Model updating in rotor dynamic system using FEM based neural network

Executable Digital Twins (xDTs) assist maintenance engineers by calculating current and future health status of physical assets based on operational conditions. In xDTs, the simulation describes the ideal physical behavior of the system, whereas the real data from sensors assist in updating the xDTs following the changes to the physical asset. We introduce a novel algorithm combining neural networks (NN) and finite element method (FEM) for the system updating in rotor dynamic systems xDTs. In the presented algorithm, the known parts of the system are modeled using FEM package, whereas the unknown parts such as bearing parameters is modeled using a neural network. The discretized form of the residual of the governing partial differential equation (PDE) is backpropagated to learn the neural network parameters. The estimated parameters are used to predict the vibration in an unseen scenario. The algorithm facilitates an ecosystem combining numerical methods and data science to result in more accurate and continuously-updating xDTs.