<p>Traditional data-driven fault diagnosis methods often suffer from performance degradation when applied across different machines or operating conditions. To overcome this challenge, we propose a dynamics-enhanced vision transformer (Dynformer) framework for cross-domain rolling bearing fault diagnosis. The method integrates deterministic learning and transfer learning in a semi-supervised manner. First, deterministic learning is employed to extract intrinsic system dynamics from vibration signals, which are then encoded into dynamics information images. These images serve as domain representations for both source and target domains. Second, a vision transformer with adversarial pre-training is designed to learn domain-invariant features from these representations. Finally, a small number of labeled samples from the target domain are used to fine-tune the pre-trained model, enabling accurate and robust fault classification under varying machines and working conditions. The effectiveness of Dynformer is verified through experiments on two public bearing datasets, where it demonstrates competitive accuracy and generalization across cross-domain scenarios.</p>