Campbell Diagram Prediction for an Overhung Rotor through Physics-Informed Neural Networks
پذیرفته شده برای پوستر
کد مقاله : 1059-ISAV2025 (R1)
نویسندگان
1دانشکده مهندسی مکانیک، دانشگاه صنعتی امیرکبیر، تهران، ایران
2آزمایشگاه تحقیقاتی آکوستیک، دانشکدهی مهندسی مکانیک، دانشگاه صنعتی امیرکبیر، تهران، ایران
چکیده
The accurate prediction of rotor dynamic behavior, encapsulated in Campbell diagrams, is critical for the design and safe operation of rotating machinery. This paper presents a novel methodology employing a Physics-Informed Neural Network (PINN) to predict the Campbell diagram of a five-degree-of-freedom overhung rotor system with gyroscopic coupling and anisotropic bearings. The developed PINN model maps rotational speed to natural frequencies using a deep network architecture that directly embeds the governing physical equations as residual constraints during training. The results demonstrate that the model successfully generates a high-fidelity Campbell diagram, accurately identifying five distinct vibrational modes—including forward/backward whirl and a speed-independent torsional mode—with an average absolute error of 8.2 Hz (3-5% relative error) against a numerical benchmark. The training process was computationally efficient, converging stably in approximately 15 minutes on a standard CPU. Crucially, the model exhibits exceptional interpolation accuracy within the trained speed range (300-3000 rpm), with a mean absolute error of 0.13 Hz, confirming its ability to learn the underlying system dynamics without overfitting. While a moderate decrease in accuracy was observed during extrapolation beyond 3000 rpm, the model's predictions remained physically consistent, correctly preserving the whirl characteristics of all modes. This work establishes the proposed PINN framework as a robust, efficient, and accurate tool for rotor dynamics analysis, providing a solid foundation for future research aimed at enhancing generalizability across a wider range of operational conditions and system configurations.
کلیدواژه ها
Title
Campbell Diagram Prediction for an Overhung Rotor through Physics-Informed Neural Networks
Authors
Amirpasha Afsari, Emadaldin Sh Khoram-Nejad
Abstract
The accurate prediction of rotor dynamic behavior, encapsulated in Campbell diagrams, is critical for the design and safe operation of rotating machinery. This paper presents a novel methodology employing a Physics-Informed Neural Network (PINN) to predict the Campbell diagram of a five-degree-of-freedom overhung rotor system with gyroscopic coupling and anisotropic bearings. The developed PINN model maps rotational speed to natural frequencies using a deep network architecture that directly embeds the governing physical equations as residual constraints during training. The results demonstrate that the model successfully generates a high-fidelity Campbell diagram, accurately identifying five distinct vibrational modes—including forward/backward whirl and a speed-independent torsional mode—with an average absolute error of 8.2 Hz (3-5% relative error) against a numerical benchmark. The training process was computationally efficient, converging stably in approximately 15 minutes on a standard CPU. Crucially, the model exhibits exceptional interpolation accuracy within the trained speed range (300-3000 rpm), with a mean absolute error of 0.13 Hz, confirming its ability to learn the underlying system dynamics without overfitting. While a moderate decrease in accuracy was observed during extrapolation beyond 3000 rpm, the model's predictions remained physically consistent, correctly preserving the whirl characteristics of all modes. This work establishes the proposed PINN framework as a robust, efficient, and accurate tool for rotor dynamics analysis, providing a solid foundation for future research aimed at enhancing generalizability across a wider range of operational conditions and system configurations.
Keywords
Campbell Diagram, PINN Diagram, Jeffcott System, Rotor Dynamics