Statistical Manifold Generation-Driven Equipment Collaborative Personalized Fault Diagnosis

Area of Science:

• Industrial Internet of Things (IIoT)
• Machine Learning
• Data Privacy

Background:

• Existing intelligent fault diagnosis models struggle with generalization in IIoT due to data privacy constraints and distribution discrepancies.
• Time-varying working conditions in IIoT lead to performance degradation in fault diagnosis models.

Purpose of the Study:

• To propose a federated generalization fault diagnosis method driven by statistical manifold generation for IIoT.
• To enhance the robust generalization performance of fault diagnosis models across diverse and unknown working conditions.

Main Methods:

• A closed-loop collaborative strategy involving local extraction and cloud-based generation of statistical information (SI).
• Utilizing a global Gaussian mixture model with covariance expansion for multi-source distribution fitting and virtual SI generation.
• Employing a difference-aware mechanism and instance normalization for domain-invariant augmented training.

Main Results:

• Achieved over 85% average diagnostic accuracy on rolling bearing and gearbox datasets for completely unknown working conditions.
• Demonstrated diagnostic accuracy exceeding 90% in specific tasks.
• Maintained a minimal communication payload of 1.25 KB per round.

Conclusions:

• The proposed method significantly enhances cross-domain generalization capability of local models in IIoT.
• It strictly preserves data privacy with minimal communication overhead.
• Offers an efficient and robust collaborative intelligent diagnosis solution for resource-constrained IIoT edge devices.