During the Late Paleozoic marine-continental transitional stage, the extensive coal-bearing strata were formed in Ordos Basin. Among them, the No. 8 coal seam (hereafter 8# coal) in Benxi Formation is the primary target for exploration and development, and horizontal wells are widely employed to enhance production. However, subtle micro-faults frequently develop around the coal measures, and highly concealed which leads to the risks of lost-circulation and borehole-collapse. Therefore, it’s necessary to research the concealed-faults identify method to support rapid and optimized drilling. Conventional fault-detection techniques rely mainly on geometric attributes such as coherence, variance and curvature, but their noise immunity is poor and lateral continuity is limited, making them ineffective for identifying these subtle faults. Taking advantage of high-precision pre-stack depth-migrated seismic volumes, the fault-constrained structure-oriented filter was used to sharpen fault-tip imaging. Then the Transformer network algorithm was introduced, and the localized training label set that honors the fault characteristics of coal-bearing intervals was established to build an AI-driven fault-prediction model. The resulting intelligent fault volume achieves F₁-scores and intersection-over-union values significantly higher than those of traditional coherence and curvature attributes, providing a reliable basis for horizontal-well trajectory optimization. After completing the support for the eight horizontal coalbed methane wells, no well leakage or collapse incidents occurred, and the research findings contributed to the rapid production development of deep coalbed methane in Ordos Basin.