Real-time stress field prediction of planetary gears based on POD model order reduction and deep residual networks

Abstract

To overcome the high computational cost of finite element simulations for planetary gear stress fields under complex operating conditions, this paper proposes a fast prediction method combining Proper Orthogonal Decomposition (POD) and a Deep Residual Network (Res-DNN). High-fidelity stress snapshots under angular velocities of 1–9 rad/s (1 rad/s increment, 9 conditions in total) are generated using LS-DYNA. POD extracts principal modes, compressing the high-dimensional field into a low-dimensional space. A Res-DNN then maps time and angular velocity to POD coefficients. With 20 modes (97.6% energy), the model achieves a mean absolute error of 0.43 MPa and a speedup of ∼1.7×106 over LS-DYNA. Note: All results are based on LS-DYNA simulations. Experimental validation on a vibration test rig (Section 4.3) shows good agreement with surrogate predictions for in-range operating conditions (average MAE 0.59 MPa), confirming the model’s practical potential.