Data-Model Jointly Driven Fault Diagnosis for Wind Turbine Planetary Gearboxes

• • The proposed method achieves superior diagnostic accuracy under fault sample scarcity compared to classical approaches, as validated on a wind turbine planetary gearbox fault diagnosis test rig, directly addressing the industrial bottleneck where fault data are scarce due to automatic shutdowns. • • A high-fidelity lumped-parameter dynamic model generates pseudo-fault data that supplements the training set, mitigating the imbalance between abundant healthy data and scarce fault data without relying on oversampling or undersampling techniques that risk overfitting or information loss. • • The integration of convolutional block attention modules and local maximum mean discrepancy aligns pseudo and real fault data distributions at the fault-category level, reducing domain shift and enhancing feature transferability, which is critical for deploying models across different operating conditions. • • The Kolmogorov-Arnold network module improves the model's ability to learn complex data relationships, enabling robust classification of different fault types; however, the framework currently addresses only known fault types, and unknown fault identification remains an open challenge requiring open-set domain generalization.