Multifeature Representation Distribution Adaptation Based on Data Fusion for Intelligent Fault Diagnosis of Planetary Gearboxes

Planetary gearboxes are widely used in many sorts of machinery, and data-driven intelligent fault diagnosis is one of the important development directions for planetary gearbox maintenance; however, the existing data-driven methods suffer from limitations such as weak fault information, insufficient training data, and uneven feature distribution in common industrial diagnosis scenarios. Although transfer learning-based approaches are anticipated to resolve the above issues, it remains challenging for the fault diagnosis of gearboxes with small datasets; thus, we propose a multifeature representation distribution adaptation (MFRDA) method for the cross-condition fault diagnosis of planetary gearboxes under time-varying conditions. We dimensionally reorganize the time-frequency feature matrices of multiple sensor signals under the same condition during signal preprocessing, thus extracting the feature information of multiple signals simultaneously for data fusion. The issue of weak fault feature loss within single signals is addressed by data enhancement of the fault-related features. We adopt a multifeature representation hybrid structure to extract multiple fault-related feature representations for feature learning, thus alleviating the issue of network overfitting caused by insufficient training data; moreover, a distribution assessment algorithm is presented in the hybrid structure to assess the feature distribution between source and target tasks, which minimizes the feature differences among different tasks by the distribution adaptation corresponding to different feature representations. This method is validated experimentally. Analysis results indicate that the diagnostic accuracy and training robustness of this method outperform the residual network and other adaptation methods.