A Novel k-Nearest Neighbor Method with Distribution Discrepancy and Differential Feature Importance for Rolling...

Area of Science:

• Mechanical Engineering
• Signal Processing
• Machine Learning

Background:

• Rolling bearings are critical in rotary machinery, and their failures can be costly.
• Traditional k-nearest neighbor (KNN) methods struggle with non-stationary vibration signals and varying feature importance in fault diagnosis.
• Accurate fault detection and localization are essential for preventing machinery downtime.

Purpose of the Study:

• To develop an improved KNN-based fault diagnosis method for rolling bearings.
• To address limitations of traditional KNN by incorporating distribution discrepancy and differential feature importance.
• To enhance the accuracy and effectiveness of rolling bearing fault diagnosis.

Main Methods:

• Vibration signals decomposed using three-level wavelet packet decomposition.
• Fault features extracted based on energy distribution at the third level.
• Mean Impact Value (MIV) algorithm used for feature importance weighting.
• Earth Mover's Distance (EMD) applied to measure spatial distribution differences.
• Integrated KNN with Euclidean distance, MIV, and EMD for diagnosis.

Main Results:

• The proposed method achieved a diagnostic accuracy of 99.43%.
• This represents a 5.97% improvement over traditional KNN methods.
• Demonstrated accurate and effective fault diagnosis on rolling bearing datasets.

Conclusions:

• The novel KNN approach effectively diagnoses rolling bearing faults.
• Incorporating distribution discrepancy and differential feature importance significantly improves diagnostic accuracy.
• The method offers a robust solution for machinery health monitoring.