Bearing Fault Diagnosis Based on Randomized Fisher Discriminant Analysis.

Area of Science:

• Mechanical Engineering
• Signal Processing
• Machine Learning

Background:

• Bearing faults are critical in rotating machinery, necessitating accurate diagnostic methods.
• Traditional Fisher Discriminant Analysis (FDA) struggles with nonlinear data.
• Kernel Fisher Discriminant Analysis (KFDA) addresses nonlinearity but incurs high computational costs.

Purpose of the Study:

• To propose a novel randomized Fisher discriminant analysis (RFDA) for bearing fault diagnosis.
• To address the non-linearity issue in vibration signals using random feature maps.
• To reduce the computational burden compared to KFDA while maintaining diagnostic accuracy.

Main Methods:

• Extraction of representative time-domain features from raw vibration signals.
• Extension of linear FDA to nonlinear RFDA using random feature maps.
• Application of Bayesian inference for classifying vibration signals to diagnose bearing status.

Main Results:

• RFDA effectively handles non-linearity in vibration data.
• Comparative experiments on CWRU and PU datasets demonstrate RFDA's superior performance.
• RFDA achieves comparable or higher accuracy than KFDA with significantly reduced computation time.

Conclusions:

• The proposed RFDA method offers an efficient and accurate approach for bearing fault diagnosis.
• RFDA provides a computationally advantageous alternative to KFDA for nonlinear problems.
• This method holds promise for real-world applications in condition monitoring of rotating machinery.