Conventional data-driven fault diagnosis methods rely on large amounts of labeled data and stable operating conditions. However, in practical industrial scenarios, fault sample scarcity and frequent condition variations are commonly encountered. This coexistence of data deficiency and domain shift severely limits the model’s generalization ability in few-shot cross-domain tasks. This work proposes a mechanical fault diagnosis method based on a meta-transfer relation network. First, a relation network is employed to transform the fault data classification task into a similarity measurement problem, and a multi-task training mechanism based on meta-learning is introduced to reduce the model’s dependence on sample quantity. Second, to address domain shift, the model incorporates a parameter transfer strategy, in which the general fault knowledge obtained through meta-learning from source-domain tasks is used as the initial model parameters for target-domain tasks, enabling rapid adaptation through a small number of gradient updates. In addition, to enhance the model’s ability to recognize complex fault patterns, a lightweight multi-scale convolutional module is designed in the feature extraction stage to improve multi-scale feature capture while controlling computational cost. In the relation measurement stage, a class traversal mechanism is designed to extract highly discriminative features by utilizing the global information of the support set, further strengthening class representations. Finally, the model is evaluated under various few-shot cross-domain scenarios, including speed transfer, load transfer, noise interference, and cross-device transfer. The results demonstrate that the proposed method achieves higher accuracy and robustness than existing fault diagnosis models.
Hong-wei et al. (Mon,) studied this question.