Replicating Existing Axial Magnetic Bearing Controller With a Neural Network

Abstract

In various industrial applications, neural network-based control solutions can present a viable alternative to traditional control laws. The adaptability of these solutions allows the control law to be trained through data observations by considering the tools of deep learning. One of the example fields is replacing an existing controller with a neural network with the idea that the network is trained to mimic the control law. This paper focuses on the replacement of the axial active magnetic bearing (AMB) controller with a nonlinear autoregressive with external input (NARX) neural network structure. The learning process is treated as a black box, meaning there is no prior knowledge of the controller, and it utilizes input/output data for training. A step-by-step fitting procedure is applied and the obtained neural network structures are linearized to enable frequency domain analysis of the control performance. The obtained controllers are evaluated with electrical machine with AMB suspended rotor system.