Diseño de un controlador aplicado a un péndulo invertido utilizando estrategias basadas en aprendizaje de máquina

Abstract

This document presents the development of a controller applied to a simple inverted pendulum, using strategies based on machine learning. For the development of this project, a simulated platform is used in the Simulink software, which performs a respective characterization of the system and a representation of the plant from a block diagram implemented in this Software. In this document, the non-linear equations of the plant are shown so that, based on its behavior, a control based on neural networks is designed that is capable of stabilizing the angular position of the inverted pendulum around a specific work point. The development of the project is divided into 4 important phases.

The first phase consists of knowing the input and output variables of the plant that will be used for the design of the neural network. The input variable refers to the voltage injected into the plant, while the output variables refer to the angles of the pendulum and the rotating arm. The equations that describe the behavior of the inverted pendulum system are also shown. In the second phase, a review of the state of the art is made, this in order to observe methodologies implemented in previous works, taking as a starting point some machine learning methods that can be used for angular position control in a pendulum. The third phase consists of obtaining the plant data that will be used in the design of the neural networks as a set of training and testing data. From the data obtained in the third phase, in the fourth phase the neural network is implemented with its respective training and it will be in charge of stabilizing the pendulum.

The experimental results that were carried out on the implemented neural networks are presented, taking into account different tests carried out, making changes in parameters such as delays, cycles, sampling frequency or input data. An analysis of these tests and of the graphs obtained from the system from the error parameter is also made, in order to observe the behavior of the pendulum once the neural network has been trained.