[home] [Personal Program] [Help]

---

Session: Classical and Machine learning methods in flow control II
Session starts: Tuesday 04 November, 11:40
Presentation starts: 11:40
Room: Lecture room B

---

Satoshi Tsuruhara (Graduate School of Engineering and Science, Shibaura Institute of Technology)
Kazuhisa Ito (Department of Machinery and Control Systems, Shibaura Institute of Technology)

Abstract:
A tap-water-driven artificial muscle has the advantages of both McKibben-type artificial muscles and tap-water-driven systems, which mainly have low environmental burden and cleanliness. However, it is difficult to model and control due to the strong asymmetric hysteresis characteristics of the muscle, so data-driven controls have been actively studied because they do not explicitly require the mathematical model. However, designing controllers from only data is often impractical, e.g. difficult to consider constraints such as the applied voltage to a valve, so we consider a novel control method that combines model-based and data-driven methods for muscle. In this study, we propose a novel two-layered forgetting factor-based indirect adaptive model predictive control (AMPC) that approximates the nonlinear characteristics of the muscle with hysteresis characteristics as a linear time-varying system using only the linear model structure of the muscle. The proposed method shows a fast parameter convergence rate to the neighborhood of true values for this time-varying system without requiring prior system identification. As a result, the proposed method effectively compensates the effect of hysteresis of the muscle at each time step and allows input constraints to be considered via AMPC. The experiment demonstrates the effectiveness of the proposed method by comparing the four methods.