Abstract: To address the challenge of cross-domain fault diagnosis for ship thrust bearings caused by scarce labeled samples in the target domain under cross-equipment and cross-operational conditions, a fault diagnosis method based on Zero-Shot Guided Discriminative Adaptation (ZSGDA) is proposed. The framework initially extracts cross-domain task-irrelevant feature pairs as prior knowledge, which are jointly optimized with labeled source fault samples to build distribution-consistent feature subspaces for the target domain. Additionally, a Guided Discriminative and Correlation Subspace Learning (GDCSL) framework is introduced to plan the feature mapping path and optimize the distribution of the shared feature space for cross-domain data. Finally, a robust mapping from fault features to the semantic space is achieved under the condition of zero labels in the target domain. Experiments designed using bearing datasets verify that the proposed method achieves an average diagnostic accuracy of 99.3% in zero-shot scenarios and can significantly shorten the convergence cycle. This method realizes zero-shot transfer for ship thrust bearings, providing a high-precision and high-robustness solution for fault diagnosis of ship thrust bearings in zero-shot scenarios, with significant engineering application value.

Key words: fault diagnosis, subspace learning, zero-shot, ship shafting, domain adaptation