A digital twin-driven cross-domain adaptation method for bearing intelligent fault diagnosis

Subject to frequent operation in extreme working conditions such as high loads and rotational speeds, bearings are prone to unpredictable failures. Deep learning has improved diagnostic capability, but its demand for large amounts of labelled data limits its application under small sample or zero-shot conditions. Digital twin technology can alleviate this issue by generating simulated signals through virtual models, but traditional cross-domain adaptation methods cannot be directly applied due to the distribution difference between simulated and measured signals. To address this issue, this paper would propose a digital twin-driven cross-domain adaptation method, aiming to solve the small sample problem and domain distribution difference. This method constructs a virtual model for bearing faults based on Hertz contact theory, calibrates simulated and measured signals using cosine similarity to generate high-fidelity labelled simulated signals, and combines adversarial domain adaptation technology with a kurtosis weighting strategy to effectively transfer diagnostic knowledge from simulated signals to real scenarios, reducing the distribution difference. Experimental results show that under small-sample conditions, the proposed method achieves an average diagnostic accuracy of 98.88% with relatively short training time, significantly outperforming other comparative methods. This highlights its viability for practical implementation in scenarios with limited data.