A Federated Adversarial Fault Diagnosis Method Driven by Fault Information Discrepancy

Area of Science:

• Machine Learning
• Artificial Intelligence
• Industrial Fault Diagnosis

Background:

• Federated learning (FL) enables collaborative model training but is sensitive to local model quality.
• Poor local models can cause negative transfer, degrading global model performance.
• Existing methods struggle with varying fault data distributions across different working conditions.

Purpose of the Study:

• To propose a federated adversarial fault diagnosis method (FedAdv_ID) for optimal global model construction.
• To address challenges posed by multiple working conditions and fault information discrepancies.
• To enhance the applicability and robustness of federated fault diagnosis models.

Main Methods:

• Introduced a consistency evaluation metric to quantify local vs. global fault information discrepancy.
• Developed a federated adversarial training mechanism to minimize feature discrepancies.
• Implemented an optimal aggregation strategy with adaptive weights for reduced global feature discrepancy.

Main Results:

• FedAdv_ID achieved 93.09% fault diagnosis accuracy on benchmark and real-world motor-bearing datasets.
• The proposed method significantly outperformed traditional regularization-based FL methods by 17.89%.
• Demonstrated improved global model performance under various motor operating conditions.

Conclusions:

• FedAdv_ID effectively mitigates negative transfer and improves global model accuracy in federated fault diagnosis.
• The method enhances model applicability across diverse working conditions by addressing fault information discrepancies.
• Adversarial training and optimal aggregation are key to building robust federated fault diagnosis systems.