Intelligent control using confidence interval networks: Applications to robust control of active magnetic bearings

Robust control synthesis requires an explicit mathematical description of the system dynamics (a model) and uncertainty bounds associated with that model. These uncertainty bounds are usually chosen arbitrarily and conservatively for guaranteed stability, frequently at the expense of controller performance. This research demonstrates the application of Confidence Interval Networks (CINs), unique artificial neural networks that utilize asymmetric bilinear error cost functions, for estimating the bounds of model uncertainty required for robust control synthesis. A highly nonlinear and unstable active magnetic bearing (AMB) system is considered. A high-speed flexible rotor supported by AMBs is modeled using analytical approaches, finite element analysis, and system identification. CINs "learn" the statistical bounds of model uncertainty resulting from unmodeled dynamics and parameter variations. These bounds are incorporated into the synthesis of multivariable robust controllers based on two approaches, linear time invariant and linear parameter varying. Experimental results on a multivariable AMB test rig reveal the benefits of this combination of intelligent system identification and robust control: significant performance improvements vs. conventional robust control with and without mass imbalance (process disturbances).