A balanced and weighted alignment network for partial transfer fault diagnosis.

Area of Science:

• Mechanical Engineering
• Artificial Intelligence
• Machine Learning

Background:

• Domain adaptation is crucial for mechanical fault diagnosis.
• Existing methods fail when target labels are a subset of source labels (partial transfer).
• This limitation hinders practical applications in mechanical fault diagnosis.

Purpose of the Study:

• To propose a novel network for partial transfer fault diagnosis.
• To address the challenge of differing label spaces between source and target domains.
• To improve the accuracy and robustness of mechanical fault diagnosis systems.

Main Methods:

• A balanced and weighted alignment network is proposed.
• Augmenting the target domain to balance class distributions.
• Shortening class-center distances to mitigate conditional distribution shifts.
• Weighted adversarial alignment to filter irrelevant source samples and minimize marginal distribution discrepancy.

Main Results:

• The proposed method effectively handles partial transfer scenarios in fault diagnosis.
• Demonstrated reduction in negative transfer and enhancement of positive transfer.
• Achieved promising performance on experimental test rigs.
• Outperformed existing state-of-the-art partial transfer methods.

Conclusions:

• The balanced and weighted alignment network is effective for partial transfer fault diagnosis.
• The method successfully overcomes the limitations of identical label space assumptions.
• This approach offers a significant advancement for real-world mechanical fault diagnosis.