Data-driven modelling of touch-down bearing forces

To reduce CO2 emissions, the share of renewable energies in the grid increases. At the same time, many sectors like the transport and the building sector are changing to be powered by electricity. Especially electric cars demand high peaks in current from the grid. Storage is needed to balance demand and supply of electric energy. Flywheels can be part of the solution as they can be charged and discharged with high power and do not suffer from losing significant capacity even after thousands of cycles. Minimum loss of energy is crucial for a flywheel therefor active magnetic bearings (AMB) are used. If a malfunction of the AMB occurs the rotor falls into a touch-down bearing (TDB). To decide whether further maintenance in case of a drop-down event is needed information about the forces stressing the TDB is important. To avoid costs for physical sensors soft sensors are a suitable solution. In this research, a data-driven soft sensor based on recurrent neural networks is created to calculate the forces during the drop-down event. As input data only the position of the rotor is used. A test rig with physical sensors applied to every TDB supplies the force data to train, validate, and test the soft sensor model. Three different network architectures are compared. The results show that the sensor can calculate whether the rotor hits a TDB and is also capable of predicting the peaks in the force signal.