A dual-enhanced generative framework for zero-shot bearing compound fault diagnosis

Rolling bearings are critical components in rotating machinery, and their faults may threaten system safety and stability. Although data-driven methods have achieved promising results, they usually require labeled samples from all target categories, which limits their application to unseen compound faults. To address the scarcity of labeled compound-fault samples, this paper proposes a dual-enhanced generative framework for zero-shot bearing compound fault diagnosis, which enhances the diagnosis from two aspects: semantic feature generation and global-local feature interaction. First, ensemble empirical mode decomposition is employed to construct physically meaningful fault semantics from single-fault vibration signals, and unseen compound fault semantics are inferred by fusing corresponding single-fault semantics. Then, a Regressor-Enhanced Generative Adversarial Network (R-GAN) synthesizes unseen compound-fault features under semantic guidance, with a Discriminative Regressor ensuring semantic consistency. Finally, a Convolutional Neural Network (CNN)–Transformer Feature Interaction Network (CT-FIN) extracts discriminative global and local fusion features. The proposed method is validated on the comprehensive compound fault dataset and the HDU dataset, demonstrating that the proposed method can effectively identify unseen compound faults and outperforms the compared zero-shot diagnosis methods.