Experimental Testing of a Magnetically Levitated Rotor With a Neural Network Controller

Magnetic bearings represent a solution for high rotating speeds and sterile environments where lubrication fluids could contaminate. They can also be used in systems where maintenance is difficult or inaccessible, because they don’t require auxiliary lubrication systems and don’t suffer mechanic wear as they work with no contact between rotor and bearing stator. An important part of magnetic bearings is the controller; which is needed to stabilize the system. This controller is generally a PID in which tuning and/or filters design can be complicated for not well known systems. This work presents results of the development of a neural network controller, which is potentially easier to implement, to control the position of a magnetically suspended rotor. The proposed controller is based in the identification of the system inverse model. This is achieved first by implementing a simple PID capable of levitating the rotor, and then some excitations are applied to the rotor in order to acquire data of the position of the rotor and current in the actuators. Current and position data is used to train the artificial neural network for the controller. The controller was implemented in a numerical model and also in an experimental system with a rotor of 1.06kg and 300mm in length. The implementation of SISO, MISO and MIMO neural controllers (both with offline and online training) and a conventional PID with neural network compensation are compared. Structures and architectures of networks are shown. Vibration responses to: a constant force; a controlled impact and a constant acceleration ramp between 0 and 12500rpm are compared. Results in both, numeric model and experimental system, show that neural network controllers are capable of hovering the rotor and control vibrations. Peak-Peak amplitudes vs. rpm plots are similar to a conventional PID. In most cases, the neural network controllers show amplitudes slightly lowers on low frequencies and slightly higher on higher frequencies, except the conventional PID with neural network compensation case, were the system responses as with higher damping. Finally, a discussion is made about future steps in research to improve implementation of a neural controller that is potentially simpler and faster in terms of tuning and with a comparable performance to a conventional magnetic bearing PID controller.