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

Abstract

The working conditions of planetary gearboxes are complex, and their structural couplings are strong, leading to low reliability. Traditional deep neural networks often struggle with feature learning in noisy environments, and their reliance on one-dimensional signals as input fails to capture the interrelationships between data points. To address these challenges, we proposed a fault diagnosis method for planetary gearboxes that integrates Markov transition fields (MTFs) and a residual attention mechanism. The MTF was employed to encode one-dimensional signals into feature maps, which were then fed into a residual networks (ResNet) architecture. To enhance the network’s ability to focus on important features, we embedded the squeeze-and-excitation (SE) channel attention mechanism into the ResNet34 network, creating a SE-ResNet model. This model was trained to effectively extract and classify features. The developed method was validated using a specific dataset and achieved an accuracy of about 98.1%. The results demonstrate the effectiveness and reliability of the developed method in diagnosing faults in planetary gearboxes under strong noise conditions.