Semi-supervised domain generalization for fault diagnosis using adaptive pseudo-label selection and distributionally robust optimization

Deep learning's potential for intelligent fault diagnosis (IFD) is constrained by inadequate model generalization to unseen working conditions, which motivates research in domain generalization-based fault diagnosis (DGFD). However, most DGFD methods rely on multiple labeled source domains during the training phase, conflicting with the prevalent scarcity of labeled industrial data. To bridge this gap, this paper introduces a novel semi-supervised domain generalization-based fault diagnosis (SemiDGFD) method using adaptive pseudo-label selection and distributionally robust optimization (Ada-DRO). The Ada-DRO only requires one labeled source domain and multiple unlabeled source domains. Furthermore, this method combines the primary branch model with auxiliary branch models. Firstly, it employs multiple auxiliary branch models to assign pseudo-labels to unlabeled data through distribution alignment. Subsequently, an adaptive threshold selects high-confidence pseudo-labels to mitigate noise. Finally, an uncertainty set is constructed using mask-augmented labeled source data and selected pseudo-labeled data. Wasserstein distance constrains the set scope, and distributionally robust optimization (DRO) is performed over this set to enhance the primary branch model's cross-domain generalization performance. Extensive experiments demonstrate the proposed method's superior accuracy over state-of-the-art SemiDGFD methods.