Universal source-free domain adaptation method for cross-domain fault diagnosis of machines

Abstract

Cross-domain machinery fault diagnosis aims to transfer enriched diagnosis knowledge from a labeled source domain to a new unlabeled target domain. Most existing methods assume that the prior information on the fault modes of the target domain is known in advance. However, in engineering practice, prior knowledge of fault modes is rare in a new domain, in which there may be only partial source fault modes or some new fault modes. Furthermore, up to the present, almost all existing cross-domain fault diagnosis methods require the labeled source data during the model training process, which restricts their deployment on certain devices with limited computing resources. To this end, we propose a universal source-free domain adaptation method that can handle cross-domain fault diagnosis scenarios without access to the source data and is free of explicit assumptions about the target fault modes. More specifically, we develop a convolutional network with a Transformer as the attention module to extract discriminative feature information from the source data and then send the model and parameters to the target domain. In target domain training, we first propose a supervised contrastive learning strategy based on source class prototypes, which utilizes high-confident predictions to achieve source-free domain alignment and class alignment. Then, we also introduce a threshold-based entropy max–min loss to further align known class samples in the target domain or reject target outlier samples as an unknown class. Furthermore, we introduce self-supervised learning to further learn feature representations of the target domain to reduce the previous misclassification. A series of experiments on two rotating machine datasets demonstrate the effectiveness and practicability of the proposed method.