basin

Fault detection holds significant importance for the exploration and development of strike-slip fault-controlled fracture-cavity reservoirs in the Tarim Basin. Strike-slip faults in seismic data are typically characterized by small fault displacements, weak seismic response energy, and ambiguous reflection features, rendering fault detection particularly challenging. With recent advancements in deep learning technology, convolutional neural network-based fault identification methods have proliferated. Noting that deep learning approaches are fundamentally data-driven, the generation of extensive high-precision fault labels becomes crucial. This study employs 3D fault structure modeling to generate training data and corresponding labels. Through feature migration techniques, the synthetic training data is imbued with realistic seismic amplitude characteristics and noise features from actual field data. A Res-Unet architecture is subsequently constructed for network training and testing, with validation conducted using real seismic data from the Tarim Basin. Test results demonstrate that the network effectively captures essential features from field data, with method reliability confirmed through standardized testing protocols. Comparative analysis shows superior fault detection performance over conventional deep learning approaches using generic fault labels, thereby validating the methodology's effectiveness. Furthermore, this approach enables the generation of substantial reliable training data with field-consistent fault labels, significantly enhancing the generalization capability of deep learning-based fault detection methods. The proposed workflow provides an innovative technical framework for precise identification of subtle strike-slip fault systems in complex carbonate reservoirs.