Intelligent fault diagnosis of bearings based on domain generalization class incremental transfer learning

Deep learning has been widely applied to intelligent fault diagnosis. However, most existing methods are trained offline in static settings and struggle to handle dynamic, open environments with evolving data distributions. In actual industrial applications, changing operating conditions produce non-independent and identically distributed (non-IID) data streams, while access to target domain data is often unavailable, limiting the effectiveness of conventional transfer learning methods. To address these issues, this paper proposes a domain generalization class incremental transfer learning network (DGCITL). A novel loss function combining CORAL and subspace robust embedding (SSRE) is introduced to align features across domains without target data. A domain generalization incremental learning strategy is then designed to preserve previous knowledge while accommodating new classes, leveraging information aggregation and dynamic adaptation. Instead of traditional sample replay, DGCITL employs a Graph Replay Mechanism using graph convolution to capture and reuse learned knowledge, effectively mitigating catastrophic forgetting. Experimental results show that the proposed method performs well in the incremental task of variable operating condition class, providing an efficient solution for cross-domain intelligent fault diagnosis without target domain.