Abstract: Aiming to address the challenge of insufficient fault data for planetary gearboxes in practicalengineering, which results in low diagnostic aeeuracy of traditional machine learning algorithms, a fauldiagnosis method based on the CAGNet (Conv Attention Gear Net) network is proposed. Firstly, vibration simulation signals under different fault conditions are generated using an established phenomenologieal model. By segmenting these simulation signals, a network training dataset is constructed to overcomethe issue of limited training samples for planetary gearboxes. Secondly, to enhance network training effi.cieney and convergence speed , two convolutional layers are utilized for elfective feature extraction and in.lormation transmission , aecelerating the training process. Additionally, the convolutional attention mech.anism is embedded in the Residual Network to capture deep features of the signal. By selecting these fea.tures, the CAGNet network model is developed, which improves the model's diagnostic accuracy. Final-ly, based on both simulation and experimental data, the proposed model is compared with the ResidualNetwork ( ResNet), Deep Residual Shrinkage Network ( DRSN), and MSACNN network. The diagnosticaccuracy of the proposed model is improved by 1.5% , 2.5% , and 4. 5%6 , respectively.

Key words: planetary gearbox, Residual Network : attention mechanism, fault diagnosis