A Two-Stage Interpretable Fault Diagnosis Approach for Bearings Based on EBM

Area of Science:

• Engineering
• Artificial Intelligence
• Machine Learning

Background:

• Explainable artificial intelligence (XAI) is crucial for bearing fault diagnosis.
• Existing methods like SHAP struggle with feature redundancy from extensive signal processing.
• Feature redundancy increases computational cost, overfitting risk, and biases interpretability.

Purpose of the Study:

• To develop a two-stage interpretable fault diagnosis approach.
• To address feature redundancy and improve diagnostic accuracy and transparency.
• To provide a trustworthy methodological reference for industrial fault diagnosis.

Main Methods:

• A two-stage approach using Explainable Boosting Machine (EBM) for core feature selection.
• Enhancing EBM with Random Forest (RF) via residual learning to create the RF-EBM model.
• Utilizing both EBM and Shapley Additive Explanations (SHAP) for dual interpretability analysis.

Main Results:

• The proposed RF-EBM model achieved good diagnostic performance on benchmark datasets.
• The approach outperformed traditional EBM in diagnostic accuracy.
• Feature selection effectively reduced redundancy, enhancing model efficiency.

Conclusions:

• The developed method reduces feature redundancy and enhances diagnostic performance.
• The approach improves the transparency of decision-making processes in fault diagnosis.
• This offers a valuable reference for trustworthy industrial fault diagnosis systems.