Optimal tuning of controller parameters for a magnetic levitation system using radial basis based neural network metamodeling approach

• Main Author: Mohammad Abdalhadi, Abdualrhman Daw
• Format: Thesis
• Language: English
• Published: 2021
• Subjects: TK7885-7895 Computer engineer. Computer hardware
• Online Access: http://eprints.utm.my/id/eprint/97301/1/AbdualrhmanDawMohammadAbdalhadiMSEE2021.pdf.pdf; http://eprints.utm.my/id/eprint/97301/; http://dms.library.utm.my:8080/vital/access/manager/Repository/vital:144545

The Magnetic Levitation System (MLS) is a challenging nonlinear mechatronic system in which an electromagnetic force required to suspend an object (metal sphere) in the air. The electromagnetic force is very sensitive to the noise, which can create acceleration forces on the metal sphere, causing the sphere to move into the unbalanced region. Maglev’s benefits the industry, and the system has reduced power consumption, has increased power efficiency, and reduced maintenance cost. The typical applications for Maglev’s Power Generation, for example, wind turbine, Maglev’s trains, and Medical Device (magnetically suspended Artificial Heart Pump). This project presents a comparative assessment of controllers for the magnetic levitation system and the way of optimally tune of the PID parameter. The magnetic levitation system divided into two types, attractive and repulsive, in this project attractive type has been chosen. The analysis will be performed after finding the state space model of magnetic levitation system, and simulation will be performed using MATLAB Simulink. The optimal tuning based PID controller will offer a transient response with better overshoot and rise time than the standard optimization methods. For the trained networks, metamodel radial basis function networks perform more robustly and tolerantly than the gradient descent method even when dealing with noised input data set. The simulation output using the radial basis based metamodel approach showed an overshoot of 9.34% and rise time 9.84ms, which is better than the gradient descent and conventional PID methods.