Neural network model based control of nonlinear flexible link systems
Abstract
The presence of link flexibilities in multilink manipulators increases the system order by the number of modes retained when the assumed modes approach is taken or by the number of independent coordinates of a finite element description of the elastic member. The resulting model is a highly coupled nonlinear model of higher complexity than the rigid manipulator model. Such complex dynamics are ignored in non-model based controllers which are simpler and limited in performance compared to model based controllers. Recently, model based predictive controllers based on linear models have been shown to give performance improvements over non-model based controllers in plants with variable dead times and non-minimum phase effects. When the plant is highly nonlinear, it can be advantageous to consider a nonlinear model such as the multilayer perceptron (MLP) network that can approximate to arbitrary precision the nonlinearities over an operating range. Based on recursive long range predictions of the neural model, an input sequence that minimizes a cost function is approximated by a second order descent method. This is equivalent to a Backpropagation where the derivatives are calculated with respect to future control inputs rather than the network weights. The predictive control framework allows variations in model order, variable dead times and nonminimum phase effects in the plant. The modeling part presents a derivation of the equations of motion based on finite element analysis that results in a more structured set of equations for the flexible link manipulator compared to previous derivations. The new set of equations preserves the structure of the rigid manipulator model but single terms corresponding to a rigid linkage are expanded into submatrices of a chosen dimension when link flexibilities are introduced. Furthermore, the new dynamic equations satisfy the passivity properties needed for many nonlinear adaptive control schemes.
Degree
Ph.D.
Advisors
Koivo, Purdue University.
Subject Area
Electrical engineering
Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server.