Neural Network Limitations and Database Requirements

Using neural networks for plant control is challenging. The objective was to show how to make a neural network effective, even with noise, lag times, and unmeasured disturbances in the plant. The procedure was to evaluate the neural network in a laboratory flotation vision system and in a plant refractory gold autoclave system. Although the resulting neural network easily conquered simple laboratory data, the noisy plant data was much more difficult, and adaptive linear models were equally effective. We concluded that the neural network needs support by combining it with other tools found in the best control packages.