Aiming at the serious negative transfer brought about by the excessive domain discrepancy of virtual-real data in the simulated data-driven transfer fault diagnosis method, this paper proposes an unsupervised fault diagnosis method based on a virtual-to-real domain multi-level Joint adaption network (VDMJAN). First, the dynamic simulation model of various types of bearing faults with irregular patterns was established, and the virtual signals of the faults under real working conditions were generated and analyzed in terms of transportability modes in combination with the fault mechanism. Then, deep convolutional neural networks with different sizes of convolutional kernels are constructed to perform feature extraction at different granularities for virtual and real domain signals, which enhances the domain invariance and comprehensiveness of the features used for state discrimination. Finally, based on the pseudo-label labeling of the simulated samples using the probabilistic fusion of parallel outputs from multiple classifiers, the simulated and measured samples are finely aligned with domain features at different levels to achieve cross-domain consistency of the real and virtual data. The introduction of the effectiveness loss of VDMJAN training guarantees the certainty and consistency of the multi-classifier recognition of the state of real samples in the target domain. The class-aligned measured samples are filtered and used to calibrate the classifier to better adapt to the target domain classification task. The results of two experimental analyses show that the proposed VDMJAN, driven by bearing fault simulation data in a single source domain, can realize effective simulation-reality fault diagnosis and has good application prospects in the field of equipment fault diagnosis where samples are scarce in special environments.