Abstract
Active Magnetic Bearings (AMB) guide a rotating shaft without mechanical contact by means of magnetic levitation. They enable high speeds without friction, and meet the challenges of energy efficiency and decarbonization.
Unlike radial bearings, which are made up of stacks of laminations, axial PMAs are massive, inducing high eddy currents; magnetic saturation adds strong non-linearities, making transient simulation costly. We aim to build real-time-compatible reduced dynamic models from finite element simulations.
We first introduce an order-reduction method based on orthogonal eigenvalue decomposition (POD), combined with hyper-reduction of nonlinearities. We then propose a hybrid approach "model+data" based on an equivalent electrical circuit of the Cauer network type. A reduced space of model error is identified, then a recurrent network predicts correction weights, improving magnetic flux prediction and enabling field reconstruction.
Finally, we explore physics-informed hybrid models where the resistances and inductances of the equivalent circuit are corrected by neural networks, leading to a universal differential equation (UDE) for accurate, real-time-compatible flux prediction.
