Invariant Feature Learning and Open-Space Feature Synthesis for Open-Set Domain Generalization Fault Diagnosis of Rotating Machinery

Current domain generalization-based diagnosis methods mostly assume that label spaces among source and target domains are identical. However, in real industry, what fault occurs is hard to know in advance. It is possible that fault types unseen in the source domains occur during target equipment operation. Take gearbox diagnosis as an example, source domains may only contain bearing faults while the target machine may occur gear faults. To address the above issue, this paper proposes an open set domain generalization fault diagnosis method for rotating machinery, where known classification and unknown detection are both considered for the unseen target domain. In the proposed method, class-wise and known-universal invariant features are learned to solve domain shifts and improve the separation between known classes and unknown faults. Furthermore, a novel open set feature synthesis strategy is developed, introducing open space information to jointly refine invariance feature learning and known class boundaries. In two diagnosis case studies, the superiority of the proposed method has been verified with 3.2% and 10.0% improvement in the comprehensive performance of known classification and unknown detection, measured by H-score, over the compared methods.