Research on Quantitative Diagnosis of Helical Gear Pitting Fault Severity Based on CycleGAN and Improved DANN

To tackle the issues of poor diagnostic accuracy and even diagnosis failure caused by insufficient samples and significant discrepancies between sample distributions for variable working conditions in gear pitting fault quantification diagnosis, a two-stage transfer diagnosis method driven by a dynamic model is proposed. First, a dynamic model of helical gears with varying severities of pitting faults is established based on the actual helical gear system structure. The diverse and sufficient simulated data are obtained by solving differential equations and used as the source domain. In the first stage, CycleGAN is applied to reduce the feature discrepancy between the simulated and experimental data, and thereby, the distribution of the simulated data is effectively adjusted. In the second stage, a domain alignment strategy is developed using a Beta-distribution-enhanced DANN method to align features between source domain simulation data and target domain experimental data. A transfer diagnosis model for pitting fault severity across multiple operating conditions is built. The results from nine cross-condition transfer diagnosis tasks demonstrate that varying severities of pitting faults can be accurately identified by the proposed method, achieving an accuracy of up to 93.18%. The method significantly outperforms comparative methods, effectively addresses the issue of insufficient samples, and improves fault diagnosis accuracy.