Conformal residual score extrapolation diffusion for domain generalization fault diagnosis under unseen working conditions

Abstract For fault diagnosis of rotating machinery under unseen working conditions, existing generative domain generalization methods commonly suffer from the entanglement between fault semantics and domain shifts, the lack of explicit constraints on the direction and strength of extrapolation, and the difficulty of quantitatively screening the reliability of pseudo samples. To address these issues, this paper proposes a Conformal Residual Score Extrapolation Diffusion (CRSED) method for domain generalization fault diagnosis. The method explicitly decomposes the noise prediction of the conditional diffusion model into a fault-shared score and a domain-residual score, disentangling cross-domain stable fault semantics from source-domain-related distribution shifts; a residual score extrapolation consistency constraint is then introduced so that the domain-residual branch maintains smooth and predictable responses beyond the source-domain boundary; on this basis, a semantic-reliable adaptive boundary extrapolation strategy for pseudo-domain generation is proposed and combined with a class-conditional conformal semantic boundary criterion to adaptively determine the safe extrapolation radius and screen reliable pseudo unseen-domain samples, enlarging the coverage of the source-domain distribution while preserving the semantic reliability of the generated samples. Experiments on public bearing datasets and a bearing dataset collected on a self-built fluid system simulation platform show that the average diagnostic accuracy of CRSED is superior to that of the compared methods. The average diagnostic accuracies on the JNU, HUST, and Fluid System Simulation Platform (FSSP) datasets are 94.52%, 95.63%, and 94.02%, respectively.