Improved Physics-Guided PINN Design and its Application in Bearing Acceleration Response Modeling

In engineering, dynamic and fault modeling is crucial for predicting system behavior and diagnosing faults to ensure reliable and efficient operation. Recently, physics-informed neural networks (PINNs) have gained significant attention and demonstrated great potential in modeling applications. However, general PINNs (G-PINNs) are not well-suited for long-period simulations, and embedding complex dynamic equations can be challenging, leading to difficulties in training. This paper proposes an improved physics-guided PINN (IP-PINN) approach to address these issues. The method utilizes three key techniques: frequency domain loss function, transfer learning, and sine activation function. To demonstrate the advantages of the IP-PINN approach, the acceleration responses of bearings with outer race fault, inner race fault, and ball fault are modeled and compared with simulation data. The results show that the IP-PINN method provides accurate solutions even for long-duration simulations. Additionally, analyzing the impact of different weights for the frequency domain loss and activation functions allows for the selection of suitable parameters for various fault types. This approach addresses the limitations of G-PINNs and offers a novel method for bearing fault signal modeling.