USING A NEURAL NETWORK SYSTEM FOR ADVANCED PROCESS-CONTROL

Like the human brain, an artificial neural network learns by example. In process control, neural networks are useful because of their ability to learn from experience, generalize from previous examples, and extract the essential characteristics from process data containing irrelevant information. Rudd explains neural network technology and describes a mill trial in which an artificial neural network system was used on a brownstock washer to determine values for mat consistency and mat density. The purpose was to control the dilution factor to stabilize the black liquor solids, and laboratory data were collected to train the network. In an eight-day trial, the standard deviation of the black liquor solids was reduced by 25%. When the network was trained to also supply soda loss values, control performance was just as accurate. Other expert systems require that a set of rules be defined for each process situation, but the neural network learns by example and makes generalizations by training on actual data. This is an advantage for process control engineers, who must control variables that cannot be directly or instantaneously measured and which respond slowly to adjustments.