Zero-Shot Fault Diagnosis for Smart Process Manufacturing via Tensor Prototype Alignment.

Area of Science:

• Process Manufacturing
• Artificial Intelligence
• Machine Learning

Background:

• Identifying unseen faults is critical for digital transformation in process manufacturing.
• Existing methods struggle with unseen faults due to limited generalizability.
• Conventional zero-shot recognition performs poorly in dynamic manufacturing scenarios.

Purpose of the Study:

• To develop a tensor-based zero-shot fault diagnosis framework (MetaEvolver) for process manufacturing.
• To improve fault diagnosis accuracy and generalize to unseen fault domains.
• To address the limitations of current methods in dynamic and unseen fault scenarios.

Main Methods:

• Proposed the concept of the uncertain meta-domain and constructed sample prototypes guided by class-level attributes.
• Developed a tensor-based framework (MetaEvolver) for evolving dual prototype distributions.
• Employed a strategy to inject prototype distribution information from another modality to enhance alignment.
• Utilized devised loss functions for knowledge transfer and unseen domain generalization.

Main Results:

• MetaEvolver demonstrates superior performance in zero-shot fault diagnosis across five process manufacturing datasets.
• Achieved significant improvements in accuracy and generalizability for unseen fault identification.
• Validated the framework's potential on five zero-shot benchmarks.

Conclusions:

• MetaEvolver offers a robust solution for zero-shot fault diagnosis in smart process manufacturing.
• The framework effectively handles unseen faults and enhances domain generalizability.
• This tensor-based approach shows great superiority and potential for practical applications.