Recognising patterns in rejection data for diagnosis

Abstract

A new method for representing the causal relationship has been proposed by Ransing.12 The causes for the occurrence of defects are classified into two categories. The first category is the “rootcauses”, comprising the causes which are related to the actual process, material and design parameters and these can be directly monitored. The second category is more abstract in nature and is classified as “metacauses”. These capture the scientific rationale implicit in the cause and defect relationships. The knowledge on causes and defects is represented in the form of a network rather than by a set of rules. A new diagnostic paradigm which takes into consideration the occurrence and non-occurrence of defects among the defective components during the diagnosis has been implemented. The pattern of defects is presented to the defect-metacause-rootcause network. The authors have presented two semantically constrained network algorithms to recognise and diagnose these defect patterns. The first is based on the Bayesian probability whereas the second is based on the “feed forward” neural network. Both algorithms have been compared. The semantically constrained neural network algorithm has been found to be a better technique for analysing casting defects.