Metodologia e síntese de controladores híbridos clássico-neurais aplicados a sistemas mimo não lineares: estudo de caso em helicóptero 2-dof

Abstract

This dissertation presents the development and application of a hybrid control methodology based on the integration of decentralized classical RST-PI structures with artificial neural networks of the Multilayer Perceptron (MLP) type, applied to the position control of a didactic two-degrees-of-freedom (2-DOF) helicopter system. The studied plant exhibits characteristics of a nonlinear multivariable system with dynamic coupling between the yaw and pitch axes, subject to disturbances and parametric variations that hinder the derivation of accurate analytical models and the efficient tuning of local conventional controllers. To overcome these limitations, system identification was initially carried out using the Recursive Least Squares (RLS) algorithm in state space, allowing for the estimation of linearized models for different operating regions. A model selection strategy based on a cost function from the RLS method was proposed to choose the most suitable model. Based on the estimated linear model, classical controllers with RST-PI and RST-PID structures were designed using pole placement techniques. These controllers were tested in both simulations and real-time experiments, and their performance was evaluated using metrics such as mean integral of the squared error (ISE), mean integral of the squared control signal (ISU), variance of control and error signals, and robustness indicators based on sensitivity function analysis. Using experimental data collected from the real process, a neural network was trained offline to identify the plant’s operating region at each instant and dynamically adapt the controller gains, aiming to enhance transient response and reduce control effort. The experimental results demonstrated that the proposed hybrid approach effectively reduced oscillations, smoothed the control signal, and improved tracking performance under abrupt variations and coupling effects. The developed methodology proved to be both technically and practically viable, offering a promising direction for the application of artificial intelligence to enhance classical control strategies in nonlinear multivariable systems with uncertain and time-varying dynamics. Future work includes exploring neural-based adaptation in centralized or more sophisticated controllers.