| The objective of this research study is to develop a method for adaptive state filtering in complex process systems using recurrent neural networks (NN). Nonparametric and nonlinear adaptive state filtering methods, such as the one developed here, have many applications in the areas of inferential process control, adaptive process control, condition monitoring, and fault diagnosis.; Conventional state filtering methods make use of assumed linear models of the underlying process, and thus have restricted applicability to complex process systems, if any. In numerous industrial processes, however, empirical nonlinear process models have much greater applicability than linear process models. In this study, certain classes of NNs, such as the Feedforward Multilayer Perceptron (FMLP) and the Recurrent Multilayer Perceptron (RMLP), are cast in the form of nonlinear and nonparametric model structures, and they are shown to be very effective in modeling complex process systems.; The proposed nonlinear and nonparametric state filtering method is based on the fundamental principles of the standard Kalman Filter. However, minimal assumptions are placed upon the process model used in the developed estimator. The key assumptions on the process model utilized stem from the limitations of the NNs to approximate arbitrary nonlinear functions. Further, no assumptions are placed on the process model noise throughout the development. In fact, the developed method formulates a minimum variance filter, which is designed using least-squares algorithms. Specifically, nonadaptive and adaptive forms of the proposed state filtering method using recurrent NNs is developed, and its applicability is investigated. Additionally, a hybrid form of the state filtering method is developed, which uses both the nonadaptive and adaptive developments.; The proposed state filtering method is applied to three simulated nonlinear process systems, with increasing levels of complexity, and utilizing different forms of system models. The first of these systems is an artificial problem. The second is a DC Motor-Pump system. The third, and most complex system, is a U-Tube Steam Generator (UTSG) system. In addition to estimating dynamic states, the proposed filtering method is used to estimate constant parameters of the DC Motor-Pump and the UTSG systems. It is found that the proposed filtering method performs well in all cases studied. As anticipated, the accuracy of the state estimates are directly dependent upon the fidelity of the process models used in the filter development. The choice between an FMLP or an RMLP depends on the specific application. However, as the system complexity increases, the need for the RMLP is established.; This research study and the accompanying simulation results demonstrate that certain NNs can be effectively used for adaptive nonlinear state filtering, when very little is known explicitly about the dynamics of the process under consideration. This conclusion complements the well-known capabilities of certain NNs in modeling the input-output behavior of complex systems, and it is, again, attributed to their superior function approximation capability. Experimental verification of these simulation results is warranted, and it should be pursued as the next step towards building confidence in these nonlinear computational tools. |
