Abstract: Deep learning provides a novel approach for automatic fault recognition in seismic data, and has achieved remarkable results. However, several challenges remain in practical applications, including weak cross-regional generalization capability, large discrepancies between synthetic samples and actual faults, limited diversity of synthetic sample types, and occasional unexpected fault-recognition results. To address these issues, a modified ResU-Net network was developed by combining the strengths of ResNet and U-Net. The model’s initial parameters were obtained through pretraining on a synthetic dataset. Using the fault probability volumes calculated by the pretrained model and considering the actual fault characteristics in the study area, artificial samples were interpreted for transfer learning, and the network parameters were optimized. Given the subjectivity of manual interpretation, these artificial samples were updated and refined based on the first transfer learning results and the observed fault development patterns. Additional artificial samples were created at locations of previously unrecognized faults, and iterative transfer learning was performed until the results met the desired accuracy. Using this approach, faults in Structure A of the Xihu Sag were automatically identified with high precision. The fault development characteristics of Structure A were reinterpreted, revealing that faults in the anticlinal wings have higher grades than those in the anticlinal core and that hydrocarbon migration conditions are more favorable in these areas. This finding highlights a promising direction for structure-lithology and lithology exploration. The method supported the drilling of three exploration wells, all of which resulted in large-scale commercial discoveries, confirming its reliability.