Self learning process control systems by neural networks

Abstract

The authors discuss the development of a backpropagation neural network control scheme with application to two kinds of processes: a water bath and a multi-input multi-output furnace. The scheme can be easily implemented to solve two different control problems using only the input-output characteristics of the plants without the need for any initial conventional controller or knowledge regarding dynamics. The neurocontrol schemes developed on the two processes have been compared to self-tuning control and conventional digital-PID (proportional plus integral plus derivative) control schemes. Several experiments have been conducted to show the reliability of the neurocontrol scheme. The experimental results also show that the neural-network-based processes are superior and robust. Drawbacks in the neuro-control scheme are the requirement for prior training and the need for a judicious selection of the neural network models. >