为了充分提高小断层识别精度，研发了人工智能小断层预测方法，通过基于小断层解释成果的样本标签制作，构建基于解释成果的样本标签库；研发了针对小断层的建模和正演方法，构建基于模型正演的标签库；形成针对小断层识别的特殊神经网络，可直接形成小断层属性数据。此外，结合去强轴、构造保护平滑滤波等预处理技术提高原始地震数据的品质，并利用基于主成分分析的多属性融合反映多尺度断层信息，最后通过3种手段验证预测结果，形成了基于人工智能的小断层预测流程。本文方法和技术在鄂尔多斯盆地大牛地气田展现了良好的应用成效，相比于常规属性，小断层识别数量增加30%，预测结果的分辨率和连续性显著提升，与地震数据、井数据和常规属性吻合较好，并揭示了小断层发育特征：优势方位为北西走向，工区西部和东北部的小断层相对集中，断层间的裂缝也较为密集，部分小断层沿着储层的边界分布，或许和优质储层的发育有关。预测结果为评价下古生界优质储层、开展井位部署提供可靠的技术支撑，在类似地区具有广阔的应用前景。

In order to improve the accuracy of subtle fault identification, an artificial intelligence subtle fault prediction method based on the seismic data was developed. By making sample labels based on subtle fault interpretation results, a sample label library based on interpretation results was built. The subtle fault modeling and forward methods for subtle faults were developed, and a label library based on model forward was built. The special neural network for identifying subtle faults was developed, which can directly generate attribute data for subtle faults. In addition, the seismic preprocessing approaches such as removing strong seismic events and structure oriented smoothing filtering were added to improve the original seismic data. Multi-attribute fusion based on principal component was used to reflect multi-scale faults. Finally, the prediction results were verified through three steps, forming the subtle fault prediction workflow with artificial intelligence. This study demonstrates good application in the Daniudi gas field in the Ordos Basin. Compared with conventional attributes, the number of subtle fault identification has increased by 30%, and the resolution and continuity of the subtle faults have significantly improved. The prediction results are consistent with seismic data, well data, and conventional seismic attributes. Based on the subtle faults identification, the new understandings of regional structure are revealed: the dominant orientation of subtle faults is northwest, the most subtle faults are relatively concentrated in the western and northeastern parts of the survey, and the fractures between faults are also very dense. Some subtle faults are distributed along the boundary of the high-quality reservoir, perhaps related to the development of reservoirs. The prediction results contribute to evaluating high-quality reservoirs in the Lower Paleozoic and well deployment, and have application prospects for similar areas.