Active magnetic bearings (AMBs) employ electromagnets to levitate the rotor in rotating machinery so that it spins without contacting the stator. Conventionally, magnetic forces are controlled using feedback from sensors that measure shaft position. Although rotor position sensors work well in most applications, their cost can be prohibitive. Additionally, it is often necessary to operate sensors in extreme environments, which increases sensor cost. Sensors also cannot be placed at the same location as the bearings (colocation), which can be a source of control problems and even control instability. A further disadvantage in utilizing sensors in AMB applications is that they increase the number of points of potential failure. The sensorless AMB exploits the fact that the magnetic actuators themselves can act as sensors. That is, the relationship between magnetic coil driving current and voltage is affected by the electrical inductance, which (ideally) is inversely proportional to the clearance between the stator and the rotor. Neural networks can learn this nonlinear mapping between the driving current and voltage and rotor position. By training the neural network on data from an operating machine, the mapping will necessarily also account for the effects of flux leakage, magnetization curve nonlinearity, eddy currents, and hysteresis.
