Investigation of Feature Engineering Methods for Domain-Knowledge-Assisted Bearing Fault Diagnosis

Area of Science:

• Mechanical Engineering
• Data Science
• Machine Learning

Background:

• Rolling bearing condition monitoring is crucial for industrial machinery.
• Traditional methods often rely on vibration data and machine learning for fault diagnosis.
• Deep learning methods are gaining popularity for their data-driven approach, reducing the need for domain expertise.

Purpose of the Study:

• To investigate the effectiveness of traditional feature engineering methods combined with domain knowledge for rolling bearing fault diagnosis.
• To evaluate novel combinations of signal processing techniques and mathematical feature formulas.
• To compare the performance of feature engineering approaches against purely data-driven deep learning methods.

Main Methods:

• A comprehensive feature engineering study involving 42 mathematical feature formulas.
• Application of preprocessing methods including envelope analysis, empirical mode decomposition, wavelet transforms, and frequency band separations.
• Utilizing a random forest classifier for fault prediction and evaluation on the CWRU bearing fault dataset.

Main Results:

• Feature calculation using frequency band separation yielded particularly high prediction accuracies.
• The proposed feature engineering method demonstrated high efficiency with low computational effort.
• The approach provided excellent accuracies comparable to deep learning methods, with the added benefit of explainability.

Conclusions:

• Feature engineering combined with domain knowledge offers a robust and explainable alternative for rolling bearing fault diagnosis.
• Frequency band separation is a highly effective preprocessing technique for this application.
• The proposed method presents a competitive and interpretable solution compared to complex deep learning models.