A novel joint distinct subspace learning and dynamic distribution adaptation method for fault transfer diagnosis

• A joint distinct subspace learning and dynamic distribution adaptation framework. • Distinct subspace learning for minimizing geometrical shift. • Dynamic distribution adaptation for minimizing distribution divergence. • Manifold Laplacian regularization term for improving classifier performance. • Learn an adaptive transfer classifier for cross-domain fault diagnosis. Domain adaptation (DA) have achieved phased results in fault transfer diagnosis. However, there is still no unified framework that not only minimize geometrical shift but also minimize distribution shift of source and target data. Besides, existing DA approaches only exploit shared features between domains, which will degenerate when the two domains have a big shift, since a common subspace may not exist. In this paper, a novel joint distinct subspace learning and dynamic distribution adaptation (JDSDDA) method is proposed for rolling bearing fault transfer diagnosis. JDSDDA seeks two coupled projections to map source and target data into distinct subspaces, while adjusting the relative significance of marginal and conditional distributions by an adaptive weight factor. Finally, JDSDDA learns an adaptive classifier by optimizing structural risk minimization (SRM), distinct subspace alignment, dynamic distribution adaptation simultaneously. The experimental results demonstrates that the JDSDDA has significant superiority compared with other DA approaches.