Learning to Imbalanced Open Set Generalize: A Meta-Learning Framework for Enhanced Mechanical Diagnosis.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Mechanical Engineering

Background:

• Domain generalization (DG) is used in mechanical diagnosis to handle varying data distributions.
• Open set DG (OSDG) addresses unknown fault states but suffers from class imbalance due to data collection challenges.
• Class imbalance in OSDG skews decision boundaries, leading to misclassification of unknown states and security risks.

Purpose of the Study:

• To develop a method that simultaneously tackles domain shift and class imbalance in unknown domains for OSDG.
• To improve the accuracy and security of mechanical diagnosis systems operating under diverse and imbalanced conditions.

Main Methods:

• Proposes a multisource domain-class gradient coordination meta-learning (MDGCML) framework.
• MDGCML coordinates gradients between interdomains and interclasses to learn generalized decision boundaries.
• Implements a joint learning paradigm with shared parameters between open-set and closed-set classifiers for rapid adaptation.

Main Results:

• The proposed MDGCML framework demonstrates superior performance in handling imbalanced OSDG.
• The method effectively learns generalized boundaries, mitigating the impact of domain and class shifts.
• Verified effectiveness on two benchmark datasets, showing significant improvements.

Conclusions:

• The MDGCML framework offers a robust solution for imbalanced OSDG in mechanical diagnosis.
• The joint learning approach enhances model adaptability to unknown domains and fault states.
• This work contributes to more reliable and secure AI applications in critical engineering domains.