Probabilistic evaluation of failure events with artificial neural networks

An emphasis of plant-specific probabilistic safety analysis (PSA) for nuclear power plants consists in the data extraction and the appropriation of reliability characteristics for the considered components. For its characteristic calculations, component collectives must be formed by descriptive features. Their assignments are normally chosen by operating and component task aspects. Then, if no plant-specific failures of a component group are available, a zero failure statistic is carried out, and the preserved values are afflicted with great uncertainties. Another problem in PSA application is complexity of calculation. Then, a procedure to solve the mentioned problems was proposed. Its application can be used in plant maintenance in a kind of risk monitor by component part failures. And, influence from such failures on the plant safety may be shown on display as a verbal statement. By means of the procedure leading in this work inferences of failure rates at component parts to system safety may close. Calculation with neural nets necessary to this is very simply feasible in kind of a black box. Its application fields are plant maintenance with possibility of immediately recognizing and eliminating relevant failure mechanisms, it can be also use for online risk monitor. (G.K.)