Lifeisgood: Learning Invariant Features via In-Label Swapping for Generalizing Out-of-Distribution in Machine Fault...

Area of Science:

• Machine Learning
• Artificial Intelligence
• Mechanical Engineering

Background:

• Conventional data-driven models for machine fault diagnosis struggle with generalization due to domain shifts.
• Empirical Risk Minimization (ERM) focuses on label informativeness but neglects feature invariance, hindering performance across different operating conditions.

Purpose of the Study:

• To propose a new learning framework, Lifeisgood (learning invariant features via in-label swapping for generalizing out-of-distribution), to enhance generalization in machine fault diagnosis.
• To enable models to learn invariant features that maintain performance across domains with varying data distributions.

Main Methods:

• Introduced the Lifeisgood framework, inspired by assessing feature invariance through label-preserving feature swapping.
• Developed a theoretical guarantee for improved out-of-distribution performance using a novel swapping 0-1 loss.
• Derived a surrogate swapping cross-entropy loss to address training difficulties associated with the swapping 0-1 loss.

Main Results:

• Lifeisgood demonstrated superior performance compared to state-of-the-art methods in machine fault diagnosis.
• Achieved higher average accuracy and significantly increased the frequency of outperforming generic Empirical Risk Minimization (ERM).

Conclusions:

• The Lifeisgood framework offers a convenient and effective approach for developing robust data-driven fault diagnosis models.
• The proposed method successfully addresses the generalization limitations of conventional ERM by incorporating feature invariance.