Few-Shot Fault Diagnosis of Rolling Bearings Based on Category-Enhanced Siamese Attention Hierarchical-Embedded UNet Network

In engineering practice, not only is it difficult to obtain large amounts of labeled training data, but also strong noise interference degenerates the data quality, leading to the limited performance of deep learning-based fault diagnosis methods. Therefore, this paper proposes a category-enhanced Siamese attention hierarchical-embedded UNet network for diagnosing rolling bearing faults under the limited samples. Firstly, the channel attention is hierarchically embedded into the UNet model, thereby enhancing the model’s capacity to discern and retain key features at different levels. Secondly, a classifier is introduced into the Siamese network architecture, and a comprehensive loss function is designed to improve the model’s capacity to learn differentiated features by utilizing the category information of limited samples. Finally, in the testing stage, the outputs of distance metric and the classifier are combined to predict the category of the samples, thus increasing the fault classification accuracy. Two sets of experimental data are analyzed, and the comparison results verify the effectiveness and practicality of the investigated method in few-shot fault diagnosis of rolling bearings.