Abstract: In response to the problems of low efficiency and prominent contradiction between quality and efficiency in the traditional manual inspection mode of the railway Trouble of moving Freight car Detection System (TFDS), this paper proposed a TFDS fault intelligent recognition scheme based on large model technology. It adopted a distributed architecture of "central training, edge inference", combined with large model pre training, hierarchical recognition, and workflow integration technology to solve the problem of fault recognition accuracy and sample imbalance in complex scenarios, and constructed a sample library using multi-source data, combined semi-automatic annotation with forward and backward iterative training to optimize the model, adopted strategies such as data augmentation and local detection boxes to improve inference performance. The experimental results and on-site trials show that the proposed scheme achieves recognition rates of 100%, 99.91%, and 99.85% for Class A, B, and C faults, respectively. The false alarm rate and recognition time in on-site testing are both better than industry standards. The efficiency of homework has increased by 43%, with personnel decreased by 28.6%, which verifies the effectiveness and practicality of this scheme in intelligent detection of railway freight cars.