Insulated bearing fault diagnosis method based on shape-aware attention and dynamic physical information guidance

Abstract Most of the existing physical models of insulating bearings ignore the coupling dynamic effects between the insulating coating and the substrate, and the commonly used static physical guidance methods are difficult to adapt to the dynamic changes between data during the training process, which aggravates the domain offset between simulation data and actual data. To this end, an insulating bearing diagnosis method (SAKA-DPG) that constructs shape-aware attention (SAKA) and dynamic physical information guidance (DPG) is proposed in this paper. Firstly, the SAKA attention mechanism is constructed based on convolutional Kolmogorov–Arnold network (CKAN), and the geometric structure of the control points of the B-spline function in its kernel is analyzed in real time to achieve adaptive modulation of the contribution of each element inside the kernel function; secondly, based on the stiffness and damping of the outer ring body and the insulating coating, and simplifying the mechanical coupling and interface effects between the two, the dynamic model of the insulating bearing is constructed; then, based on the generated simulation data and actual data, the DPG method is designed, and according to the deviation changes between the simulation and actual data in different training batches, the dynamic physical constraint regularization term and dynamic data matching are constructed respectively, so as to achieve the synergistic integration between data-driven learning and physical prior knowledge. Finally, experiments show that this method significantly outperforms the baseline CKAN and other attention enhancement methods, and can still maintain a high accuracy under noisy conditions. The effectiveness and practicality of SAKA and DPG are verified through ablation experiments.