Transformer-based meta learning method for bearing fault identification under multiple small sample conditions

Most fault identification methods based on deep learning rely on a large amount of data, and their effects are limited in the actual production environment. In the case of multiple classification, limited data and complex working conditions, the bearing fault identification is very difficult. In this paper, a novel method called ensemble transformer meta-learning (ETML) is proposed for bearing fault identification with few samples. Based on the base-learner of vision transformer (ViT) and the meta-learner of model-agnostic meta-learning (MAML), a novel transformer meta-learning model is constructed to enhance the extracting feature ability. Finally, the weighted voting of multiple models determines the final result. By meta-training of limited data, the model can obtain excellent initial parameters and identify bearing faults quickly and accurately with fewer samples under varying working conditions. The effectiveness of ETML is verified by two cases of bearing fault identification with a few labeled fault samples.