A semantically constrained Bayesian network for manufacturing diagnosis

Abstract

The diagnostic problem is posed as recognizing patterns in rejection data and the subsequent mapping to causes. A new network architecture has been proposed which should overcome many of the disadvantages of the existing diagnostic tools. The network is based on the authors' earlier work (Ransing et al . 1995) on representing the causal relationship in the defect-metacause-rootcause form. Although the algorithm is based on the Bayesian analysis, many of the laws of probability have been altered to suit the complexities involved. For example, the notion of conditional probability has been generalized to enable the belief revision even in the presence of partial evidence. The inherent presence of the degree of ignorance or uncertainty in the quantification of a relationship has also been considered. Rigorous constraints, again based on the laws of probability, have been developed to check the consistency among the network values. The network is required to be initialized with only a few values or the range for the same and then a set of globally consistent values is generated automatically and efficiently. Using the most suitable set of consistent values, the diagnosis is performed using the generalized Bayesian analysis. The network has been tested for a pressure die casting process, however, it is generic in nature and can also be applied to other manufacturing processes.