Badania porównawcze wielomodalnych sieci stwierdzeń

Artykuł dotyczy badań porównawczych wielomodalnych sieci stwierdzeń. Przedstawiono definicje sieci stwierdzeń oraz sieci wielomodalnej. Pokazano, że szczególną wersją sieci stwierdzeń może być sieć przekonań lub sieć warunków koniecznych i dostatecznych. Prezentowane sieci wielomodalne zostały opracowane w szkieletowym systemie doradczym MMNET i dotyczą diagnozowania stanu technicznego obiektu stanowiącego fizyczną miniaturyzację instalacji przemysłowej stosowanej, np. w przemyśle chemicznym. Przeprowadzone badania porównawcze pozwoliły określić zalety i wady badanych sieci.

The paper deals with comparative studies of multimodal statement networks. Section 1 contains description of the concept of a multimodal network and a statement. A belief (Bayesian) network as well as a network of necessary and sufficient conditions (exact and approximate) are briefly described in Section 2. Section 3 is the most important part of the paper. It shows that the considered networks can be constructed by means of the MMNET system [4]. Fig. 3 presents the considered technical object which exemplifies a physical miniaturization of an industrial installation used, e.g. in a chemical or food industry. The object operation is illustrated in Fig. 4. It was decided to consider 12 possible technical states of the object including so-called complex states, i.e. the cases when at least 2 elements of the object were faulty. Active diagnostic experiments were conducted in order to obtain sets of data necessary for identification of the examined networks (diagnostic models). Comparison was performed on the basis of the results obtained from the identified models (e.g. Tabs. 2, 3 and 4). In consequence, it was possible to determine advantages and disadvantages of the networks. Tab. 5 shows that comparison. Belief networks are used in many different domains also in technical diagnostics. It seems that multimodal statement networks exemplify an interesting alternative because they have got many advantages especially a possibility of constructing models as a result of integration of independently identified component networks and a possibility of reasoning on the basis of inaccurate and partly contradictory data.