Wasserstein Generative Adversarial Networks with Meta Learning for Fault Diagnosis of Few-shot Bearing

Abstract

In practical work situations, the bearing fault diagnosis is a small and imbalanced data challenge. However, the intelligent fault diagnosis model relies on a mass of label data. This research, presents a different method, Wasserstein GAN with Meta Learning, for overcoming the difficulty of few-shot fault diagnosis under imbalanced data constraints. The WGAN module can generate synthetic samples for the data argument, and the first-order model agnostic meta-learning (FOMAML) to initialize and modify the network parameters. Validation of the comparative performance has been made using a benchmark dataset, i.e. CWRU datasets, which show that can achieve excellent diagnostic accuracy with small data. It's successfully overcome that the imbalanced data lead to the sample distribution bias and over-fitting. In addition, it can leverage that can precisely identify the bearing fault health types in a variety of working environments, even with noise interference. It is also found that the proposed model performs better in the testing set after training difficult datasets.