Physics-Informed Neural Networks for the Condition Monitoring of Rotating Shafts.

Area of Science:

• Mechanical Engineering
• Computational Science
• Artificial Intelligence

Background:

• Condition monitoring of rotating shafts is crucial for industrial machinery reliability.
• Deep learning excels at pattern recognition but often lacks physical interpretability.
• Integrating physics with data-driven models addresses the 'black-box' nature of deep learning.

Purpose of the Study:

• To apply physics-informed neural networks (PINNs) for rotating machinery condition monitoring.
• To estimate key health parameters of an extended Jeffcott rotor model.
• To evaluate PINNs' performance against traditional methods, including noisy data scenarios.

Main Methods:

• Utilized PINNs to model an extended Jeffcott rotor with damping and anisotropic supports.
• Estimated five health-related parameters: unbalance (radial, angular), stiffness (principal axes), and damping coefficient.
• Employed displacement signals from the rotor disk center for parameter estimation.
• Analyzed performance across various rotational speeds and in the presence of data noise.

Main Results:

• PINNs successfully estimated the five characteristic parameters of the rotor system.
• The methodology demonstrated efficacy and precision in identifying system health states.
• Performance was robust across different rotational speeds and under noisy data conditions.
• PINNs showed comparable or superior performance to traditional optimization algorithms.

Conclusions:

• PINNs offer a powerful, physics-aware approach for rotating machinery condition monitoring.
• This method enhances the interpretability and reliability of deep learning in mechanical systems.
• PINNs provide a promising alternative to conventional parameter estimation techniques, especially in complex industrial settings.