Temperature prediction of rub-impact in a rotor-stator system based on physics-informed neural networks

Rub-impact of the rotor-stator system adversely affects the reliability and safety of the system. Accurately predicting of the temperature field is crucial for ensuring service performance during operation of the rotor-stator system. A rub-impact physics-informed neural network (RI-PINN) is proposed to predict the temperature field of rub-impact based on Fourier feature mapping (FFM) and physics-informed neural network (PINN). The FFM significantly enhances the ability of PINN to learn high-frequency features and effectively improves the prediction accuracy of the proposed network. The results indicate that compared to deep neural network (DNN) and PINN models, the proposed RI-PINN model achieves significantly lower relative error, root mean square error (RMSE), and maximum error than DNN and PINN models, demonstrating substantial advantages in computational efficiency and prediction accuracy. The impact of FFM parameters on network training is investigated, and it is found that the highest prediction accuracy and computational efficiency are achieved when the scaling factor is set to 32 and the mapping dimension to 128.