Simulation-to-real transfer learning for bearing fault diagnosis across working conditions: A hybrid approach combining physical modeling and data-driven techniques

To address the challenge of limited fault samples in the source domain for intelligent fault diagnosis of rotating machinery, this study proposes a simulation-to-real transfer learning framework that integrates physics-based modeling with data-driven methods. Specifically, a dynamic bearing model is constructed under realistic boundary and loading conditions to simulate vibration signals corresponding to various health states. These signals are transformed into time–frequency representations using wavelet transform to form the source domain dataset, while real-world vibration data serve as the target domain. A Multi-scale Kolmogorov–Arnold Convolutional Network (MKANC) is developed to extract discriminative features across multiple semantic scales. To enhance feature representation, a Feature Enhancement module with Cross Attention (FECA) is introduced, which emphasizes fault-related patterns and facilitates information exchange across scales. Additionally, independent domain discriminators are deployed for each scale to construct a multi-scale collaborative adversarial transfer architecture, enabling effective domain adaptation from simulation to real-world scenarios. Experimental results demonstrate that the proposed framework achieves high diagnostic accuracy in the target domain, exhibiting strong generalization capability and significant practical applicability.