A contrastive generative network with feature-attribute consistency for zero-shot fault diagnosis in process industries

Abstract

In fault diagnosis tasks for process industries, comprehensively identifying all potential fault types poses significant challenges. Therefore, zero-shot fault diagnosis has gradually become a research hotspot. Currently, existing zero-shot fault diagnosis methods commonly face domain shift issues, which limit diagnostic performance. To address this shift, this paper proposes a feature-attribute consistency contrastive generative network (FAC-CGNet). This method combines attribute supervision with a contrastive learning mechanism to simultaneously maintain attribute consistency and decouple the feature space during feature generation. FAC-CGNet constructs an attribute-guided feature generation framework that integrates attribute information into the feature transformation process, ensuring that the generated features in the feature space remain consistent with their corresponding attributes. Furthermore, to prevent excessive overlap of generated features with similar attributes in the feature space, the paper designs a contrastive decoupling module. This module optimizes the feature space distribution through feature separation constraints and further enhances feature representation discrimination by combining a feature concatenation strategy. Finally, experiments on the public TEP dataset show that FAC-CGNet achieves an average accuracy of 83.1% in unknown fault diagnosis and significantly optimizes the feature representations in the feature space, confirming the effectiveness and superiority of the proposed method.