An Interactive Zero-Shot Compound Fault Diagnosis Framework Based on Fuzzy Semantics and Fuzzy Neural Networks

Abstract

In recent years, zero-shot compound fault diagnosis (ZSCFD) methods have been widely developed. However, existing methods still suffer from limitations, such as the rigidity of overly discriminative semantics and deterministic mapping mechanisms when facing the nonlinearity and uncertainty of compound faults, resulting in poor domain generalization. To address these limitations, an interactive ZSCFD framework based on fuzzy semantics and fuzzy neural networks (FNNs) is proposed. First, a convolutional neural network (CNN)-based feature extraction module introduces the center loss and the decouple loss to address module reusability. Second, in addition to discriminative semantics, a fuzzy semantic construction is proposed, forming a two-level semantics based on statistical semantics for discriminability and domain generalization tradeoff. Third, an FNN is introduced into the semantic embedding module to flexibly map fault features into the semantic space, alleviating the rigidity of deterministic mappings and enhancing domain generalization. Finally, an interactive extension module is proposed to generate interpretable engineering prompts. Comprehensive experimental results on two datasets demonstrate the effectiveness and domain generalization of this method for compound fault diagnosis (CFD).