Process control of nonlinear time variant processes via artificial neural network

One of the major difficulties in analyzing the dynamic responses of industrial processes is the fact that the characteristics of processes frequently change over time. For time varying processes, conventional neural networks which are trained off-line cannot accurately predict process outputs over time. Hence, it is important to introduce recent process data to the network and update their weights as soon as new data is received. In this paper we introduce an algorithm to train the neural network model recursively. A controller is developed using a mathematical model and is employed as a nonlinear controller. The resulting architecture controls the process extremely well in the region of interest and has ability to learn continuously from new incoming data