Adaptive Intermediate Class-Wise Distribution Alignment: A Universal Domain Adaptation and Generalization Method for...

Area of Science:

• Machine Learning
• Artificial Intelligence
• Mechanical Engineering

Background:

• Transfer learning methods for fault diagnosis often require task-specific parameter tuning.
• Existing domain adaptation mechanisms suffer from dynamic alignment goals, leading to instability and poor robustness.

Purpose of the Study:

• To propose a novel and simple transfer learning diagnosis method, the adaptive intermediate class-wise distribution alignment (AICDA) model.
• To overcome limitations of existing methods, including parameter tuning and dynamic alignment issues.

Main Methods:

• Developed the AICDA mechanism for adaptive intermediate distribution alignment across multiple domains.
• Introduced a dynamic intermediate alignment (DIA) adaptive layer for domain confusion without explicit distribution loss.
• Utilized AdaSoftmax loss to enhance classification performance and separability.

Main Results:

• The AICDA model simultaneously aligns global and class-wise distributions of source and target domains.
• The DIA layer achieves domain confusion adaptively, reducing reliance on distribution distance and regularization losses.
• Experimental results on wind turbine planetary gearboxes demonstrate superior diagnosis accuracy and generalization ability compared to state-of-the-art methods.

Conclusions:

• The AICDA model offers a robust and effective solution for multisource mixed fault transfer diagnosis.
• The proposed method exhibits strong generalization capabilities in both domain adaptation and domain generalization tasks.
• AICDA significantly improves fault diagnosis performance in complex machinery systems.