Inverse Identification of Contact Forces and Prediction of Roller Stress Fields for Large Main Bearings Based on Physics-Informed Neural Network

The main bearings of large rotating machinery often encounter heavy overloads, significant offset loads, and variable loads during operation. Accurately sensing the internal contact force and the force field distribution of the rollers is crucial for evaluating the health status of the main bearing. To address the difficulty of precisely measuring contact forces of the main bearing, this study proposes an inverse calculation framework for the contact force within the main bearing based on Physics-Informed Neural Networks (PINN). The proposed method embeds the equilibrium-physical-geometric equations and boundary constraints that describe the linear elastic contact boundary value problem into the loss function. By leveraging physical laws and limited observable data from roller end faces, the loss function is optimized through a two-stage process involving Adam and L-BFGS-B optimizers. This method predicts the distribution of displacement and stress fields, enabling the inverse identification of contact forces. Simulation experiments demonstrate that, driven by shear stress data, the coefficient of determination R2 for the inverse contact force reaches 93.92%.