<p>Accurate and effective calculation of the range of equivalent stress intensity factors is crucial for assessing the risk level of crack propagation in rolling bearings. In order to improve the calculation efficiency of the initial crack stress intensity factor range of rolling bearings, this study proposes a finite element neural network decoupling framework. This approach preserves physical mechanisms via finite element analysis while enabling real-time prediction through neural networks. First, a precise finite element model of GCr15 bearing steel was established. Stress intensity factors at the crack tip were computed through interactive analysis between ABAQUS and FRANC3D, with FRANC3D verifying finite element mesh convergence (Hertz contact error &lt; 5%) prior to stress intensity factor calculation. Subsequently, a dataset of 216 samples was constructed, incorporating loading conditions, crack geometric parameters, and corresponding stress intensity factor range values. Bayesian regularization was employed to mitigate overfitting in this limited sample size (216 data points). A dual-hidden-layer Bayesian Regularized Backpropagation Neural Network (15 nodes/layer) was designed, with the two-hidden-layer architecture determined through sensitivity analysis, to achieve rapid prediction of stress intensity factor range. The proposed model achieved a mean absolute percentage error of just 3.61% on the test set, with prediction errors below 7% for unseen parameters. Compared to conventional simulations, this framework improved computational efficiency by over 90%, reducing the stress intensity factor range prediction time from hours to seconds. This method provides a robust and effective engineering solution for rapidly assessing the risk level of crack propagation in rolling bearings.</p>