Abstract: New fault types may continually emerge under varying operating conditions, while conventional domain adaptation methods often fail to recognize them as unknown. To address this issue, this study proposes PC-AED, an open-set cross-domain fault diagnosis method driven by prototype-guided contrastive learning and adversarial entropy decoupling. The method uses multi-view augmentation in the time, frequency, and time–frequency domains to improve the stability of vibration features under operating-condition changes and noise. Dynamic class prototypes are used to constrain the feature distributions of the source and target domains. This promotes the aggregation of same-class features and the separation of different-class features. Source-domain predictive entropy is used to determine a dynamic uncertainty threshold. Adversarial learning then refines the decision boundary between known and unknown classes. Nineteen open-set cross-domain diagnosis tasks were constructed using the CWRU and SEU bearing datasets. PC-AED achieved mean H-scores of 98.04% and 93.83% on the CWRU and SEU datasets, respectively. These values exceeded those of the best-performing baseline method by 6.22 and 4.48 percentage points, respectively. The results show that PC-AED maintains its ability to recognize known faults under varying operating conditions while effectively detecting unknown faults. It provides a feasible approach to rotating machinery fault diagnosis in open industrial environments.