Abstract
Industrial plants have stringent requirements on product quality, therefore real time measurements such as online analyzers and laboratory analysis are used as process quality indicators. However, online analyzers are very costly and maintenance intensive. These concerns motivate the development of prediction models. Yet, the highly nonlinear relationships between the process variables (inputs) and the product (outputs) have limited the chances to come up with reliable mathematical models. The implementation of intelligent control technology based on artificial neural networks has shown significant results, especially in highly nonlinear applications. This project discusses the methodology and implementation of inferential model based on artificial neural networks using various backpropagation learning algorithms such as gradient descent, scaled conjugate gradient, Bayesian regularization and Lavenberg-Marquardt. The objective is to enhance the online prediction and reduce the necessity of costly online analyzers. This project also discusses alternative approaches to model an inferential where artificial neural networks, extended Kalman filter and process noise estimation model are used in conjunction to solve learning problems. The project addresses the drawback of backpropagation learning algorithms and proposes different learning approach. The results show significant potential for this algorithm to be used in industrial applications.