A Novel Multi-Source Domain Knowledge Joint Transfer Network for Aero-Engine Bearing Fault Diagnosis

In recent years, various deep learning models and transfer learning strategies have been applied to machine fault diagnosis, demonstrating efficient performance. However, the reliance on large amounts of labeled data has not been completely eliminated. In real industrial scenarios, it is impractical to comprehensively collect and label a sufficient number of training samples. Cross-domain fault diagnosis of bearings has practical significance. Addressing the issue of limited labeled data, a Multi-source Domain Knowledge Joint Transfer Network (MDKJTN) method is proposed for sharing knowledge across multiple domains. This approach is designed for fault diagnosis tasks in cross-domain scenarios where no samples are available in the target domain. To enhance the domain generalization performance of the network, a multi-source domain joint discrepancy loss function based on the maximum mean discrepancy is designed to reduce the marginal distributions and conditional distributions discrepancy between different domains. By learning domain-invariant features, the shared diagnostic knowledge is jointly transferred from both labeled and unlabeled data to the target domain. It is combined with supervised classification loss to update network parameters together, and ensuring excellent diagnostic performance when facing unseen working condition samples. Finally, an aero-engine inter-shaft bearing fault dataset was used to establish multiple sets of experiments in different cross-working condition scenarios. Through comparisons with other methods, the effectiveness of the proposed method in cross-domain fault diagnosis performance is validated.