Alarm processing based on associative neural memories with explanatory capability

System diagnosis can be divided in two parts: alarm processing and fault lo cation. The former is concerned with the identification of the cause(s) for the firing of alarms. The latter is related to finding the faulty componen t(s). This paper presents an expansion of the Itaipu power plant operation guidance system. This project has begun with the development of a fault loc ation system for the Itaipu 500 kV gas-insulated substation. It has been cl aimed that the application of artificial neural networks to fault location is more reliable if associative memory models are used instead of pattern r ecognizers. In this paper, the same argument is extended to alarm processin g. Besides this extension, the major shortcoming of the neural network appr oach for alarm processing is overcome, i.e., its opacity. An algorithm for qualitatively justifying a neural network inference is applied. Initially, the paper compares the performance for alarm processing of the Optimal Nonl inear Associative Memory with other artificial neural network models. After wards, it describes the alarm processing system developed for the Itaip' ge neration units.