Dual-Path Conditional Diffusion Model With Attribute Consistency for Zero-Shot Fault Diagnosis

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Traditional data-driven fault diagnosis struggles with acquiring sufficient training data for all fault types.
• Existing zero-shot learning (ZSL) methods, often based on generative adversarial networks (GANs), face challenges like training instability and lack of robustness.

Purpose of the Study:

• To propose a novel dual-path conditional denoising diffusion probabilistic model with attribute consistency (DP-CDDPM-AC) for robust zero-shot fault diagnosis (ZSFD).
• To enhance the generation of realistic and diverse samples for unseen fault classes, addressing limitations of current ZSL approaches.

Main Methods:

• Introduced a dual-path diffusion mechanism: a feature-based path for sample generation and an attribute-based path for robust attribute representation.
• Leveraged intermediate attribute representations to improve attribute consistency and mitigate domain shift issues in ZSFD.
• Integrated an attribute regressor with an attribute-consistent loss and employed a clustered KL-guided filter with feature concatenation for enhanced feature synthesis.

Main Results:

• The proposed DP-CDDPM-AC model demonstrated superior performance in zero-shot fault diagnosis across three benchmark datasets.
• The dual-path diffusion approach effectively generated diverse and realistic samples for unseen fault classes.
• Attribute consistency mechanisms significantly improved robustness against noisy attributes and domain shift.

Conclusions:

• The DP-CDDPM-AC model offers a significant advancement in zero-shot fault diagnosis, overcoming limitations of previous GAN-based methods.
• The attribute consistency enhancement is crucial for robust and accurate diagnosis in scenarios with limited or no prior data for certain faults.
• This approach provides a more reliable and effective solution for complex industrial fault diagnosis applications.