Search

Data-driven fault diagnosis techniques are frequently applied to ensure the reliability and safety of industrial systems. However, as a common challenge, the class-imbalance problem reduces the performance of data-driven methods due to the lack of data information. We propose a weighted modified conditional variational auto-encoder (WM-CVAE) as a novel data augmentation technique to tackle the issue. The modified structure can alleviate the existing Kullback–Leibler (KL) divergence vanishing by an adaptive loss. Meanwhile, kernel mean matching (KMM) is proposed on weight computation to reduce the negative effect of dissimilar generated samples. Constructing the WM-CVAE data augmentation framework can effectively improve the data quality and learning capability in class-imbalance fault diagnosis. To validate the proposed WM-CVAE model, three real-world industrial datasets are used as study objects, and the random forest is used as the base learner in the fault classification tasks. The diagnostic results demonstrate that the proposed WM-CVAE data augmentation framework can improve learning results in class-imbalance fault diagnosis. • A general adaptive loss-based learning algorithm to solve the KL divergence vanishing problem widely existing in the original CVAE generators. • An adaptive weight case is proposed as an example to be analyzed and validated on real-world class-imbalance fault diagnosis datasets. • Kernel mean matching is applied to realize weight assignment by metric learning. • Research results show the WM-CVAE data augmentation approach is excellent for class-imbalance fault diagnosis tasks. [ABSTRACT FROM AUTHOR]

Domain adaptation has been widely applied in data-driven fault diagnosis tasks to address the domain shift problem between source and target data. However, conventional domain adaptation methods require both domains to be known, which is not always feasible due to privacy concerns and big data transmission. To overcome this limitation, a dedicated method called source-free domain adaptation (SFDA) has been developed to ensure reliable performance without relying on source data during target model adaptation. SFDA can achieve accurate classification tasks under domain shift problems and source data-free scenarios. We propose a generalized source model with manifold Mixup data augmentation and label smoothing techniques to avoid overfitting during the source model training. Based on this model, a novel self-training framework is proposed to implement the domain adaptation task and achieve accurate prediction performance. The experimental results from three real-world datasets demonstrate the effectiveness of the proposed approach. • A novel source data-free domain adaptation approach to enable fault diagnosis in industrial systems. • A well-designed manifold Mixup module is internal to the model and trained jointly during source training. • A self-training strategy is applied with pseudo-labels and uncertainty reduction in target model adaptation. • Experimental results show the proposed method is excellent for fault diagnosis tasks under domain shift conditions. [ABSTRACT FROM AUTHOR]