Intelligent Supervision System for Underactuated Mechanical Systems
Underactuated mechanical systems, Sliding Mode Control, Artificial Neural Networks, Reinforcement Learning, Upper Confidence Bound.
Underactuated mechanical systems (UMS) are frequently encountered in several industrial and real-world applications such as robotic manipulators with elastic components, aerospace vehicles, marine vessels, and overhead container cranes. The design of accurate controllers for this kind of mechanical system can become very challenging, especially if a high level of uncertainty is involved. Furthermore, to increase reliability and productivity, it is important that control algorithms can make the system capable of adapting to different conditions from those of the design, such as changes in setpoints and failures that may occur in the actuators. In this sense, an online tuning of the controller parameters is necessary to improve the control performance when the operating settings are changed. This work presents a Sliding Mode controller for UMS using Artificial Neural Networks to represent and compensate unmodeled dynamics and modeling inaccuracies and a new intelligent supervision system based on Reinforcement Learning to adjust the controller parameters. The stability properties are proven by means Lyapunov Stability Theorem. Regarding the Intelligent Supervisor, it was used the MAB (Multi-Armed Bandit) problem approach to be the basis for the controller parameters adjustment algorithm based on the UCB (Upper Confidence Bound) method. This methodology is tested with the trajectory tracking problem of an overhead container crane.