PC-CIL: a physical consistency class-incremental learning framework for fault diagnosis of rotating machinery under variable operating conditions

When rotating machinery encounters a continuous flow of new fault categories under variable operating conditions, traditional deep learning-based diagnostic methods fail to handle feature distribution shifts and suffer from catastrophic forgetting, significantly compromising long-term monitoring reliability. To address this problem, we propose a physical consistency class-incremental learning (PC-CIL) framework. Physical consistency representation learning serves as the model backbone, incorporating universal physical statistical features as prior constraints to bridge the gap between data-driven models and physical laws. We then introduce a pseudo-sample generation mechanism based on prototype boundary sampling and confidence filtering to tackle data isolation in incremental scenarios. The generated pseudo-samples enable a hybrid replay strategy to dynamically update decision boundaries while suppressing negative transfer with low computational overhead. To mitigate catastrophic forgetting, we implemented a topology-preserving relational knowledge distillation approach, thus facilitating the structural consistency of the feature space. Experimental results on two publicly available benchmark datasets, namely the HUST and CWRU bearing datasets, as well as a self-built bearing dataset, demonstrate that PC-CIL consistently maintains high diagnostic accuracy while effectively mitigating catastrophic forgetting during sequential fault-category expansion. The consistent improvements across different bearing types, test rigs, sampling systems, and operating conditions further verify the robustness and generalization capability of the proposed framework.