LAUSR

Methods for fault diagnosis based on metric learning, in which a query sample is classified by picking the closest prototype from the support set based on their feature similarities, have been the subject of many studies. In real-world applications of in-orbit products, such as circulating pumps, the computation of similarity between different pairs is prone to different degrees of inaccuracy, especially epistemic uncertainty. Knowing and considering the uncertainty of similarity may improve fault detection accuracy. This article provides a unique approach to fault diagnosis based on Prototypical Network (Pro-Net) and Uncertainty Theory. In particular, we use epistemic uncertainty by altering the representation of prototypes from a deterministic scalar to an uncertain representation. To assess the similarity between a query and the prototypes in a support set, we calculate the uncertain distance between the pairs using cross-entropy. Experiments with symmetrical structures reveal that our proposed method significantly enhances classification precision and achieves state-of-the-art performance. It improves the reliability of fault diagnosis and reduces the risk of making erroneous judgments in safety-critical systems, decreasing the possibility of adverse consequences.