An uncertainty-aware prototype learning framework with structural constraints for open-world semi-supervised fault...

Area of Science:

• Industrial process monitoring
• Machine learning for fault detection
• Semi-supervised learning

Background:

• Real-world industrial fault diagnosis struggles with unknown fault types and scarce labeled data.
• Existing methods often face prototype collapse and unreliable clustering, hindering performance.
• Addressing these limitations is crucial for robust industrial monitoring systems.

Purpose of the Study:

• To propose an uncertainty-aware prototype learning framework with structural constraints for open-world semi-supervised fault diagnosis.
• To enhance the accuracy and adaptability of fault diagnosis systems in complex industrial environments.
• To overcome the challenges of prototype collapse and unreliable clustering in limited data scenarios.

Main Methods:

• Developed an uncertainty-aware prototype learning framework (OpenUPS) incorporating structural constraints.
• Utilized prototypes based on simplex equiangular tight frames for uniform distribution and maximal separation of class centers.
• Implemented an uncertainty-aware contrastive strategy for adaptive selection of informative unlabeled data pairs.

Main Results:

• OpenUPS effectively prevents prototype collapse, even with limited labeled data.
• The uncertainty-aware contrastive strategy ensures robust alignment of known classes and progressive clustering of novel faults.
• Experimental validation on the Tennessee Eastman process and a polyester esterification process confirmed superior performance over existing methods.

Conclusions:

• OpenUPS demonstrates strong generalization and adaptability for open-world industrial fault diagnosis.
• The proposed framework offers a robust solution for handling unknown fault types and limited data.
• This approach significantly advances the capabilities of semi-supervised learning in industrial settings.