A Multiscale Recursive Attention Gate Federation Method for Multiple Working Conditions Fault Diagnosis

Area of Science:

• Machine Learning
• Artificial Intelligence
• Engineering

Background:

• Federated learning (FL) enables multi-client data aggregation for bearing fault diagnosis when individual clients lack sufficient samples.
• Disparities in client working conditions (e.g., lifecycle stage, load) can lead to valuable data being overlooked in traditional FL.
• Unequal client data distributions pose challenges for building robust fault diagnosis models.

Purpose of the Study:

• To develop a multiscale recursive federated learning (FL) framework to enhance bearing fault diagnosis.
• To enable the FL server to prioritize useful information from diverse clients.
• To improve the reliability of multi-condition fault diagnosis models in FL settings.

Main Methods:

• Implementation of a multiscale recursive federated learning (FL) framework.
• Local multiscale feature fusion to maximize server-side information utilization.
• Focusing on client data relevance within the FL process.

Main Results:

• The proposed FL method demonstrated improved accuracy in multi-condition bearing fault diagnosis.
• Achieved a 23.21% accuracy improvement over existing FL fault diagnosis methods.
• Validated effectiveness using the Case Western Reserve University benchmark dataset, even with limited local data and complex fault types.

Conclusions:

• The multiscale recursive FL framework effectively addresses challenges posed by heterogeneous client data in bearing fault diagnosis.
• Prioritizing useful client information and leveraging local feature fusion leads to more reliable diagnostic models.
• This approach significantly enhances fault diagnosis accuracy, offering a promising solution for real-world industrial applications.