Deep learning based model predictive controller on a magnetic levitation ball system.

Area of Science:

• Control Systems Engineering
• Artificial Intelligence
• Nonlinear Dynamics

Background:

• Magnetic levitation (maglev) systems are crucial industrial devices known for nonlinearity and open-loop instability.
• Accurate modeling of these systems is essential for effective control design.

Purpose of the Study:

• To develop a precise pseudo-linear model for the maglev ball system using deep learning.
• To design an efficient model predictive controller (MPC) and a predictive functional controller (PFC) for real-time maglev control.

Main Methods:

• A Long Short-Term Memory (LSTM) network was modified into an auto-regressive model with exogenous input variables (LSTM-ARX) to deduce a locally linearized model.
• The LSTM-ARX model was converted into a linear parameter varying (LPV) state-space model.
• A model predictive controller (MPC) was derived, linearized for quadratic programming, and enhanced with a predictive functional controller (PFC) to reduce online optimization complexity.

Main Results:

• The proposed LSTM-ARX model accurately describes the maglev ball system dynamics.
• The LSTM-ARX-based Predictive Functional Controller (LSTM-ARX-PFC) was successfully implemented for real-time control.
• Both simulation and experimental results demonstrated improved transient performance and efficiency.

Conclusions:

• The LSTM-ARX-PFC offers a robust and efficient solution for controlling nonlinear, unstable systems like maglev.
• Deep learning-based modeling combined with advanced control strategies provides significant advantages in system performance.