Confidence interval networks for bounding model uncertainty: experimental evaluations on an active magnetic bearing system

Robust control techniques require bounds on model uncertainties to provide closed-loop stability in the face of unmodeled dynamics and parameter variations. To guarantee stability, it is customary to choose uncertainty bounds (weighting functions) that are somewhat arbitrary and overly conservative, usually at the expense of controller performance. In this paper, an intelligent approach for estimating additive uncertainty bounds associated with linear, time invariant models is presented. Confidence interval networks (CINs) provide a non-parametric method for identifying bounds on modeling error for use as uncertainty weighting functions. The CIN is a "soft-computing" variation of the model error modeling (MEM) technique, a parametric approach based on recursive least squares. By combining these methods with the "hard computing" aspects of H_∞ control, the size of the uncertainty model is optimized, thus improving performance while maintaining robust stability. A multivariable, flexible-rotor active magnetic bearing system is used to experimentally demonstrate the benefits of intelligent uncertainty identification.