Pyramid-type zero-shot learning model with multi-granularity hierarchical attributes for industrial fault diagnosis

For zero-shot fault diagnosis, we need to use seen faults to diagnose the class of faults that have never been seen before. Zero-shot learning (ZSL) transfers the knowledge of seen categories to unseen categories by constructing the attribute space, which can solve this problem. However, traditional ZSL methods fail to explore the hierarchical characteristics among attributes. In this paper, a pyramid-type ZSL (PZSL) model with multi-granularity hierarchical attributes is proposed to handle the above-mentioned problem based on the following recognitions: (1) the granularity of attribute information is different, and (2) coarse-grained attribute information can provide guidance for the prediction of fine-grained attributes. For the first time, the concept of information granularity in attributes is proposed, which can reveal the correlation of different faults at multiple levels. A hierarchical constrained network (HCNet) is designed to predict the attributes layer by layer. In addition, an attribute feature-guided (AFG) module is developed, which can integrate coarse-grained attribute information into fine-grained attribute recognition and transfer knowledge from easy-to-recognize attributes to hard-to-recognize attributes. Finally, a multi-layer fusion inference strategy is proposed, which can blend multi-granularity information of attributes. Results of experimental verification in thermal power plant processes prove the effectiveness of PZSL.