Multi-Source Diagnosis of Bearing Faults Using Interpretable Boosted Trees

Area of Science:

• Electrical Engineering
• Artificial Intelligence
• Industrial Maintenance

Background:

• Induction motors are critical in industrial applications, and early fault detection prevents significant financial losses.
• Bearing faults are the most common issue in induction motors, necessitating reliable diagnostic techniques.
• Traditional diagnosis relies on vibration analysis, but supply current and sound are also explored, in both time and frequency domains.

Purpose of the Study:

• To employ explainable artificial intelligence (XAI) for identifying optimal sensor data and domains for induction motor fault diagnosis.
• To quantify the diagnostic improvements achievable through multisensor data fusion.
• To objectively compare different diagnostic approaches using a rigorous model selection process.

Main Methods:

• Utilized explainable artificial intelligence (XAI) techniques, including SHAP values, for interpreting diagnostic models.
• Applied boosting techniques for precise fault diagnosis.
• Implemented a structured model selection procedure for objective comparison of diagnostic variables and domains.

Main Results:

• Identified specific sensor variables and domains that significantly contribute to accurate induction motor fault diagnosis.
• Demonstrated the diagnostic benefits of integrating data from multiple sensors (multisensor data fusion).
• Achieved highly precise diagnoses through the application of boosting techniques and XAI interpretation.

Conclusions:

• Explainable AI provides valuable insights into the most effective data sources and domains for induction motor fault diagnosis.
• Multisensor data fusion enhances diagnostic accuracy, leading to improved industrial reliability.
• SHAP value interpretation of boosting models offers a transparent and effective approach to understanding complex diagnostic decisions.