Data-driven observer design for an inertia wheel pendulum with static friction

Abstract

An indirect data-driven state observer design approach for the inertia wheel pendulum considering static friction of the actuated inertia disc is presented. The frictional forces occurring in a real laboratory setup are characterized by the Stribeck effect as well as the transition between two different dynamic behaviors, sticking and non-sticking. These switching nonlinear dynamics are identified with various machine learning methodologies in a data-driven manner, i.e., the unsupervised separation and feature clustering of measured state trajectories into two dynamic classes, and the supervised classification of a state-dependent switching condition. The identified system with the interior switching-structure of two dynamics is combined with a moving horizon estimator.