Cross-domain Remaining Useful Life prediction under unseen condition via Mixed Data and Domain Generalization

Cross-domain Remaining Useful Life (RUL) prediction is an important issue in Prognostic and Health Management (PHM) and has attracted widespread academic attention. However, most of the current methods require access to the target domain at training time, limiting their application in real industrial environment. In this paper, we propose a Mixed Data and Domain Generalization (MDDG) framework for cross-domain RUL prediction under unseen conditions. Our approach implements data augmentation from both class space and domain space to increase the diversity of data. A domain discriminator is designed to distinguish semantic information between multiple data domains to help the model improve semantic discrimination. In addition, a domain discrepancy metric module is employed to balance the distance between data domains, thus suppressing distributional differences. Extensive experiments on two benchmark databases validate the superiority of our method.