Federated incremental prognostics for system-specific degradation tracking in spacecraft propulsion systems

Accurate degradation prediction is essential for safe spacecraft operations during long-duration and autonomy-driven missions, including proximity operations around deep-space natural bodies where maintenance and intervention are limited. In these scenarios, propulsion systems must sustain stable performance over extended periods, making continuous health monitoring indispensable. However, propulsion units often exhibit distinct degradation behaviors due to differences in design, manufacturing tolerances, operational modes, and mission profiles, making cross-unit health assessment highly challenging. These challenges are further amplified by limited telemetry, early-stage data scarcity, and temporal drift, which restrict the ability of centralized or static learning approaches to capture consistent long-term degradation evolution. To address these challenges, this study proposes a federated incremental learning framework for spacecraft propulsion health monitoring that is specifically designed for system-level degradation tracking across heterogeneous fleets and evolving mission conditions. The framework incorporates a Global Momentum Aggregation (GMA) strategy to preserve long-term temporal inertia at the fleet level and a Multi-Agent Game (MAG) collaboration mechanism to coordinate contribution-aware knowledge sharing among heterogeneous propulsion systems, thereby suppressing inter-client inconsistency and alleviating early data scarcity. An enhanced LoRA-based incremental learning module further enables lightweight system-specific model evolution while maintaining global consistency, allowing newly observed degradation behaviors to be assimilated without disrupting previously learned trends. Collectively, these components stabilize long-horizon learning, respect system-specific degradation differences, and promote coherent fleet-wide evolution under heterogeneous operating conditions. Validation on the JAXA spacecraft propulsion dataset show that the proposed method outperforms existing approaches in accuracy and reliability.