Конференции ИВТ СО РАН

Abstracts

At the analysis of potentially dangerous events (technological catastrophes, extreme phenomena of nature etc) there are often used models which have the form of fault trees (also called error trees). Fault tree allows to present in the graphical form a hierarchy of consecutively occurring stochastic events which cause certain undesirable event. On the basis of fault tree analysis, the forecast of the undesirable event can be made, as well as measures for its prevention elaborated.

Usually fault tree is designed by an expert on the basis of his experience and other information on the given event and on similar events. Often the expert may use statistical information describing the object (phenomena) under investigation and similar objects, and this information can also be used for the construction of fault tree.

In the given work we suggest a method for fault tree design (or construction of some of its parts, for finding of which it is necessary to make data analysis) by means of the analysis of multivariate time series describing the dynamic properties of object under investigation and similar objects. Suppose that at each moment the object is described by some features (part of them can be qualitative). The essence of the proposed method is in using the algorithms for dynamic objects recognition based on decision trees. Decision tree has a structure similar to fault tree, and so it is suitable for using as a basic element for fault tree construction. In the nodes of decision tree, the truth of some statements on lagged features are checked. Leaves of decision tree correspond to forecast classes (i.e. presence or absence of the undesirable event). Sets of statements, corresponding to undesirable events, are selected for forming the events of the next layer of fault tree (under the event we understand the true or false of corresponding statement). The event descriptions obtained can be corrected and complemented by the expert.

The Monte-Carlo modeling was used to check the efficiency of the method.

Note. Abstracts are published in author's edition

Comments

[Home] [Conference]

© 1996-2000, Institute of computational Techologies SB RAS, Novosibirsk
© 1996-2000, Siberian Branch of Russian Academy of Science, Novosibirsk