Cross-condition fault diagnosis under heterogeneous label spaces via consistent discriminative feature learning and reconstruction

In practical cross-condition fault diagnosis, the fault data available under different operating conditions often exhibit incomplete and heterogeneous class compositions because of rare fault occurrences and data-acquisition constraints. Existing domain generalization methods commonly reduce inter-domain discrepancies through global feature alignment. However, when the source domains contain different fault classes, discrepancies in class composition become entangled with operating-condition shifts. Global alignment may therefore incorrectly suppress class-discriminative information. Moreover, incomplete source-domain label spaces make it difficult to establish reliable decision boundaries for the global fault label space. To overcome these limitations, a cross-domain consistent and intra-domain discriminative feature learning framework with missing-feature reconstruction (CDID-MFR) is developed in this study. A domain–class feature grid is constructed through confidence-weighted aggregation to explicitly characterize observed and missing domain–class combinations. Collaborative attention along the domain and class dimensions jointly captures cross-domain same-class consistency and intra-domain inter-class relationships. A correlation-guided projection is introduced to suppress confusion-related feature components and enhance class separability. Missing domain–class features are reconstructed by conditionally modulating class-specific fault-feature bases with source-domain operating-condition representations, thereby complementing the incomplete discriminative structure. Experiments on the Shanghai Maritime University motor data and the Paderborn University bearing benchmark confirm that CDID-MFR achieves superior diagnostic accuracy and transfers reliably to previously unobserved operating regimes. Its effectiveness is maintained across varying degrees of label-space heterogeneity, with particularly pronounced advantages under severe class missingness.