Parametric and Neural Network Models for Fault Detection and Isolation of Industrial Process Sub-Modules

Abstract

The paper presents results of a linear Auto-Regressive with eXogenous input (ARX), Nonlinear Auto-Regressive with eXogenous input (NARX), and Hammerstein model identification. The NARX and Hammerstein models are of a neural network type. The models approximate a dependence of the juice steam pressure in stage one on the juice steam pressure in stage three in a five stage sugar evaporator. While the NARX model has been trained with the backpropagation algorithm, a combined backpropagation and recursive least squares learning algorithm has been employed for the neural network Hammerstein model. The models have been identified using real process data recorded at the normal operation conditions and can be used for fault detection directly. To perform a fault isolation step, models of the sub-module at its faulty conditions can be identified using a similar approach. For all models, the order of two and the delay of 20s have been assumed. From comparison of model testing results, it follows that the highest accuracy has been obtained for the Hammerstein model for the training set, and the ARX model for the testing set.