Nonlinear and learning control of a one-dimensional magnetic manipulator

Nonlinear and learning control of a one-dimensional magnetic manipulator

Damsteeg, J.

Babuska, R. (mentor)
Nageshrao, S. (mentor)

Mechanical, Maritime and Materials Engineering

Delft Center for Systems and Control

2015-06-12

This thesis describes the control design for a magnetic manipulator. The experimental setup has four electromagnets (coils) which can be used to shape the magnetic field above the magnets by controlling the currents through all coils. By doing this, the steel ball can be positioned horizontally in one degree of freedom. The control objective consists of accurate and fast regulation of the ball. The magnetic force created by the coils is highly nonlinear. An empirical model is used to approximate the force exerted by each coil. This leads to a constrained nonlinear control problem. Multiple nonlinear controllers are designed: a Feedback Linearization (FL) controller as benchmark, a State Dependent Riccati Equation (SDRE) controller, a Constrained-SDRE (MPC approach) and a Nonlinear Model Predictive Controller (NMPC). Furthermore, two learning controllers are designed: a Reinforcement Learning controller and an Imitation Learning controller based on Local Linear Regression (LLR). All controllers are evaluated in a simulation study. A satisfactory performance was achieved for both the FL controller and the Constrained-SDRE controller. The NMPC is not feasible in real-time and the RL controller did not achieve a satisfactory performance. The other four controllers were successfully implemented on the experimental setup. From the results of the model-based controllers on the setup it can be concluded that the Constrained- SDRE performs best in terms of settling time, overshoot and control effort. The imitation learning controller is able to match the performance of the C-SDRE controller, and performs better than the C-SDRE in terms of adapting to a different ball size.

nonlinear control
magnetic manipulation
learning control

http://resolver.tudelft.nl/uuid:68284ce3-8a04-4ff7-9cd5-8f222f5d5e95

2017-06-30

Student theses

master thesis

(c) 2015 Damsteeg, J.