Adapting Without Source Data: High-Confidence Knowledge-Guided Source-Free Domain Adaptation for Rotating Machinery Fault Diagnosis

Accurate condition monitoring of rotating machinery relies on the stability of the vibration measurement-to-decision mapping. In practical deployments, variations in operating conditions, together with source-free constraints where historical source data are unavailable, often induce distribution shifts in vibration measurements, leading to unstable representations and unreliable fault diagnosis. To address this challenge, a High-Confidence Knowledge-Guided Source-Free Domain Adaptation network (HCKNet) is proposed. Within this framework, a Confident Anchor-Induced Self-Refinement (CASR) mechanism is designed to purify noisy pseudo-labels by exploiting high-confidence target anchors as structural priors, which mitigates error accumulation and stabilizes the adaptation process. In addition, a Cross-Space Consistency Regularization (CSCR) mechanism is formulated as a complementary constraint to suppress residual local prediction inconsistency and improve decision consistency under domain shift. Experimental results on two datasets demonstrate that HCKNet achieves a consistently more reliable and robust vibration measurement-to-decision mapping under source-free deployment.