In traditional data-driven fault diagnosis, acquiring training examples for all potential fault classes is highly challenging. Zero-shot learning (ZSL) methods based on generative adversarial networks (GANs) have shown promising results; however, these approaches suffer from training instability, mode collapse, and a lack of robustness. To overcome these limitations, this article proposes a novel dual-path conditional denoising diffusion probabilistic model with attribute consistency (DP-CDDPM-AC) for zero-shot fault diagnosis (ZSFD). Specifically, a dual-path diffusion mechanism is introduced, comprising a feature-based diffusion path and an attribute-based diffusion path. The feature-based diffusion path employs a conditional reverse denoising process to generate realistic and diverse samples for unseen fault classes. Simultaneously, the attribute-based diffusion path explicitly models the complex attribute distribution, providing robust and informative attribute representations. By leveraging intermediate noisy attribute representations from the attribute-based path, the feature-based path effectively enhances attribute consistency, robustness to noisy or ambiguous attributes, and mitigates the domain shift problem inherent in zero-shot scenarios. In addition, an attribute regressor is integrated to introduce an attribute-consistent loss, further ensuring generated features align well with their corresponding attributes. A clustered KL-guided filter and feature concatenation operation are adopted to select qualified generated features and synthesize more discriminative fault features, respectively. Extensive experiments on three fault diagnosis datasets demonstrate that the proposed DP-CDDPM-AC significantly outperforms other methods.
Liao et al. (Thu,) studied this question.