Planetary gearbox fault classification based on tooth root strain and GAF pseudo images

Area of Science:

• Mechanical Engineering
• Condition Monitoring
• Machine Learning

Background:

• Traditional acceleration signal analysis for planetary gearbox fault detection is hindered by interference.
• Accurate fault identification in gearboxes is crucial for preventing catastrophic failures and ensuring operational reliability.

Purpose of the Study:

• To develop and validate a robust gear fault classification method for planetary gearboxes.
• To overcome the limitations of acceleration-based methods by utilizing strain data and advanced machine learning.

Main Methods:

• Direct acquisition of ring gear root strain data using fiber optic sensors.
• Preprocessing of strain signals via resampling and time-domain synchronous averaging.
• Encoding strain data into Gramian Angular Fields (GAF) images for analysis.
• Utilizing CN-EfficientNet with contrast learning for feature extraction and classification.

Main Results:

• Achieved a classification accuracy of 96.84% for various fault types in planetary gearboxes.
• Demonstrated superior performance compared to other common classification models.
• Grad-CAM visualization provided interpretability of the fault recognition network's decisions.

Conclusions:

• The proposed method effectively classifies faults in planetary gearboxes using root strain and pseudo images.
• The integration of strain data, GAF, and CN-EfficientNet offers a promising approach for enhanced gearbox condition monitoring.
• The study highlights the potential of deep learning with contrastive learning for complex machinery fault diagnosis.