A Few-Shot Machinery Fault Diagnosis Framework Based on Self-Supervised Signal Representation Learning

The intelligent fault diagnosis method based on deep learning has achieved promising results in recent years; however, the performance of most models requires many labeled samples for training, which is usually impractical in real industry situations. At the same time, large amounts of unlabeled operational data are easily available. It is of great significance to efficiently harness and leverage the wealth of information encapsulated within unlabeled data and build a robust deep learning model with limited labeled samples; thus, we propose a novel few-shot learning model that combines the unlabeled signal representation learning idea with the few-shot learning algorithm. The proposed model first employs self-supervised learning (SSL) to obtain inherent features of the signal from massive unlabeled samples. Subsequently, the acquired features are transferred to an improved Siamese network to enhance its robustness and generalization on few-shot datasets. This method not only provides a novel solution for unlabeled signal feature learning but also further promotes the few-shot learning method to become a more robust and practical technique. We verify the proposed method on two fault diagnosis data sets, and the experiments verify that the proposed model achieves excellent performance under extremely limited labeled training samples.