Novel joint transfer fine-grained metric network for cross-domain few-shot fault diagnosis

Traditional deep learning fails to identify new faults when the number of faulty samples is limited. Existing meta-learning studies on cross-domain small-sample fault diagnosis do not fully account for the differences in the distribution of faults across domains between training data and new fault classes, which limits further performance improvement of the diagnosis model. In this study, we propose a new joint transfer fine-grained metric network for cross-domain few-shot fault diagnosis. First, a hybrid attention is used to enhance the feature extraction ability of the model and suppress redundant features. The proposed joint transfer function is used to align the corresponding subdomains of the source and target domains in small samples. Finally, cross-domain few-shot fault diagnostics is implemented using a modified fine-grained metric network. The proposed joint transfer fine-grained metric network outperformed other common meta-learning fault diagnosis methods on three mechanical device datasets.