Decentralized control of complex dynamic systems employing function emulation by neural networks

Abstract

A novel robust adaptive control design synthesis, which employs both high-order neural networks and math-analytical results for a class of mechatronic nonlinear systems possessing similarity property has been derived. This approach makes adequate use of the structural feature of composite similarity systems and neural networks to solve the representation issue of uncertainty interconnections and subsystem gains by updating online the weight of the neural networks . Lyapunov stability theory and attraction domain analysis are used. This synthesis guarantees the real stability in closed loop but also requires skills to obtain larger attraction domains around the operating equilibrium. The benchmark example of elastically interconnected two inverted pendulums on carts, thus creating a complex nonlinear dynamic system possessing inherent uncertainties, is investigated. The decentralized control of this benchmark plant is solved and its simulation results are given to illustrate the proposed technique.