Digital Twin Implementation for Three-Dimensional Rotordynamic Response via Physics-Informed LSTM Neural Networks

Abstract

The rotating assemblies of critical machinery are complex dynamical systems and rotordynamic model response prediction inaccuracy risks machinery failure leading to high production losses. Jeffcott, Euler beam, and high-fidelity 3D solid finite element models are frequently utilized for rotordynamic analyses. Even though the 3D rotor has the higher accuracy, beam models are still widely used in industrial applications. To improve prediction accuracy of the lower-fidelity Jeffcott and beam models, a rotordynamics physics-informed neural network (R-PINN) is proposed. This models physics-informed long short-term memory (LSTM) neural networks that utilize partial or limited measured data, by incorporating physical laws. This approach enables the creation of a Digital Twin, which can produce additional data and help remove noise and outliers. In the current study, two R-PINNs are introduced to validate the superior capability of the model for both low- and high-fidelity physics. Random noise of 10% is introduced into the measured data produced by the Digital Twin to replicate real-world noisy measurements. The result shows that both low- and high-fidelity physics R-PINNs can achieve high accuracy even with high noise data, thereby increasing the robustness of the model. The results clearly demonstrate the ability of the proposed R-PINN algorithm to enhance an Euler beam model's predicted response to the level of accuracy of a 3D solid element model's predicted response, the latter acting as a surrogate for test measurements in an actual application.