Cross-Condition Fault Diagnosis of Planetary Gearboxes Driven by Data-Model Fusion Based on Improved Domain-Adversarial Transfer Learning

Data-driven methods have been extensively applied in the intelligent fault diagnosis of planetary gearboxes. Building reliable deep learning models typically requires a substantial amount of labeled data. However, in real-world industrial environments, obtaining a large volume of sample data under various working conditions and for different types of faults is often challenging. Mechanistic models can generate simulated data under various working conditions and for different types of faults through numerical simulations, thereby addressing the issue of insufficient labeled data. However, the significant distributional discrepancies between simulated and measured data may compromise the diagnostic performance of the models. To address the aforementioned issue, this paper proposes an improved domain-adversarial neural network (DANN) model that integrates mechanism and data fusion for cross-condition fault diagnosis of planetary gear systems. First, a translation-rotation dynamics model considering the impact of crack faults is developed to generate simulation data, compensating for the deficiency of experimental data. Furthermore, envelope preprocessing is applied to reduce noise interference in the real data, while a DANN model with an integrated subdomain discriminator is designed. By aligning the conditional distributions of simulation and experimental data, classification accuracy and robustness are enhanced. Finally, the method is applied to the operational condition transfer of the 2K-H planetary gearbox, validating its diagnostic performance under different conditions. Comparative studies with classical models show that, even with limited data, the proposed method is capable of fault diagnosis under different operating conditions and demonstrates superior diagnostic accuracy compared to other methods.