Early Bearing Fault Evolution Modeling with a Dynamics-Constrained Graph Neural Network

During the incipient localized-fault stage, impact signatures in ball bearings are weak. Accurate degradation labels are difficult to construct. Purely data-driven models often lack physical consistency. To address these issues, a dynamics-constrained graph neural network (DC-GNN) was developed for early fault evolution modeling in ball bearings. Firstly, a bearing dynamic model was established based on nonlinear ball–raceway contact mechanics. A fault evolution parameter was then introduced. The incipient localized fault was represented by contact stiffness degradation. It was also described through increased local additional displacement. This formulation captured the continuous transition from a healthy condition to slight localized damage. Secondly, the bearing was represented as a graph structure. The inner ring, outer ring and balls were defined as nodes. Ball–raceway contacts were treated as edges. Contact deformation, contact stiffness, contact force, and a fault-region indicator were embedded into node or edge features. A DC-GNN was then constructed. Finally, dynamic simulations were performed under different fault evolution levels. Bearing vibration data were also used for validation. The results show that contact-force fluctuations intensified as the fault evolution parameter increased. The time-domain impact amplitude also increased. Fault characteristic components became more prominent in the envelope spectrum. The degradation indicator exhibited a consistent evolution trend. These variations agreed well with incipient localized fault progression. This study provides an effective approach for early fault evolution characterization in rolling bearings.