On the identification of neural network models during closed-loop operation

In recent years there has been considerable interest in developing nonlinear dynamic models in the form of neural networks (NN) from input-output data. These empirical models can then be used for a variety of purposes which include process monitoring and on-line control (Hunt et al, 1992); Su and McAvoy, 1997). Virtually all of the large number of publications on neural net models have been restricted to model identification under open-loop conditions. But in many practical applications, closed-loop identification is prefered by plant personnel. This important and timely problem is the subject of the present paper. To the best of our knowledge, the only previous publication concerned with closed-loop identification of NN dynamic models is the very recent paper by Cheng et al (1996) which was published during the final stage of our research. In this paper, we explore the feasiblity of closed-loop identification of neural net, dynamic models and investigate whether such models can be used to design nonlinear, modelbased controllers. Since analytical techniques are not available for this type of investigation, we use a case study approach based on a simulated pH neutralization process that has been utilized in many previous studies (Seborg, 1994). A number of practical identification issues are considered which include: (1) type of input excitation, (2) effect of controller settings; (3) effect of noise level, and (4) comparison of models obtained from both open-loop (OL) and closed-loop (CL) operation. We then evaluate nonlinear controllers designed using models identified from both OL and CL operation.