Two-stage dual-domain clustering fusion model for serviceability assessment in flywheel bearings

Flywheel bearings, as core rotating components, require accurate serviceability assessment to ensure system reliability. Conventional fault monitoring often suffers high false alarm rates due to the volatility and data scarcity of cage friction faults, causing over-rejection and waste. This paper proposes an online assessment model based on two-stage dual-domain clustering fusion. First, an improved ensemble empirical mode decomposition (EEMD) is applied to denoise vibration signals and enhance fault features. Characteristic frequency ratio (CFR) and other features are constructed and fused. A normal–fault bipolar positioning strategy trains two clustering domain models with both normal and fault samples, and serviceability is quantified via bidirectional spatial mapping distance, reducing ambiguous judgments. An enhanced two-stage clustering fusion mitigates domain overlap, improving identification accuracy. Engineering validation confirms the method’s capability to identify bearing states with limited samples, offering a high-accuracy, low–false-alarm solution for aerospace maintenance.