Robustness barriers in LQG/LTR Controller via Hybrid Model Genetic-Neural in Robotic Manipulator.

Ivanildo Silva Abreu¹; Alfredo Costa Oliveira Junior¹; Ismael Silva de Melo¹; José Charles Medeiros¹; Paula da Costa Sousa¹

¹ Universidade Estadual do Maranhão

doi:10.20906/CPS/COB-2015-0799

Resumo

The objective of this research is to design an LQG/LTR controller to retrieve the robustness properties that are lost with the inclusion of a stochastic estimator to estimate the state variables and to design the performance and stability barriers of robustness. The plant used is a robotic manipulator, wherein the model is the fourth order and represented in state space. The design methodology uses a controller positioned in the right branch of the control loop, so that robustness requirements are met via Bode diagram. The dynamics of the closed loop system is decoupled, the dynamics of the nominal plant with direct feedback of states and observer states, that is the property of separation. The LQG/LTR design is developed separately from the controller and observer states. In the first stage, we use the LQR controller for achieving the states. For the LQR design, obtaining of the gain depends on the solution of Algebraic Riccati Equation (ARE). Its resolution is carried out by two approaches of Computational Intelligence, Genetic Algorithm (GA) and Artificial Neural Networks (ANN), for allocation of state and control weighting matrix, and resolution of the ARE, respectively. The GA selects the weighting matrices in a unrestricted and nonlinear optimization structure, and the same is assessed by sensitivity and statistical metrics. For ARE solution uses up a recurring ANN, whose architecture is formulated by an unrestricted optimization problem, and the associated energy function is convex and depends on the ARE solution and Cholesky factor. The ANN weights are evaluated by the surface of the infinite norm. The LQG controller is based on the formulation of a stochastic control problem, Kalman filter, to estimate the variables states that are not available for the design. In this robust controller to determine the target is a control law which minimizes the expected value of a quadratic performance index through a feedback output. We use a GA to determine the covariance matrices, which is assessed by the sensitiv

Palavras-chave: genetic algorithm; neural network; performance and stability barriers; singular value decomposition