Multi-level confusion augmentation and joint domain generalization for fault diagnosis under unseen conditions

Abstract Current domain generalization (DG) methods predominantly rely on single-level confusion strategies to enhance data diversity and extract domain-invariant features solely through marginal distribution alignment, limiting their ability to capture both global and local discriminative information. To address these limitations, this paper proposes multi-level confusion augmentation and joint DG(MCA-JDG) for fault diagnosis of rotating machinery under unseen working conditions. Specifically, we first construct a composite domain by randomly sampling from multiple real working conditions, and apply global Dir-Mixup and local Dir-FMixup to all original and composite domains to generate diverse augmented domains and enrich multi-scale feature representations. Furthermore, we introduce a joint distribution alignment scheme that simultaneously aligns marginal and conditional distributions, enhancing the domain invariance of learned features across scales. To mitigate negative transfer caused by low-quality samples during initial training phases, we design ConfMix, a confidence-based sample selection strategy that adaptively filters unreliable samples. Extensive experiments on three public rotating machinery fault diagnosis datasets demonstrate that MCA-JDG significantly outperforms state-of-the-art methods in diagnostic accuracy and generalization capability, achieving accuracies of 99.83%, 96.51%, and 92.43% on the JUST, Ottawa, and SDUST datasets, respectively.