The state space bounded derivative network superceding the application of neural networks in control

摘要

This paper introduces a challenge to the general acceptance of neural networks being ‘ideally suited’ for use in nonlinear control schemes. The paper briefly outlines 10 significant reasons as to why neural networks should not be used in any control system that directly affects process plant. The State Space Bounded Derivative Network will then be presented as a universal approximating architecture that encompasses the power of approximation of neural networks but without the failings. This algorithm has now been widely applied to the industrial control of polymer plants worldwide and has been the key enabling technology for Aspen Apollo — the Worlds' first commercial truly universal model based controller. The unique features of the SSBDN include globally guaranteed invertibility; global constraints on the model gains; robust, elegant and intelligent extrapolation capability and the capability of modelling both positional and directionally dependent dynamic nonlinearities. A commercial application of this technology to an industrial polyethylene unit will be given.