A robust diagnosis system based on neural networks

Summary form only given, as follows. The authors are interested in developing robust and reliable diagnosis systems for nuclear power plants. Traditional tools are not well suited for these tasks because they were not designed to handle large amounts of redundant information. A connectionist architecture for representing symbolic knowledge is proposed. The network is composed of two layers; the first one corresponds to intermediary conclusions and the second one to the final conclusions. The units in both layers account for several symbols and each symbol is represented several times within a layer. The robustness against the destruction of units is improved, and a method for parallel matching and rule firing is devised