Data-driven fault diagnosis framework of taper roller bearings using statistically ranked feature sets and machine...

Area of Science:

• Mechanical Engineering
• Condition Monitoring
• Machine Learning Applications

Background:

• Taper roller bearings are critical components in industrial machinery.
• Accurate and efficient fault diagnosis is essential for operational reliability.
• Current methods may lack the necessary precision or speed for real-time applications.

Purpose of the Study:

• To develop a novel, dependable, and computationally efficient fault diagnosis method for taper roller bearings.
• To enhance the automation of condition monitoring in industrial plants.

Main Methods:

• Collected vibration signals from an SKF 32,206 taper roller bearing under various conditions (healthy, inner race fault, outer race fault, roller fault, cage fault).
• Derived time-domain and frequency-domain features, ranking them using one-way ANOVA and Kruskal-Wallis tests.
• Evaluated six machine learning classifiers (SVM, neural networks, discriminant analysis, naive bayes, decision trees, nearest neighbor) with optimized feature sets.

Main Results:

• The Kruskal-Wallis statistical test proved effective for feature selection, improving both accuracy and computational efficiency.
• An optimal feature set comprising 18 features was identified.
• The linear Support Vector Machine (SVM) classifier achieved 99% classification accuracy and an Area Under the Curve (AUC) of 1.
• The SVM classifier demonstrated the shortest training time, indicating suitability for real-time fault detection.

Conclusions:

• The proposed vibration-based method offers a reliable and computationally efficient solution for taper roller bearing fault diagnosis.
• This approach significantly contributes to the automation of condition monitoring systems in industrial settings.
• The high accuracy and efficiency make it suitable for real-time industrial applications.