Generalized zero-shot bearing compound fault diagnosis based on reserved gate pre-classifier and physical information semantics

The diagnosis of bearing compound faults, particularly zero-shot compound fault diagnosis, has become a significant focus in recent research. While semantic learning methods have introduced novel approaches for addressing zero-shot diagnosis tasks, their diagnostic accuracy can be compromised by the mutual interference between single and compound faults in practical scenarios. To overcome these challenges, this article presents a generalized zero-shot compound fault diagnosis method that integrates a reserved gate classifier and physical information-constrained semantics. By effectively distinguishing single faults from compound faults and implementing tailored diagnostic strategies, this method achieves robust and accurate fault diagnosis under generalized zero-sample conditions. First, a reserved gate pre-classifier (RGPC) and a global-local feature extractor (GLFE) are trained using interpolation strategies to classify both single and compound faults. Physical information from single faults is leveraged to construct a physical information-constrained autoencoder (PICAE), which generates high-fidelity semantic representations. Finally, samples identified as compound faults by the reserved gate outputs are mapped into the semantic space using global-local fusion features. The compound fault types are then determined by comparing the mapped features with predefined compound fault semantics. The proposed method was validated on a bearing compound fault dataset, achieving diagnostic accuracies of 86.71% for three compound fault types and 78.63% for four compound fault types.