Residual Prototype Network-based Method for Few-Shot Fault Diagnosis of Bearing

In engineering scenarios, it is often hard to gather enough fault data, making few samples for fault diagnosis a worthwhile hot area. Fortunately, through advanced meta-learning theory, there is the potential to decrease the need for model training data. Inspired by this, a residual prototype network (RPN) is proposed for few-shot diagnosis. By adopting the concept of meta-learning, the meta-learning task is used to select random samples from the training dataset to build the support set and query set. The model parameters are further updated by the query set. Therefore, an effective fault diagnosis model with few samples is developed. The behavior of RPN was demonstrated on a bearing dataset with few samples. The results show that our proposed model outperforms PN in terms of classification accuracy for the bearing dataset.