Neural-Network based Swing-up and Stabilization Control of Rotary Inverted Pendulum Systems

Abstract

We propose the neural-network based control (NNC) approach for rotary inverted pendulum (RIP). The control structure for RIP consists of two phases: (i) swing-up phase, which drives the pendulum up towards the desired upright position; and (ii) stabilization phase, which enables the pendulum to keep the desired upright position. In our paper, the swing-up controller is designed based on energy control approach with feedback linearization technique, where the corresponding control gain Kec is determined by the proposed NNC. Next, the stabilization controller is designed based on modified super-twisting sliding-mode control (MSTSMC) algorithm, which requires to design a new sliding surface in order to achieve the asymptotic stability of RIP. Similar to the swing-up phase, the control gain KSS of the MSTSMC is determined intelligently by the proposed NNC. In fact, the proposed NNC allows to select Kec and KSS adaptively based on the operation of RIP. Finally, we provide the experimental comparisons between the proposed NNC and the control system of RIP without the neural-network framework. Through the experimental results, we show that by the NNC, the proposed control system of RIP provides the better swing-up and stabilization performances.