NONLINEAR MODELING OF A REACTOR-EXCHANGER BY USING NARX NEURAL NETWORKS
Systematic methods and tools for managing the complexity
Tools Integration - CAPE Methods & Tools (T4-10P)
Keywords: Neural Network; NARX; Modeling; Nonlinear identification
Abstract
Process development and continuous request for productivity led to an increasing complexity of industrial units. In chemical industries, it is absolutely necessary to control the process and any drift or anomaly must be detected as soon as possible in order to prevent risks and accidents. Moreover, detecting a fault appearance on-line is justified by the need to solve effectively the problems within a short time (Villermaux, 1996; Chetouani et al., 2004; Chetouani, 2006).
The intrinsic highly nonlinear behavior in the industrial process, especially when a chemical reaction is used, poses a major problem for the formulation of good predictions and the design of reliable control systems (Cammarata et al., 2002). Due to the relevant number of degree of freedom, to the nonlinear coupling of different phenomena and to the processes complexity, the mathematical modeling of the process is computationally heavy and may produce an unsatisfactory correspondence between experimental and simulated data. Similar problems arise also from the uncertainty for the parameters of the process, such as the reaction rate, activation energy, reaction enthalpy, heat transfer coefficient, and their unpredictable variations. In fact, note that most of the chemical and thermo-physical variables both strongly depend and influence instantaneously the temperature of the reaction mass (Chetouani et al., 2006). One way of addressing this problem is the use of a reliable model for the on-line prediction of the system dynamic evolution. However, designing empirical models like the black-box models is unavoidable (Leontaritis et al., 1985).
The purpose of this identification is to establish a reliable model of the dynamic behavior of a process like a reactor-exchanger. This reliable model enables to reproduce the process dynamics under different operating conditions in a normal mode. We are interesting in the anomaly detection module intended to supervise the functioning state of the system (Chetouani et al., 2006). The former has to generate on-line information concerning the state of the automated system. This state is characterized not only by control and measurement variables (temperature, rate, etc.), but also by the general behavior of the process and its history, showing in time whether the behavior of the system is normal or presents drifts. In the context of numerical control, fault detection and isolation (FDI) proves a vital complement to the adaptive means of dealing with instationarities in nonlinear highly non-stationary systems. Under normal conditions, the fault detection module allows all information to be processed and managed in direct liaison with its general behavior. In other case, it detects any anomaly and alerts the operator by setting on the appropriate alarms.
The main aim of this paper is to establish a reliable model of a process behavior under its normal operating conditions. The use of this model should reflect the true behavior of the process and allow to distinguish a normal mode from an abnormal one. In order to obtain a reliable model for the process dynamics, the black-box identification by means of a NARX (Nonlinear Auto-Regressive with eXogenous input) model has been chosen in this study. It is based on the neural network approach. This paper shows the choice and the performance of the neural network in the training and test phases. An analysis of the inputs number, hidden neurons and their influence on the behavior of the neural predictor is carried out. Three statistical criteria; Aikeke’s Information Criterion (AIC), Rissanen’s Minimum Description Length (MDL) and Bayesian Information Criterion (BIC) are used for the validation of the experimental data. A reactor-exchanger is used to illustrate the proposed ideas concerning the dynamics modeling. The outlet temperature is modeled according to the inlet one. The model is implemented by training a Multi-Layer Perceptron (MLP) Artificial neural network with input-output experimental data. Satisfactory agreement between identified and experimental data is found and results show that the model successfully predicts the evolution of the outlet temperature of the process.
See the full pdf manuscript of the abstract.
Presented Thursday 20, 13:30 to 14:40, in session Tools Integration - CAPE Methods & Tools (T4-10P).
