PD-FedOS: Prototype-driven federated open-set learning framework for collaborative intelligent fault diagnosis of aero-engine rotor systems

Aero-engine rotor systems fault diagnosis is a key component in safeguarding flight safety. In practice, monitoring data is distributed across multiple engines or organizations, motivating the widespread adoption of federated learning (FL). However, unseen fault states inevitably emerge during operation, violating the closed-set assumption of most FL-based methods and leading to elevated false alarm rates and degraded diagnostic performance. To address this challenge, a prototype-driven federated open-set learning framework (PD-FedOS) is proposed to enable open-set fault diagnosis of aero-engine rotor systems using heterogeneous and distributed data. Within the proposed framework, feature prototypes are introduced as shared class representations that are locally computed at each client and aggregated at the server to encode global diagnostic knowledge, thereby mitigating inter-client feature space heterogeneity. A distance-based cross-entropy loss is further devised based on the global prototypes to align local features with their corresponding prototypes while pushing them away from non-corresponding ones, thus enhancing feature consistency and discriminability.On this basis, a distance-based rejection rule is derived to enable reliable identification of unseen fault states in open-set scenarios. Finally, extensive experiments involving six representative comparative methods conducted on two aero-engine rotor system fault datasets demonstrate that the proposed PD-FedOS achieves the highest diagnostic accuracy and superior robustness in both closed-set and open-set federated fault diagnosis scenarios.