Extended Invariant Risk Minimization for Machine Fault Diagnosis With Label Noise and Data Shift.

Area of Science:

• Machine Learning
• Data Science
• Engineering

Background:

• Supervised models in machine fault diagnosis struggle with incorrect labels and domain shifts.
• This challenge is known as the noisy label-domain generalization (NL-DG) problem, hindering model effectiveness.

Purpose of the Study:

• To develop a novel method, Extended Invariant Risk Minimization (EIRM), to address the NL-DG problem.
• To improve the robustness and generalization capabilities of machine fault diagnosis models.

Main Methods:

• EIRM incorporates flat minima seeking by shifting the gradient penalty base to the entire model.
• Theoretical analysis explores function smoothness and algorithm convergence for EIRM.
• An efficient implementation of EIRM is developed for fault diagnosis model construction.

Main Results:

• EIRM demonstrates a strong relationship with locating flat minima, crucial for label noise robustness and generalization.
• Comparative studies on actuator and gearbox fault datasets show EIRM outperforms existing benchmarks.
• The EIRM-based method proves more effective on average across multiple NL-DG tasks.

Conclusions:

• EIRM offers a robust solution for machine fault diagnosis under noisy label and domain generalization challenges.
• The method enhances model performance by improving generalization and robustness to label noise.
• The developed EIRM approach provides a significant advancement in data-driven fault diagnosis applications.