An Efficient Federated Learning Framework for Machinery Fault Diagnosis With Improved Model Aggregation and Local...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Industrial IoT

Background:

• Limited high-quality labeled data from devices hinders fault diagnosis model generalization due to operating environment constraints and data privacy.
• Existing fault diagnosis models struggle with insufficient data, impacting their ability to perform accurately in real-world industrial scenarios.

Purpose of the Study:

• To propose a high-performance federated learning framework for robust machine fault diagnosis.
• To enhance both model aggregation and local model training procedures within a federated learning context.
• To address data scarcity and privacy concerns in industrial fault diagnosis.

Main Methods:

• Developed an optimized aggregation strategy combining forgetting Kalman filter (FKF) and cubic exponential smoothing (CES) for central server model aggregation.
• Proposed a deep learning network for multi-client local model training, incorporating multiscale convolution, attention mechanisms, and multistage residual connections.
• Implemented a federated learning framework to facilitate collaborative model training without compromising data privacy.

Main Results:

• The proposed federated learning framework achieved high accuracy in machinery fault diagnosis.
• The framework demonstrated strong generalization capabilities on two distinct machinery fault datasets.
• Experiments confirmed the framework's effectiveness in protecting data privacy during distributed training.

Conclusions:

• The novel federated learning framework effectively overcomes data limitations and privacy concerns in industrial fault diagnosis.
• The integration of FKF-CES aggregation and advanced deep learning for local training significantly boosts diagnostic performance.
• This approach offers a practical solution for developing accurate and generalizable fault diagnosis models in real-world industrial settings.