Abstract
In conventional data-driven fault diagnosis scenarios, it is often challenging to acquire sufficient training instances for every potential fault category. Because many target faults cannot be collected in advance, these gaps can constrain the effectiveness of existing diagnostic models. Zero-shot learning has gained prominence as a viable solution to this problem. However, it often encounters the issue of domain shift and insufficient learning of semantic space. In this study, a novel approach named the cycle-consistent generative adversarial network with semantic distance (CycleGAN-SD) is proposed for zero-shot fault diagnosis. By defining the semantic distance between different fault classes in attribute feature space, CycleGAN-SD facilitates the transformation of fault data from seen to unseen classes. To alleviate the issue that the generated features may deviate from real faults, an attribute regressor is introduced to construct the attribute-consistent loss. Besides, this method also introduces feature concatenation to build new training data and testing data. By merging the synthesized features, a more distinctive representation is achieved, thereby enhancing subsequent fault diagnosis. To validate the effectiveness of the proposed method, three fault diagnosis scenarios were investigated. Furthermore, the experimental results demonstrate that the proposed method surpasses other state-of-the-art zero-shot fault diagnosis techniques.
Funding Information
  • U.S. Department of Commerce (123456)