Neural networks in process control

In this thesis the application of neural networks for the modeling and control of nonlinear dynamical processes is studied. The issue of lack of dynamics of standard neurons is first addressed. The development is focussed on improving the dynamics of the NN at the neuronal structure level. New neuronal structures that have memory of both its input as well as output are proposed. This greatly increases the scope of applications where such memory is beneficial. New backpropagation algorithms to train networks comprised of the new neurons are developed that include the features of adaptive learning rate and momentum terms. Dynamic modeling studies based on simulation, laboratory and industrial data confirm the excellent modeling capability of networks composed of these new neurons.; A new neural adaptive control scheme is introduced. The issue of the need for parsimonious neural network structures to facilitate online adaptive applications is dealt with. Also, the requirement of an additional neural network for modeling the dynamics of the nonlinear process is removed, thereby greatly improving the scope for real-time applications. These new controllers have very few parameters (comparable to that of the classical PID controller) and are ideal for online adaptive control applications. In addition, these controllers have the property that their outputs are bounded and hence for bounded-input-bounded-output stable systems, the closed-loop response with the new controller will always be bounded. Several control simulations of nonlinear processes such as an exothermic reaction temperature control problem and a highly nonlinear pH process demonstrate the superiority of the new adaptive neural network controllers.