<p>Continually identifying newly emerging faults in rotating machinery remains a critical challenge for intelligent industrial systems. In practical deployments, conventional continual learning methods are constrained by storage limitations and privacy restrictions, and the absence of historical data often leads to feature overlap and severe catastrophic forgetting. To address these issues, a new prospective topology-guided prototype network (PTGPN) is proposed for exemplar-free continual fault diagnosis. Specifically, a prospective embedding compression strategy is introduced to proactively allocate potential representation regions for each fault type in the feature space by enforcing intra-class compactness and inter-class separability, which reserves space for future fault type integration. In addition, a topology-guided prototype augmentation module is designed to generate more accurate prototypes of previously learned fault types by leveraging inter-class topological relationships, effectively reducing the feature overlap between old and new fault types and enhancing the model’s discriminative robustness during incremental learning. Furthermore, a prototype-constrained feature alignment module is incorporated to maintain spatial consistency of features between the current and previous models, guiding the model to retain its discriminative capability for previously learned fault types without relying on historical data. Extensive experiments on two rolling bearing datasets demonstrate that PTGPN achieves superior diagnostic performance and strong resistance to catastrophic forgetting under exemplar-free settings.</p>