A magnetic levitation testbed for development of real-time control frameworks applied in fusion

Abstract

This thesis presents the development of a magnetic levitation device as a hardware-in-theloop platform to be used for research in Control and Data Acquisition frameworks applied to fusion experiments. Specifically, the testbed is aimed to demonstrate distributed, modular control using a plasma control system framework being developed at the Plasma Science and Fusion Center at MIT. This framework integrates a real-time control framework, MARTe2, and a data management framework, MDSplus, to provide platform flexibility and robust data management for rapid prototyping of control systems. Both frameworks are widely used individually in fusion experiments worldwide. The magnetic levitation setup is centered around a single electromagnet coil which levitates a permanent disk magnet from above. Implemented with the integrated MARTe2/MDSplus framework, the controller, actuator, and sensors are distributed over the network. With the magnetic levitation testbed, this thesis achieves three objectives: 1. formulation of a physicsbased model of the system, 2. development of a controller in a modular, networked framework, and 3. training and implementation of learning-based methods within the framework. First, a state-space model for single-axis magnetic levitation is formulated based on theory and refined with magnetic field measurements. A feedback controller is then developed and implemented with MATLAB Simulink. Afterwards, a vision-based observer is developed to estimate position and tilt of the levitated magnet. Pose-image datasets are auto-labeled using fiducial markers and are used to train a convolutional neural network. Finally, the trained network will be applied in system identification of the final controlled system. Through the process of system development, this thesis proposes that the integrated MARTe2/MDSplus framework is robust in performing real-time control of a networked system, and its structural modularity is advantageous for developing and testing learning-based models.