Planetary Gearboxes Fault Diagnosis Based on Markov Transition Fields and SE-ResNet

Area of Science:

• Mechanical Engineering
• Artificial Intelligence
• Signal Processing

Background:

• Planetary gearboxes face reliability issues due to complex working conditions and strong structural couplings.
• Traditional deep neural networks struggle with feature learning in noisy environments and one-dimensional data, limiting fault diagnosis.
• Existing methods often fail to capture interrelationships within data points, impacting diagnostic accuracy.

Purpose of the Study:

• To develop an advanced fault diagnosis method for planetary gearboxes.
• To overcome limitations of traditional deep learning approaches in noisy and complex environments.
• To enhance the reliability and accuracy of planetary gearbox fault detection.

Main Methods:

• Integration of Markov transition fields (MTFs) to encode 1D signals into feature maps.
• Utilizing a residual networks (ResNet) architecture for feature extraction.
• Embedding a squeeze-and-excitation (SE) channel attention mechanism into ResNet34 (SE-ResNet) for improved feature focus.

Main Results:

• The proposed SE-ResNet model effectively extracts and classifies features from planetary gearbox data.
• The method demonstrated high performance in diagnosing faults under strong noise conditions.
• Validation on a specific dataset yielded an impressive accuracy of approximately 98.1%.

Conclusions:

• The developed fault diagnosis method integrating MTFs and SE-ResNet is effective and reliable.
• This approach significantly improves the accuracy of planetary gearbox fault detection in challenging environments.
• The study highlights the potential of attention-based deep learning for robust condition monitoring.