After decades of rapid development, our rail transit system has entered a period of frequent maintenance of equipment. It is faced with challenges of various equipment models, complex working conditions, a huge amount of fault data, etc. To meet the requirement of real-time fault diagnosis of the rail transit system to guarantee safety, a distributed training architecture is proposed to accelerate the training of a large-scale data fault diagnosis model. On the other hand, rail transit fault data is highly confidential and private, while the traditional centralized data fault diagnosis has a high risk of data leakage. Therefore, the Federated Learning (FL) paradigm is adopted in this paper and it can effectively alleviate the "data island" problem. From the experimental results of this paper, on the one hand, we successfully applied the FL paradigm to the field of rail transit fault diagnosis, ensuring the privacy and security of fault data. On the other hand, the distributed training architecture has been greatly improved in terms of training speed and computing resource utilization.