A novel inter-domain attention-based adversarial network for aero-engine partial unsupervised cross-domain fault diagnosis

Recently, domain adaptation methods have been widely applied in the field of aero-engine cross-domain fault diagnosis, which can effectively solve the problem of training and testing data coming from different distributions. However, existing methods commonly assume that the health states in the source and target domains are identical. In practice, the health states in the target domain are often a subset of the source domain, and there is no prior information on the target health states, which is defined as a partial unsupervised cross-domain fault diagnosis task. The most difficult problem in this scenario is determining the shared health state label space between the source and target domains. To address this issue, we propose an inter-domain attention mechanism that enables the neural network structure itself to dynamically learn the relationship between each source sample and each target sample by constructing a source sample attention vector. Furthermore, we propose an inter-domain attention-based encoder that can be seamlessly integrated into the widely adopted domain adversarial neural network (DANN). Finally, we conduct extensive experiments on two aero-engine datasets and the CWRU dataset, and the results show that our method achieves at least 1.05% improvement in average accuracy compared to other methods in the partial unsupervised cross-domain fault diagnosis task.