In industrial settings, rotating machinery fault diagnosis often faces continually changing operating conditions, such as motor speed and load, causing distribution shifts in the same fault type across conditions. Although traditional continual learning methods can adapt to new tasks, they struggle to balance cross-condition feature reuse with condition specific representation. To address this issue, this study proposes a Dual-Level Collaborative Path Selection incremental fault diagnosis model (DLCPS-Net) for continually changing operating conditions. DLCPS-Net adopts a dual-level archi­tecture with shallow sharing and deep differentiation. The shared layers are reused across tasks to preserve fault representations that are universal across operating conditions, while the differentiated layers are dynamically selected and expanded based on task requirements to capture condition specific features. An interstage shared representation constraint is introduced to limit excessive drift in shared representations while retaining sufficient flexibility for new task adaptation. Additionally, a multi-path fusion mechanism and path entropy regularization loss adaptively learn the importance of historical paths, enabling effective historical knowledge recombination and improving knowledge reuse efficiency. An incremental diagnostic task under continually changing conditions is constructed on the BJTU-RAO bogie dataset. Experimental results show that DLCPS-Net outperforms several repre­sentative continual learning methods in both diagnostic accuracy and knowledge retention.

fault diagnosis,continual learning,path selection,catastrophic forgetting,rotating machinery