Eventos Anais de eventos
COBEM 2021
26th International Congress of Mechanical Engineering
Physics Informed Neural Networks to Identify Unbalance Parameters in Rotating System
Submission Author:
Lucas Garpelli , SP
Co-Authors:
Lucas Garpelli, Helio Fiori de Castro
Presenter: Lucas Garpelli
doi://10.26678/ABCM.COBEM2021.COB2021-1253
Abstract
The improvement of computational resources and a large amount of data produced by different areas of science have made the neural network very relevant among researchers, exploiting their ability to deal with complex problems. However, some applications have difficulty in collecting data for neural network training, either due to the inability to obtain them or the high cost associated with their acquisition. Physics Informed Neural Networks (PINN) use the information from the model that describes the physics behaviour of the analysed system to reduce the need for large amounts of data for network training. In this work, a rotating system was modelled, in order to obtain the dynamic behaviour in frequency domain response. These responses are then used in the PINN to identify the unbalance fault parameters applied to the rotating system. The equation of motion is used to train the neural network, so it is possible to find the best weights and biases that corresponds the rotor orbit and the unbalance fault parameters. The weights and biases are trained considering two steps of solution: the direct problem, in which the guess of unbalance parameters are used to calculate the rotor orbit, and the inverse problem, in which the difference between the rotor orbit of the direct problem and the real orbit is used to find a new guess of unbalanced parameters. This type of solution is important in rotating machines because can be used to counterbalance the rotating system or identify possible residual unbalance. The knowledge of the unbalance level of a rotating system is very important to its design and maintenance. This information is used to counterbalance the rotor and reduce the vibration level, knowing the strict technical standards. Thus, PINN represents an important tool, since it provides good results with a reduced amount of data and computational cost.
Keywords
Rotordynamics, Unbalance Parameters, Neural Network, PINN
