Zero/few-shot fault diagnosis of rotary mechanism in rotational inertial navigation system based on digital twin and transfer learning

With the increasing demand for long-endurance, high-precision inertial navigation systems, rotational inertial navigation system (RINS) have become a research focus. However, the integration of rotary machinery introduces new challenges, including increased susceptibility to component failures, difficulties in collecting sufficient fault samples particularly for early stage faults and high costs and risks associated with fault injection testing. To address these challenges, this paper proposed a zero-shot fault diagnosis method for RINS based on digital-twin-assisted fault sample generation. By constructing a high-fidelity digital twin model, synthetic fault data are generated to compensate for the scarcity of actual fault samples. Furthermore, by integrating few-shot transfer learning with a small amount of real fault data, the diagnostic performance is further enhanced. Experimental results show that the proposed method achieves a fault diagnosis accuracy of 83.92% with binary classification accuracy reaching 96.86%Ẇhen few-shot transfer learning is applied, the classification accuracy exceeds 99% demonstrating the method's effectiveness in overcoming the key challenges of RINS fault diagnosis.