Dynamic Semi-Supervised Federated Learning Fault Diagnosis Method Based on an Attention Mechanism

Area of Science:

• Machine Learning
• Artificial Intelligence
• Industrial Fault Diagnosis

Background:

• Unsupervised learning struggles with accurate fault diagnosis on completely unlabeled data.
• Existing semi-supervised federated learning methods can suffer from negative transfer due to unreliable information from unlabeled clients.

Purpose of the Study:

• To propose a dynamic semi-supervised federated learning fault diagnosis method (SSFL-ATT) that prevents negative transfer.
• To enhance fault classification capabilities for unlabeled clients within a federated learning framework.

Main Methods:

• Developed a dynamic semi-supervised federated learning approach incorporating an attention mechanism (SSFL-ATT).
• Implemented a novel federation strategy driven by an attention mechanism to filter unreliable local model information.

Main Results:

• SSFL-ATT effectively prevents negative transfer in the federation model.
• Achieved significant accuracy improvements: 9.06% (Case Western Reserve University dataset) and 12.53% (Shanghai Maritime University dataset) compared to existing methods.

Conclusions:

• The proposed SSFL-ATT method enhances fault diagnosis accuracy and reliability in semi-supervised federated learning scenarios.
• This approach enables effective fault classification even with unlabeled client data, outperforming current methods.