A fault dynamics simulation-driven approach for partial domain fault diagnosis of roller bearings

Deployment of transfer learning methods for intelligent fault diagnosis has been limited by the scarcity of labeled data and label-space mismatch across domains. Here, a simulation-driven partial domain adaptation (SPDA) framework was presented that couples a validated fault-dynamics model (label-rich simulation source) with a physical test rig (unlabeled measurement target) for cylindrical roller bearings. Within SPDA, a scale-aware entropy-minimizing ambiguity domain adaptation network was developed that combines adversarial alignment with class- and instance-scale weighting and entropy minimization to mitigate negative transfer from source-only classes under partial label shift. Experiments on a NU216 benchmark covering 28 sim-to-real transfer tasks show consistent improvements over recent closed-set and partial domain adaptation baselines across diverse load–speed conditions. These results suggest that physics-informed simulation, paired with a domain adaptation algorithm, provides a practical route to reliable diagnosis when exhaustive labeled field data are unavailable.