A novel multiple-prototype and domain adversarial network for few-shot cross-domain fault diagnosis

Abstract Since deep learning has been introduced into the field of intelligent fault diagnosis it has made significant accomplishments with large amounts of data. However, in practical industrial settings there is a general lack of labeled data and operating conditions are not stable, therefore existing trained models ignore these problems and diagnostic accuracy and generalization are severely degraded. Therefore, this study proposes a multiple-prototype, domain adversarial network for fault diagnosis of rotating machinery, especially bearings, to address the issues of domain distribution shift and a shortage of labeled samples from the target domain. Firstly, the proposed method adopts a residual network with 12 layers (ResNet12) as the encoder to extract distinct features from the target domain and source domain data. Then, a domain-invariant representation module, using the domain adversarial method, is introduced to bridge the discrepancy between a source and target pair domain. Furthermore, a class typicality weight module is designed to calculate weights by constructing multiple prototypes of the source and target domains to improve the generalization of the model. Extensive experiments are executed on two datasets with variable working conditions to test and verify the feasibility and superiority of the proposed method.