Manifold Embedded Ensemble Partial Domain Adaptation: A Novel Partial-Set Transfer Mechanism for Gearboxes Fault Diagnosis

Domain adaptation (DA) models for fault diagnosis typically assume a shared label space between source and target domains. However, in real industrial applications, the target label space is often a subset of the source label space, potentially resulting in negative transfer due to uncertain propagation from source-only classes. To address this challenge, a novel manifold embedded ensemble partial DA (MEEPDA) approach is proposed. First, we proposed a novel discriminative manifold learning (DML) method based on density peak landmark selection (DPLS) to mitigate degenerated feature transformation in the PDA process. DPLS leverages the relative density metric to select density peaks as landmarks, thereby reducing the influence of interfering instances. Subsequently, DML learns a robust discriminative manifold mapping based on these landmarks to align the geometrical structures of two domains. Second, to promote positive transfer between shared categories and mitigate the risk of negative transfer from source-only classes, we propose a novel joint weighting mechanism (JWM) that incorporates entropy-based class-wise weighting and adaptive instance-wise weighting using the <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">${l}_{\text {2,1}}$ </tex-math></inline-formula>-norm structured sparsity regularizer. Then, MEEPDA combines the maximum mean and covariance discrepancy (MMCD) metric and the JWM to learn a feature adaptation matrix A by aligning the weighted marginal and class-conditional distributions. Finally, MEEPDA learns an integrated classifier by utilizing a majority voting strategy within a unified objective function, aiming to improve the prediction accuracy and generalization performance. The experimental results on two gearbox databases across 19 transfer tasks demonstrate that MEEPDA outperforms existing partial DA methods.