A cross-domain diagnosis approach combining Digital Twin (DT) and Federated Learning (FL) is proposed to further optimize the issues of sample scarcity and privacy protection in early defect diagnosis of wind turbine transmission systems. A DT which combines both mechanism- and data-driven approaches is proposed to generate multi-condition virtual fault data. Next, the feature alignment strategy is used to reduce domain difference between simulated and real data. A cross-wind FL framework is proposed, which uses dual-path feature learning and dynamic aggregation strategy combined with data quality evaluation for synchronous training. The experimental results on Coverage, Accuracy, Reliability and Earliness (CARE) dataset show that: 1) the average F1 score of the proposed DT-FL model on three wind farm test sets is 0.942, which is better than that of the FL (F1 score is 0.907) and the standard federated average algorithm (F1 score is 0.925) which only use measured data. 2) In the cross-wind migration task, the average performance attenuation of DT-FL model is 7.6%. It has better generalization ability. The above results show that the proposed model can effectively use virtual data to enhance the diagnostic capabilities and improve the early fault identification performance under the premise of protecting data privacy.