This paper proposes a Domain-Disentangled Generative Replay (DDGR) framework for lifelong fault diagnosis of rotating machinery with fault-domain increments under variable operating conditions. A dual-stream encoder is developed to explicitly model operating-domain-related variations while maintaining fault-domain-discriminative content representations through operating-domain supervision, orthogonality regularization, and Instance Normalization. In addition, a boundary-aware generative replay mechanism based on label-wise Gaussian mixture models is employed to preserve historical decision-boundary information in the latent space without storing raw vibration signals. Experiments on a high-speed train gearbox bearing dataset demonstrate that DDGR effectively mitigates catastrophic forgetting during sequential fault-domain increments and achieves a final-phase diagnostic accuracy of 87.14%, outperforming representative continual-learning baselines.

Lifelong learning, fault diagnosis, rotating machinery, fault-domain increments, variable operating conditions, generative replay