Designing neural networks for manufacturing process control systems

The design of neural networks for the control of discrete manufacturing processes is addressed. Rather than treating the networks as adaptive black boxes, an architecture that links the weights associated with the nodes and thus allows the relationships and internal structure to be tightly constrained is introduced. The constrained search space gives greater confidence in the internal representations that have been induced by the training set and therefore about the correct behavior of the network between the given limits. The method is illustrated by applying it to the dispensing of adhesives