CONTINUOUS STEEL CASTING DIAGNOSTIC FUZZY EXPERT SYSTEM IN AN ARTIFICIAL NEURAL NETWORK

This paper demonstrates the implementation of a fuzzy expert system of a realistic size in the form of a productive neural network. Productive networks are simple to construct, do not have a large number of connections, and do not need training. In contrast to the feed-forward neural networks, a meaning is assigned to each node in this kind of networks, i.e. every node performs a well-defined task, and the offsets are fixed using the problem statement. Each node in a productive network collects an offset product of inputs to that node. The offsets are either zero or one, and the nodes do not have complete connectivity across adjacent layers, unlike in feed-forward neural networks. They are somewhat more reliable since the function of each node in the network is known and well understood, and the question of sufficiency of the number of training instances does not arise. Developing an expert system is often time consuming even after a successful knowledge acquisition stage and artificial neural networks offer an alternative. Earlier work illustrated the feasibility of using a feed-forward neural network for knowledge storage and expert system-like reasoning for predicting operational problems in continuous steel casting. The inputs to the network were in terms of truth values, and information about an incoming ladle of steel. The output was about the suitability of the incoming ladle of steel for successful continuous casting, giving an indication whether problems may be encountered in the beginning and/or at the end of the casting. 66 rules generated by inductive learning from hypothetical cases were put into the productive network constiting of 12 inputs, 82 nodes in the hidden layer, and 2 outputs.