In practical engineering applications, rotating machinery typically operates under normal conditions, which can result in nonideal datasets with few or even partially missing fault samples. To address the low accuracy issue in deep learning diagnosis models trained directly on non-ideal datasets, a fault diagnosis method that incorporates finite element simulation to facilitate transfer learning is proposed. Firstly, vibration signals with different operating conditions and fault types are derived via numerical simulations. Subsequently, a large number of cost-effective and high-fidelity simulation samples are employed to pre-train a diagnostic model, and the authentic limited dataset or the hybrid dataset augmented by simulation samples is employed to fine-tune the pre-trained diagnostic model. This approach aims to implement high-precision fault diagnosis and mitigate the reliance on actual or experimental fault data. Finally, two bearing datasets are used to evaluate the effectiveness of the proposed method. Results show that the diagnostic model constructed via the proposed method achieves an accuracy exceeding 95% with a sample size of one for each fault category. In addition, in cases where the fault samples are limited and certain types of faults are missing, the accuracy is boosted by over 10% compared to the approach of supplementing the simulation samples directly.