Abstract
To address the issue that the deep coupling of bearing fault features and speed modulation information under variable speeds reduces the reliability of diagnostic models, and the difficulty of existing deep learning methods in balancing diagnostic accuracy and computational efficiency, this paper proposes a lightweight feature decoupling-knowledge transfer framework called CFRFN-iTransformer-KDNet. First, an iTransformer backbone model integrating a Channel-oriented Feature Refinement Feed-forward Network (CFRFN) is constructed. Through partial convolution and gating mechanisms, it achieves the separation and decoupling of fault impulse features from speed interference, while utilizing feature alignment loss to force the model to ignore speed fluctuation differences. Second, a knowledge distillation strategy is introduced to transfer discriminative knowledge from a high-performance TimesNet teacher model to the lightweight CFRFN-iTransformer student model, thereby achieving a unification of high accuracy and high efficiency. Experiments on the multi-source fault dataset of pitch bearings from China University of Mining and Technology (CUMTB) show that the proposed framework achieves diagnostic accuracies of 98.64%, 97.81%, and 98.70% under 1 rpm, 3 rpm, and mixed-speed conditions, respectively, with a maximum performance fluctuation of only 0.89 percentage points. Further validation on the University of Ottawa variable speed bearing vibration dataset (UOttawa) shows that the framework achieves an average diagnostic accuracy of over 97% under various continuous variable speed modes, with performance fluctuations across different conditions of less than 0.5 percentage points. Experimental results indicate that the proposed framework has only 0.16 M parameters and a computational cost of 0.03 × 10
9
FLOPs. While maintaining extremely low resource consumption, it significantly improves the diagnostic stability and generalization capability of the model under variable speed conditions.