Objectives Deep learning has shown great potential in the field of rotating machinery fault diagnosis. Its excellent performance heavily relies on sufficient training samples. However, in practical engineering applications, acquiring sufficient training data is particularly difficult, resulting in poor generalization capability and low diagnostic accuracy. Therefore, few-sample fault diagnosis methods, which can effectively extract fault-related information from limited data, have gradually become a research focus in both academic and engineering circles.

Method In this paper, the latest achievements in few-sample fault diagnosis of rotating machinery are reviewed and summarized. This paper describes the definition and learning methods for few-sample fault diagnosis. Few-sample fault diagnosis methods aim to effectively develop fault diagnosis models with strong generalization capability under limited training data conditions. Currently, according to different technical principles, existing few-sample fault diagnosis methods can be classified into five categories: meta-learning, transfer learning, domain generalization, data augmentation, and self-supervised learning. Subsequently, this paper elaborates on the applications of these five methods in rotating machinery fault diagnosis. Meta-learning-based fault diagnosis methods improve the ability of models to rapidly learn and adapt to new tasks by acquiring common knowledge from multiple related tasks. The transfer learning-based fault diagnosis methods achieve knowledge migration from the source domain to the target domain using unsupervised domain adaptation techniques. The domain generalization-based fault diagnosis methods train models using single or multiple source domains and enable the model to learn features that are common across those domains. The data augmentation-based fault diagnosis methods expand the original dataset by generating models. The self-supervised learning-based fault diagnosis methods exploit the structural information of data to construct pseudo-labels.

Results The paper summarizes the core ideas, advantages, and limitations of these five methods. Meta-learning can improve the model's generalization capability but may require significant computational resources. Transfer learning can improve learning efficiency but is limited by domain similarity. Domain generalization can enhance the model performance in unknown domains but may suffer from overfitting issues. Data augmentation can increase dataset diversity but may generate inconsistent samples. Self-supervised learning can utilize unlabeled data but faces challenges such as complex task design and potential overfitting.

Conclusions In the future, data governance, multimodal learning, federated learning, and mechanism-data hybrid-driven methods should be further explored in the field of few-sample fault diagnosis. It will overcome the limitations of existing methods and further improve the reliability of few-sample fault diagnosis.