Fault Detection Based on Multi-Dimensional KDE and Jensen-Shannon Divergence

Area of Science:

• Mechanical Engineering
• Signal Processing
• Machine Learning

Background:

• Traditional fault diagnosis systems struggle with weak signals, data coupling, and identifying unknown faults.
• Existing methods like T2 statistics and cross-entropy exhibit limitations in detection and identification performance.

Purpose of the Study:

• To propose an advanced fault diagnosis method enhancing detection performance.
• To address challenges of weak signals, data coupling, and unknown faults in industrial systems.

Main Methods:

• Developed a novel fault diagnosis approach integrating optimal bandwidth kernel density estimation (KDE) and Jensen-Shannon (JS) divergence.
• Formulated and proved the convergence of an algorithm for multidimensional KDE optimal bandwidth.
• Utilized JS divergence based on optimal KDE for effective fault detection.

Main Results:

• The proposed method demonstrated a 10% and 2% higher detection rate for known faults compared to T2 statistics and cross-entropy, respectively.
• For unknown faults, the new method achieved approximately 15% higher detection rate than cross-entropy, while T2 statistics showed ineffectiveness.
• Experimental validation using bearing data confirmed the method's superior performance.

Conclusions:

• The proposed fault diagnosis method effectively improves fault detection rates for both known and unknown fault types.
• Optimal bandwidth KDE and JS divergence offer a robust solution for complex fault diagnosis scenarios.
• This approach enhances machinery reliability and diagnostic accuracy.