Bearing Fault Diagnosis Based on an Enhanced Image Representation Method of Vibration Signal and Conditional Super...

Area of Science:

• Mechanical Engineering
• Signal Processing
• Artificial Intelligence

Background:

• Multipoint Optimal Minimum Entropy Deconvolution Adjusted (MOMEDA) uses multipoint kurtosis (MKurt) for fault detection but is sensitive to noise.
• L-kurtosis offers better noise robustness compared to traditional kurtosis methods.

Purpose of the Study:

• To propose a novel bearing fault diagnosis framework.
• To enhance the robustness and accuracy of fault detection in vibration signals.
• To address the limitations of existing deconvolution methods in noisy environments.

Main Methods:

• Developed a multipoint envelope L-kurtosis (MELkurt) method for robust temporal feature extraction.
• Utilized Gramian Angular Difference Field (GADF) to convert MELkurt series into images for enhanced representation.
• Proposed a Conditional Super Token Transformer (CSTT) model, incorporating advanced deep learning components for feature learning.
• Applied transfer learning to improve diagnostic accuracy and generalization.

Main Results:

• MELkurt demonstrated superior noise robustness and fault feature enhancement compared to traditional kurtosis.
• The proposed CSTT model achieved the highest diagnostic accuracy and stability in experiments.
• The novel framework effectively learned and extracted features from GADF images for bearing fault diagnosis.

Conclusions:

• The MELkurt method significantly improves fault feature enhancement and noise robustness.
• The CSTT model, combined with GADF image representation, provides a highly accurate and stable bearing fault diagnosis framework.
• The proposed approach outperforms existing methods like Vision Transformer and CNN-based models.