Despite the great achievements of the intelligent diagnosis methods of rotating machinery based on being data-driven, it still suffers from the problem of scarce labeled data. Therefore, this paper focuses on developing a data augmentation method of few-shot learning for fault diagnosis under small sample size conditions. Firstly, we developed the latent optimized stable generative adversarial network to adaptively augment the small sample size data without prior knowledge. Furthermore, penalty terms based on the distance metric for differences in distributions are adopted to constrain the optimization objective of the model. And self-attention and spectral normalization are applied in the model to stabilize the training process. Then, supervised classifier training is conducted based on the augmented sample set. Comparative analysis of the frequency spectrum verified the authenticity and reliability of the generated samples. Finally, the performance of the proposed method is validated with a comparative study on three cases of rolling bearing fault diagnosis experiments. The average accuracy can achieve 99.71%, 99.7%, and 96.27% in 10-shot sample fault diagnosis.

No abstract

No abstract

In practical engineering, the number of acquired fault samples from different categories might be in great difference due to the little probability of key equipment happening to malfunctioned. When training the imbalanced data, more methods focus on balancing the number of samples between different categories which may be time-consuming and easy to over-fit. To address this problem, we proposed Embedding-augmented Gaussian Prototype Network (EGPN) which applied a new training mechanism from the perspective of meta-learning. We only train the categories with large samples and the remaining categories only appeared in the testing process to calculate untrained prototypes. EGPN includes a feature embedding augmented module, weighted prototype module and metric module. Firstly, ordinary convolution and dilated convolution are mixed to capture different frequency bands simultaneously, and residual-attention module is added to highlight key features and suppress unimportant features. Different prototypes are calculated by weighting to the embedding vectors through Gaussian covariance matrix. Finally, the classification is taken according to the modified distance. The experiments in two datasets indicating that the proposed method can effectively recognize the untrained categories with only a few samples using as the prototypes, which can tackle the problem of identifying imbalanced fault data efficiently.

No abstract