A Zero-Shot Framework for Compound Fault Diagnosis in Rotating Machinery Using Orthogonal High-Order Semantics and Cross-Hierarchical Discriminative Learning

Compound faults, characterized by inherent complexity and scarce labeled data, pose significant challenges to the generalization of conventional diagnostic methods for unseen fault modes. In contrast, zero-shot learning (ZSL) leverages seen-class data and auxiliary information to diagnose unseen compound faults without requiring target samples. However, existing zero-shot diagnostic models face two key challenges. First, insufficient fault semantic discriminability and nonorthogonal representations lead to semantic redundancy and reduced separability. Second, overlapping feature spaces in compound faults degrade semantic mappings, limiting their effectiveness in recognizing unseen faults. To address these challenges, this article presents a zero-shot framework for compound fault diagnosis, termed orthogonal high-order semantics and cross-hierarchical discriminative learning (OHSCDL). OHSCDL enhances semantic discriminability by integrating high-order reasoning within a dual-stream collaborative network. Additionally, an orthogonal constraint mechanism is introduced to minimize redundancy and preserve the compositional integrity of compound fault semantics. Furthermore, OHSCDL adopts a multilevel optimization strategy that simultaneously refines feature representations and classification decisions, effectively mitigating feature overlap. Experimental evaluations on three typical rotating machinery scenarios demonstrate the effectiveness and broad applicability of the proposed method in diagnosing faults across diverse mechanical components. Specifically, OHSCDL achieves unseen compound fault identification accuracies of 94.17%, 88.89%, and 94.91% on these datasets, outperforming state-of-the-art methods.