Knowledge regroupment and preference calibration framework for unpredicted fault diagnosis under unknown working conditions

Abstract

In engineering scenarios, the absence of target domain data during model training phase leads to uncertainties in the encountered faults and working conditions. Identifying unpredicted faults including known and unknown classes under new working conditions poses a realistic and great challenge for reliability of mechanical system. The challenge is further intensified by inconsistent fault label space among multi-source domains. Current intelligent diagnostic models will be terribly powerless on such real-time diagnostic situation of open set domain generalization with category shift among multi-source domains. Therefore, to overcome this extremely challenging issue in practice, a knowledge regroupment and preference calibration framework (KRPCF) is established in this study. A knowledge regroupment scheme for generalizable feature learning and a categorical preference calibration loss for open set fault classifier training are innovatively designed in KRPCF to simultaneously solve domain shift among domains, category shift among source domains, and category shift between source domains and unseen target domains. In comprehensive experimental results based on various performance evaluations, average accuracy and average H-score of KRPCF surpass the best baseline method by more than 4.6% and 8.1%, respectively, which demonstrates the strong potential of KRPCF in practical applications. In-depth discussion on the visualized preferences of the open set fault classifier and the robustness of KRPCF demonstrates its reliable unpredicted fault diagnosis under open set domain generalization with category shift among source domains.