A MODULAR NEURAL-NETWORK APPROACH TO FAULT-DIAGNOSIS

Certain real-world applications present serious challenges to conventi onal neural-network design procedures. Blindly trying to train huge ne tworks may lead to unsatisfactory results and wrong conclusions about the type of problems that can be tackled using that technology, In thi s paper a modular solution to power systems alarm handling and fault d iagnosis is described that overcomes the limitations of ''toy'' altern atives constrained to small and fixed-topology electrical networks. In contrast to mono-lithical diagnosis systems, the neural-network-based approach presented here accomplishes the scalability and dynamic adap tability requirements of the application. Mapping the power grid onto a set of interconnected modules that model the functional behavior of electrical equipment provides the flexibility and speed demanded by th e problem. After a preliminary generation of candidate fault locations , competition among hypotheses results in a fully justified diagnosis that may include simultaneous faults. The way in which the neural syst em is conceived allows for a natural parallel implementation.