Joint attention feature transfer network for gearbox fault diagnosis with imbalanced data

Fault diagnosis methods based on deep learning have achieved remarkable success in the field of mechanical fault diagnosis. However, most data obtained in the industrial field come from imbalanced datasets, and traditional deep learning methods based on data balancing often achieve disappointing results, which limits their applications in different engineering scenarios. To solve this problem, a novel model called joint attention feature transfer network (JAFTN) is proposed. First, data are mapped to a feature space through a feature extractor that integrates a joint attention module, so as to keep the discrimination between different classes of features. After that, the constructed feature transfer module maintains the relation between different classes and transfers the common generalization representation obtained from other classes to the class with scarce samples, thereby enriching its feature space. Two gearbox fault diagnosis experiments from a test rig and an actual wind farm verify that the proposed JAFTN has higher fault diagnosis accuracy compared to several other popular imbalanced fault diagnosis methods. Meanwhile, the ablation study shows the effectiveness of each module.

Automated summary

In this article, a novel model JAFTN is proposed, which can achieve higher fault diagnosis accuracy in imbalanced datasets obtained from industrial field. This is achieved through the combination of feature extractor and feature transfer module.