Multilabel Transfer Learning Method With Dynamic Multimetric for Coupling Fault Diagnosis

Area of Science:

• Industrial fault diagnosis
• Machine learning
• Transfer learning

Background:

• Increasing system complexity leads to prevalent multicomponent failures (coupling faults), resembling multilabel data.
• Varying industrial tasks create cross-domain challenges for coupling fault diagnosis, necessitating transfer learning.

Purpose of the Study:

• To develop an advanced multilabel transfer learning method for coupling fault diagnosis.
• To address limitations in existing methods regarding multilevel similarity and complex fault features.

Main Methods:

• Proposed a novel method for dual domain alignment at global and local feature levels.
• Utilized Maximum Mean Discrepancy (MMD) for global distribution alignment across network stages.
• Introduced a dynamic multimetric structure to capture diverse local similarities and align local spatial structures.

Main Results:

• Achieved high superiority in transfer tasks across public and laboratory datasets.
• Demonstrated effective alignment of global feature distributions and local spatial structures.
• Validated the method's effectiveness in handling coupling fault diagnosis under varying conditions.

Conclusions:

• The proposed method offers a significant advancement in multilabel transfer learning for industrial fault diagnosis.
• Dual-level domain alignment effectively addresses the complexities of coupling faults in cross-domain scenarios.
• The approach shows strong potential for real-world industrial applications requiring robust fault diagnosis.